net/ipv4/ip_output.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              The Internet Protocol (IP) output module.
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Donald Becker, <becker@super.org>
  11  *              Alan Cox, <Alan.Cox@linux.org>
  12  *              Richard Underwood
  13  *              Stefan Becker, <stefanb@yello.ping.de>
  14  *              Jorge Cwik, <jorge@laser.satlink.net>
  15  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  16  *              Hirokazu Takahashi, <taka@valinux.co.jp>
  17  *
  18  *      See ip_input.c for original log
  19  *
  20  *      Fixes:
  21  *              Alan Cox        :       Missing nonblock feature in ip_build_xmit.
  22  *              Mike Kilburn    :       htons() missing in ip_build_xmit.
  23  *              Bradford Johnson:       Fix faulty handling of some frames when
  24  *                                      no route is found.
  25  *              Alexander Demenshin:    Missing sk/skb free in ip_queue_xmit
  26  *                                      (in case if packet not accepted by
  27  *                                      output firewall rules)
  28  *              Mike McLagan    :       Routing by source
  29  *              Alexey Kuznetsov:       use new route cache
  30  *              Andi Kleen:             Fix broken PMTU recovery and remove
  31  *                                      some redundant tests.
  32  *      Vitaly E. Lavrov        :       Transparent proxy revived after year coma.
  33  *              Andi Kleen      :       Replace ip_reply with ip_send_reply.
  34  *              Andi Kleen      :       Split fast and slow ip_build_xmit path
  35  *                                      for decreased register pressure on x86
  36  *                                      and more readibility.
  37  *              Marc Boucher    :       When call_out_firewall returns FW_QUEUE,
  38  *                                      silently drop skb instead of failing with -EPERM.
  39  *              Detlev Wengorz  :       Copy protocol for fragments.
  40  *              Hirokazu Takahashi:     HW checksumming for outgoing UDP
  41  *                                      datagrams.
  42  *              Hirokazu Takahashi:     sendfile() on UDP works now.
  43  */
  44
  45 #include <asm/uaccess.h>
  46 #include <asm/system.h>
  47 #include <linux/module.h>
  48 #include <linux/types.h>
  49 #include <linux/kernel.h>
  50 #include <linux/mm.h>
  51 #include <linux/string.h>
  52 #include <linux/errno.h>
  53 #include <linux/highmem.h>
  54 #include <linux/slab.h>
  55
  56 #include <linux/socket.h>
  57 #include <linux/sockios.h>
  58 #include <linux/in.h>
  59 #include <linux/inet.h>
  60 #include <linux/netdevice.h>
  61 #include <linux/etherdevice.h>
  62 #include <linux/proc_fs.h>
  63 #include <linux/stat.h>
  64 #include <linux/init.h>
  65
  66 #include <net/snmp.h>
  67 #include <net/ip.h>
  68 #include <net/protocol.h>
  69 #include <net/route.h>
  70 #include <net/xfrm.h>
  71 #include <linux/skbuff.h>
  72 #include <net/sock.h>
  73 #include <net/arp.h>
  74 #include <net/icmp.h>
  75 #include <net/checksum.h>
  76 #include <net/inetpeer.h>
  77 #include <linux/igmp.h>
  78 #include <linux/netfilter_ipv4.h>
  79 #include <linux/netfilter_bridge.h>
  80 #include <linux/mroute.h>
  81 #include <linux/netlink.h>
  82 #include <linux/tcp.h>
  83
  84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  85
  86 /* Generate a checksum for an outgoing IP datagram. */
  87 __inline__ 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
  93 int __ip_local_out(struct sk_buff *skb)
  94 {
  95         struct iphdr *iph = ip_hdr(skb);
  96
  97         iph->tot_len = htons(skb->len);
  98         ip_send_check(iph);
  99         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 100                        skb_dst(skb)->dev, dst_output);
 101 }
 102
 103 int ip_local_out(struct sk_buff *skb)
 104 {
 105         int err;
 106
 107         err = __ip_local_out(skb);
 108         if (likely(err == 1))
 109                 err = dst_output(skb);
 110
 111         return err;
 112 }
 113 EXPORT_SYMBOL_GPL(ip_local_out);
 114
 115 /* dev_loopback_xmit for use with netfilter. */
 116 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
 117 {
 118         skb_reset_mac_header(newskb);
 119         __skb_pull(newskb, skb_network_offset(newskb));
 120         newskb->pkt_type = PACKET_LOOPBACK;
 121         newskb->ip_summed = CHECKSUM_UNNECESSARY;
 122         WARN_ON(!skb_dst(newskb));
 123         netif_rx(newskb);
 124         return 0;
 125 }
 126
 127 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 128 {
 129         int ttl = inet->uc_ttl;
 130
 131         if (ttl < 0)
 132                 ttl = dst_metric(dst, RTAX_HOPLIMIT);
 133         return ttl;
 134 }
 135
 136 /*
 137  *              Add an ip header to a skbuff and send it out.
 138  *
 139  */
 140 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 141                           __be32 saddr, __be32 daddr, struct ip_options *opt)
 142 {
 143         struct inet_sock *inet = inet_sk(sk);
 144         struct rtable *rt = skb_rtable(skb);
 145         struct iphdr *iph;
 146
 147         /* Build the IP header. */
 148         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 149         skb_reset_network_header(skb);
 150         iph = ip_hdr(skb);
 151         iph->version  = 4;
 152         iph->ihl      = 5;
 153         iph->tos      = inet->tos;
 154         if (ip_dont_fragment(sk, &rt->u.dst))
 155                 iph->frag_off = htons(IP_DF);
 156         else
 157                 iph->frag_off = 0;
 158         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 159         iph->daddr    = rt->rt_dst;
 160         iph->saddr    = rt->rt_src;
 161         iph->protocol = sk->sk_protocol;
 162         ip_select_ident(iph, &rt->u.dst, sk);
 163
 164         if (opt && opt->optlen) {
 165                 iph->ihl += opt->optlen>>2;
 166                 ip_options_build(skb, opt, daddr, rt, 0);
 167         }
 168
 169         skb->priority = sk->sk_priority;
 170         skb->mark = sk->sk_mark;
 171
 172         /* Send it out. */
 173         return ip_local_out(skb);
 174 }
 175
 176 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 177
 178 static inline int ip_finish_output2(struct sk_buff *skb)
 179 {
 180         struct dst_entry *dst = skb_dst(skb);
 181         struct rtable *rt = (struct rtable *)dst;
 182         struct net_device *dev = dst->dev;
 183         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 184
 185         if (rt->rt_type == RTN_MULTICAST) {
 186                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 187         } else if (rt->rt_type == RTN_BROADCAST)
 188                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 189
 190         /* Be paranoid, rather than too clever. */
 191         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 192                 struct sk_buff *skb2;
 193
 194                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 195                 if (skb2 == NULL) {
 196                         kfree_skb(skb);
 197                         return -ENOMEM;
 198                 }
 199                 if (skb->sk)
 200                         skb_set_owner_w(skb2, skb->sk);
 201                 kfree_skb(skb);
 202                 skb = skb2;
 203         }
 204
 205         if (dst->hh)
 206                 return neigh_hh_output(dst->hh, skb);
 207         else if (dst->neighbour)
 208                 return dst->neighbour->output(skb);
 209
 210         if (net_ratelimit())
 211                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 212         kfree_skb(skb);
 213         return -EINVAL;
 214 }
 215
 216 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 217 {
 218         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 219
 220         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 221                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 222 }
 223
 224 static int ip_finish_output(struct sk_buff *skb)
 225 {
 226 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 227         /* Policy lookup after SNAT yielded a new policy */
 228         if (skb_dst(skb)->xfrm != NULL) {
 229                 IPCB(skb)->flags |= IPSKB_REROUTED;
 230                 return dst_output(skb);
 231         }
 232 #endif
 233         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 234                 return ip_fragment(skb, ip_finish_output2);
 235         else
 236                 return ip_finish_output2(skb);
 237 }
 238
 239 int ip_mc_output(struct sk_buff *skb)
 240 {
 241         struct sock *sk = skb->sk;
 242         struct rtable *rt = skb_rtable(skb);
 243         struct net_device *dev = rt->u.dst.dev;
 244
 245         /*
 246          *      If the indicated interface is up and running, send the packet.
 247          */
 248         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 249
 250         skb->dev = dev;
 251         skb->protocol = htons(ETH_P_IP);
 252
 253         /*
 254          *      Multicasts are looped back for other local users
 255          */
 256
 257         if (rt->rt_flags&RTCF_MULTICAST) {
 258                 if (sk_mc_loop(sk)
 259 #ifdef CONFIG_IP_MROUTE
 260                 /* Small optimization: do not loopback not local frames,
 261                    which returned after forwarding; they will be  dropped
 262                    by ip_mr_input in any case.
 263                    Note, that local frames are looped back to be delivered
 264                    to local recipients.
 265
 266                    This check is duplicated in ip_mr_input at the moment.
 267                  */
 268                     &&
 269                     ((rt->rt_flags & RTCF_LOCAL) ||
 270                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 271 #endif
 272                    ) {
 273                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 274                         if (newskb)
 275                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 276                                         newskb, NULL, newskb->dev,
 277                                         ip_dev_loopback_xmit);
 278                 }
 279
 280                 /* Multicasts with ttl 0 must not go beyond the host */
 281
 282                 if (ip_hdr(skb)->ttl == 0) {
 283                         kfree_skb(skb);
 284                         return 0;
 285                 }
 286         }
 287
 288         if (rt->rt_flags&RTCF_BROADCAST) {
 289                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 290                 if (newskb)
 291                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 292                                 NULL, newskb->dev, ip_dev_loopback_xmit);
 293         }
 294
 295         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 296                             skb->dev, ip_finish_output,
 297                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 298 }
 299
 300 int ip_output(struct sk_buff *skb)
 301 {
 302         struct net_device *dev = skb_dst(skb)->dev;
 303
 304         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 305
 306         skb->dev = dev;
 307         skb->protocol = htons(ETH_P_IP);
 308
 309         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 310                             ip_finish_output,
 311                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 312 }
 313
 314 int ip_queue_xmit(struct sk_buff *skb)
 315 {
 316         struct sock *sk = skb->sk;
 317         struct inet_sock *inet = inet_sk(sk);
 318         struct ip_options *opt = inet->opt;
 319         struct rtable *rt;
 320         struct iphdr *iph;
 321
 322         /* Skip all of this if the packet is already routed,
 323          * f.e. by something like SCTP.
 324          */
 325         rt = skb_rtable(skb);
 326         if (rt != NULL)
 327                 goto packet_routed;
 328
 329         /* Make sure we can route this packet. */
 330         rt = (struct rtable *)__sk_dst_check(sk, 0);
 331         if (rt == NULL) {
 332                 __be32 daddr;
 333
 334                 /* Use correct destination address if we have options. */
 335                 daddr = inet->inet_daddr;
 336                 if(opt && opt->srr)
 337                         daddr = opt->faddr;
 338
 339                 {
 340                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
 341                                             .mark = sk->sk_mark,
 342                                             .nl_u = { .ip4_u =
 343                                                       { .daddr = daddr,
 344                                                         .saddr = inet->inet_saddr,
 345                                                         .tos = RT_CONN_FLAGS(sk) } },
 346                                             .proto = sk->sk_protocol,
 347                                             .flags = inet_sk_flowi_flags(sk),
 348                                             .uli_u = { .ports =
 349                                                        { .sport = inet->inet_sport,
 350                                                          .dport = inet->inet_dport } } };
 351
 352                         /* If this fails, retransmit mechanism of transport layer will
 353                          * keep trying until route appears or the connection times
 354                          * itself out.
 355                          */
 356                         security_sk_classify_flow(sk, &fl);
 357                         if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
 358                                 goto no_route;
 359                 }
 360                 sk_setup_caps(sk, &rt->u.dst);
 361         }
 362         skb_dst_set(skb, dst_clone(&rt->u.dst));
 363
 364 packet_routed:
 365         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
 366                 goto no_route;
 367
 368         /* OK, we know where to send it, allocate and build IP header. */
 369         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 370         skb_reset_network_header(skb);
 371         iph = ip_hdr(skb);
 372         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 373         if (ip_dont_fragment(sk, &rt->u.dst) && !skb->local_df)
 374                 iph->frag_off = htons(IP_DF);
 375         else
 376                 iph->frag_off = 0;
 377         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 378         iph->protocol = sk->sk_protocol;
 379         iph->saddr    = rt->rt_src;
 380         iph->daddr    = rt->rt_dst;
 381         /* Transport layer set skb->h.foo itself. */
 382
 383         if (opt && opt->optlen) {
 384                 iph->ihl += opt->optlen >> 2;
 385                 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
 386         }
 387
 388         ip_select_ident_more(iph, &rt->u.dst, sk,
 389                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 390
 391         skb->priority = sk->sk_priority;
 392         skb->mark = sk->sk_mark;
 393
 394         return ip_local_out(skb);
 395
 396 no_route:
 397         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 398         kfree_skb(skb);
 399         return -EHOSTUNREACH;
 400 }
 401
 402
 403 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 404 {
 405         to->pkt_type = from->pkt_type;
 406         to->priority = from->priority;
 407         to->protocol = from->protocol;
 408         skb_dst_drop(to);
 409         skb_dst_set(to, dst_clone(skb_dst(from)));
 410         to->dev = from->dev;
 411         to->mark = from->mark;
 412
 413         /* Copy the flags to each fragment. */
 414         IPCB(to)->flags = IPCB(from)->flags;
 415
 416 #ifdef CONFIG_NET_SCHED
 417         to->tc_index = from->tc_index;
 418 #endif
 419         nf_copy(to, from);
 420 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 421     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 422         to->nf_trace = from->nf_trace;
 423 #endif
 424 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 425         to->ipvs_property = from->ipvs_property;
 426 #endif
 427         skb_copy_secmark(to, from);
 428 }
 429
 430 /*
 431  *      This IP datagram is too large to be sent in one piece.  Break it up into
 432  *      smaller pieces (each of size equal to IP header plus
 433  *      a block of the data of the original IP data part) that will yet fit in a
 434  *      single device frame, and queue such a frame for sending.
 435  */
 436
 437 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 438 {
 439         struct iphdr *iph;
 440         int raw = 0;
 441         int ptr;
 442         struct net_device *dev;
 443         struct sk_buff *skb2;
 444         unsigned int mtu, hlen, left, len, ll_rs, pad;
 445         int offset;
 446         __be16 not_last_frag;
 447         struct rtable *rt = skb_rtable(skb);
 448         int err = 0;
 449
 450         dev = rt->u.dst.dev;
 451
 452         /*
 453          *      Point into the IP datagram header.
 454          */
 455
 456         iph = ip_hdr(skb);
 457
 458         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 459                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 460                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 461                           htonl(ip_skb_dst_mtu(skb)));
 462                 kfree_skb(skb);
 463                 return -EMSGSIZE;
 464         }
 465
 466         /*
 467          *      Setup starting values.
 468          */
 469
 470         hlen = iph->ihl * 4;
 471         mtu = dst_mtu(&rt->u.dst) - hlen;       /* Size of data space */
 472         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 473
 474         /* When frag_list is given, use it. First, check its validity:
 475          * some transformers could create wrong frag_list or break existing
 476          * one, it is not prohibited. In this case fall back to copying.
 477          *
 478          * LATER: this step can be merged to real generation of fragments,
 479          * we can switch to copy when see the first bad fragment.
 480          */
 481         if (skb_has_frags(skb)) {
 482                 struct sk_buff *frag;
 483                 int first_len = skb_pagelen(skb);
 484                 int truesizes = 0;
 485
 486                 if (first_len - hlen > mtu ||
 487                     ((first_len - hlen) & 7) ||
 488                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
 489                     skb_cloned(skb))
 490                         goto slow_path;
 491
 492                 skb_walk_frags(skb, frag) {
 493                         /* Correct geometry. */
 494                         if (frag->len > mtu ||
 495                             ((frag->len & 7) && frag->next) ||
 496                             skb_headroom(frag) < hlen)
 497                             goto slow_path;
 498
 499                         /* Partially cloned skb? */
 500                         if (skb_shared(frag))
 501                                 goto slow_path;
 502
 503                         BUG_ON(frag->sk);
 504                         if (skb->sk) {
 505                                 frag->sk = skb->sk;
 506                                 frag->destructor = sock_wfree;
 507                         }
 508                         truesizes += frag->truesize;
 509                 }
 510
 511                 /* Everything is OK. Generate! */
 512
 513                 err = 0;
 514                 offset = 0;
 515                 frag = skb_shinfo(skb)->frag_list;
 516                 skb_frag_list_init(skb);
 517                 skb->data_len = first_len - skb_headlen(skb);
 518                 skb->truesize -= truesizes;
 519                 skb->len = first_len;
 520                 iph->tot_len = htons(first_len);
 521                 iph->frag_off = htons(IP_MF);
 522                 ip_send_check(iph);
 523
 524                 for (;;) {
 525                         /* Prepare header of the next frame,
 526                          * before previous one went down. */
 527                         if (frag) {
 528                                 frag->ip_summed = CHECKSUM_NONE;
 529                                 skb_reset_transport_header(frag);
 530                                 __skb_push(frag, hlen);
 531                                 skb_reset_network_header(frag);
 532                                 memcpy(skb_network_header(frag), iph, hlen);
 533                                 iph = ip_hdr(frag);
 534                                 iph->tot_len = htons(frag->len);
 535                                 ip_copy_metadata(frag, skb);
 536                                 if (offset == 0)
 537                                         ip_options_fragment(frag);
 538                                 offset += skb->len - hlen;
 539                                 iph->frag_off = htons(offset>>3);
 540                                 if (frag->next != NULL)
 541                                         iph->frag_off |= htons(IP_MF);
 542                                 /* Ready, complete checksum */
 543                                 ip_send_check(iph);
 544                         }
 545
 546                         err = output(skb);
 547
 548                         if (!err)
 549                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 550                         if (err || !frag)
 551                                 break;
 552
 553                         skb = frag;
 554                         frag = skb->next;
 555                         skb->next = NULL;
 556                 }
 557
 558                 if (err == 0) {
 559                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 560                         return 0;
 561                 }
 562
 563                 while (frag) {
 564                         skb = frag->next;
 565                         kfree_skb(frag);
 566                         frag = skb;
 567                 }
 568                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 569                 return err;
 570         }
 571
 572 slow_path:
 573         left = skb->len - hlen;         /* Space per frame */
 574         ptr = raw + hlen;               /* Where to start from */
 575
 576         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 577          * we need to make room for the encapsulating header
 578          */
 579         pad = nf_bridge_pad(skb);
 580         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, pad);
 581         mtu -= pad;
 582
 583         /*
 584          *      Fragment the datagram.
 585          */
 586
 587         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 588         not_last_frag = iph->frag_off & htons(IP_MF);
 589
 590         /*
 591          *      Keep copying data until we run out.
 592          */
 593
 594         while (left > 0) {
 595                 len = left;
 596                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 597                 if (len > mtu)
 598                         len = mtu;
 599                 /* IF: we are not sending upto and including the packet end
 600                    then align the next start on an eight byte boundary */
 601                 if (len < left) {
 602                         len &= ~7;
 603                 }
 604                 /*
 605                  *      Allocate buffer.
 606                  */
 607
 608                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 609                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 610                         err = -ENOMEM;
 611                         goto fail;
 612                 }
 613
 614                 /*
 615                  *      Set up data on packet
 616                  */
 617
 618                 ip_copy_metadata(skb2, skb);
 619                 skb_reserve(skb2, ll_rs);
 620                 skb_put(skb2, len + hlen);
 621                 skb_reset_network_header(skb2);
 622                 skb2->transport_header = skb2->network_header + hlen;
 623
 624                 /*
 625                  *      Charge the memory for the fragment to any owner
 626                  *      it might possess
 627                  */
 628
 629                 if (skb->sk)
 630                         skb_set_owner_w(skb2, skb->sk);
 631
 632                 /*
 633                  *      Copy the packet header into the new buffer.
 634                  */
 635
 636                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 637
 638                 /*
 639                  *      Copy a block of the IP datagram.
 640                  */
 641                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 642                         BUG();
 643                 left -= len;
 644
 645                 /*
 646                  *      Fill in the new header fields.
 647                  */
 648                 iph = ip_hdr(skb2);
 649                 iph->frag_off = htons((offset >> 3));
 650
 651                 /* ANK: dirty, but effective trick. Upgrade options only if
 652                  * the segment to be fragmented was THE FIRST (otherwise,
 653                  * options are already fixed) and make it ONCE
 654                  * on the initial skb, so that all the following fragments
 655                  * will inherit fixed options.
 656                  */
 657                 if (offset == 0)
 658                         ip_options_fragment(skb);
 659
 660                 /*
 661                  *      Added AC : If we are fragmenting a fragment that's not the
 662                  *                 last fragment then keep MF on each bit
 663                  */
 664                 if (left > 0 || not_last_frag)
 665                         iph->frag_off |= htons(IP_MF);
 666                 ptr += len;
 667                 offset += len;
 668
 669                 /*
 670                  *      Put this fragment into the sending queue.
 671                  */
 672                 iph->tot_len = htons(len + hlen);
 673
 674                 ip_send_check(iph);
 675
 676                 err = output(skb2);
 677                 if (err)
 678                         goto fail;
 679
 680                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 681         }
 682         kfree_skb(skb);
 683         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 684         return err;
 685
 686 fail:
 687         kfree_skb(skb);
 688         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 689         return err;
 690 }
 691
 692 EXPORT_SYMBOL(ip_fragment);
 693
 694 int
 695 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 696 {
 697         struct iovec *iov = from;
 698
 699         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 700                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 701                         return -EFAULT;
 702         } else {
 703                 __wsum csum = 0;
 704                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 705                         return -EFAULT;
 706                 skb->csum = csum_block_add(skb->csum, csum, odd);
 707         }
 708         return 0;
 709 }
 710
 711 static inline __wsum
 712 csum_page(struct page *page, int offset, int copy)
 713 {
 714         char *kaddr;
 715         __wsum csum;
 716         kaddr = kmap(page);
 717         csum = csum_partial(kaddr + offset, copy, 0);
 718         kunmap(page);
 719         return csum;
 720 }
 721
 722 static inline int ip_ufo_append_data(struct sock *sk,
 723                         int getfrag(void *from, char *to, int offset, int len,
 724                                int odd, struct sk_buff *skb),
 725                         void *from, int length, int hh_len, int fragheaderlen,
 726                         int transhdrlen, int mtu, unsigned int flags)
 727 {
 728         struct sk_buff *skb;
 729         int err;
 730
 731         /* There is support for UDP fragmentation offload by network
 732          * device, so create one single skb packet containing complete
 733          * udp datagram
 734          */
 735         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
 736                 skb = sock_alloc_send_skb(sk,
 737                         hh_len + fragheaderlen + transhdrlen + 20,
 738                         (flags & MSG_DONTWAIT), &err);
 739
 740                 if (skb == NULL)
 741                         return err;
 742
 743                 /* reserve space for Hardware header */
 744                 skb_reserve(skb, hh_len);
 745
 746                 /* create space for UDP/IP header */
 747                 skb_put(skb, fragheaderlen + transhdrlen);
 748
 749                 /* initialize network header pointer */
 750                 skb_reset_network_header(skb);
 751
 752                 /* initialize protocol header pointer */
 753                 skb->transport_header = skb->network_header + fragheaderlen;
 754
 755                 skb->ip_summed = CHECKSUM_PARTIAL;
 756                 skb->csum = 0;
 757                 sk->sk_sndmsg_off = 0;
 758
 759                 /* specify the length of each IP datagram fragment */
 760                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
 761                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 762                 __skb_queue_tail(&sk->sk_write_queue, skb);
 763         }
 764
 765         return skb_append_datato_frags(sk, skb, getfrag, from,
 766                                        (length - transhdrlen));
 767 }
 768
 769 /*
 770  *      ip_append_data() and ip_append_page() can make one large IP datagram
 771  *      from many pieces of data. Each pieces will be holded on the socket
 772  *      until ip_push_pending_frames() is called. Each piece can be a page
 773  *      or non-page data.
 774  *
 775  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
 776  *      this interface potentially.
 777  *
 778  *      LATER: length must be adjusted by pad at tail, when it is required.
 779  */
 780 int ip_append_data(struct sock *sk,
 781                    int getfrag(void *from, char *to, int offset, int len,
 782                                int odd, struct sk_buff *skb),
 783                    void *from, int length, int transhdrlen,
 784                    struct ipcm_cookie *ipc, struct rtable **rtp,
 785                    unsigned int flags)
 786 {
 787         struct inet_sock *inet = inet_sk(sk);
 788         struct sk_buff *skb;
 789
 790         struct ip_options *opt = NULL;
 791         int hh_len;
 792         int exthdrlen;
 793         int mtu;
 794         int copy;
 795         int err;
 796         int offset = 0;
 797         unsigned int maxfraglen, fragheaderlen;
 798         int csummode = CHECKSUM_NONE;
 799         struct rtable *rt;
 800
 801         if (flags&MSG_PROBE)
 802                 return 0;
 803
 804         if (skb_queue_empty(&sk->sk_write_queue)) {
 805                 /*
 806                  * setup for corking.
 807                  */
 808                 opt = ipc->opt;
 809                 if (opt) {
 810                         if (inet->cork.opt == NULL) {
 811                                 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
 812                                 if (unlikely(inet->cork.opt == NULL))
 813                                         return -ENOBUFS;
 814                         }
 815                         memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
 816                         inet->cork.flags |= IPCORK_OPT;
 817                         inet->cork.addr = ipc->addr;
 818                 }
 819                 rt = *rtp;
 820                 if (unlikely(!rt))
 821                         return -EFAULT;
 822                 /*
 823                  * We steal reference to this route, caller should not release it
 824                  */
 825                 *rtp = NULL;
 826                 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
 827                                             rt->u.dst.dev->mtu :
 828                                             dst_mtu(rt->u.dst.path);
 829                 inet->cork.dst = &rt->u.dst;
 830                 inet->cork.length = 0;
 831                 sk->sk_sndmsg_page = NULL;
 832                 sk->sk_sndmsg_off = 0;
 833                 if ((exthdrlen = rt->u.dst.header_len) != 0) {
 834                         length += exthdrlen;
 835                         transhdrlen += exthdrlen;
 836                 }
 837         } else {
 838                 rt = (struct rtable *)inet->cork.dst;
 839                 if (inet->cork.flags & IPCORK_OPT)
 840                         opt = inet->cork.opt;
 841
 842                 transhdrlen = 0;
 843                 exthdrlen = 0;
 844                 mtu = inet->cork.fragsize;
 845         }
 846         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
 847
 848         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 849         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 850
 851         if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
 852                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
 853                                mtu-exthdrlen);
 854                 return -EMSGSIZE;
 855         }
 856
 857         /*
 858          * transhdrlen > 0 means that this is the first fragment and we wish
 859          * it won't be fragmented in the future.
 860          */
 861         if (transhdrlen &&
 862             length + fragheaderlen <= mtu &&
 863             rt->u.dst.dev->features & NETIF_F_V4_CSUM &&
 864             !exthdrlen)
 865                 csummode = CHECKSUM_PARTIAL;
 866
 867         inet->cork.length += length;
 868         if (((length> mtu) || !skb_queue_empty(&sk->sk_write_queue)) &&
 869             (sk->sk_protocol == IPPROTO_UDP) &&
 870             (rt->u.dst.dev->features & NETIF_F_UFO)) {
 871                 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
 872                                          fragheaderlen, transhdrlen, mtu,
 873                                          flags);
 874                 if (err)
 875                         goto error;
 876                 return 0;
 877         }
 878
 879         /* So, what's going on in the loop below?
 880          *
 881          * We use calculated fragment length to generate chained skb,
 882          * each of segments is IP fragment ready for sending to network after
 883          * adding appropriate IP header.
 884          */
 885
 886         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
 887                 goto alloc_new_skb;
 888
 889         while (length > 0) {
 890                 /* Check if the remaining data fits into current packet. */
 891                 copy = mtu - skb->len;
 892                 if (copy < length)
 893                         copy = maxfraglen - skb->len;
 894                 if (copy <= 0) {
 895                         char *data;
 896                         unsigned int datalen;
 897                         unsigned int fraglen;
 898                         unsigned int fraggap;
 899                         unsigned int alloclen;
 900                         struct sk_buff *skb_prev;
 901 alloc_new_skb:
 902                         skb_prev = skb;
 903                         if (skb_prev)
 904                                 fraggap = skb_prev->len - maxfraglen;
 905                         else
 906                                 fraggap = 0;
 907
 908                         /*
 909                          * If remaining data exceeds the mtu,
 910                          * we know we need more fragment(s).
 911                          */
 912                         datalen = length + fraggap;
 913                         if (datalen > mtu - fragheaderlen)
 914                                 datalen = maxfraglen - fragheaderlen;
 915                         fraglen = datalen + fragheaderlen;
 916
 917                         if ((flags & MSG_MORE) &&
 918                             !(rt->u.dst.dev->features&NETIF_F_SG))
 919                                 alloclen = mtu;
 920                         else
 921                                 alloclen = datalen + fragheaderlen;
 922
 923                         /* The last fragment gets additional space at tail.
 924                          * Note, with MSG_MORE we overallocate on fragments,
 925                          * because we have no idea what fragment will be
 926                          * the last.
 927                          */
 928                         if (datalen == length + fraggap)
 929                                 alloclen += rt->u.dst.trailer_len;
 930
 931                         if (transhdrlen) {
 932                                 skb = sock_alloc_send_skb(sk,
 933                                                 alloclen + hh_len + 15,
 934                                                 (flags & MSG_DONTWAIT), &err);
 935                         } else {
 936                                 skb = NULL;
 937                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 938                                     2 * sk->sk_sndbuf)
 939                                         skb = sock_wmalloc(sk,
 940                                                            alloclen + hh_len + 15, 1,
 941                                                            sk->sk_allocation);
 942                                 if (unlikely(skb == NULL))
 943                                         err = -ENOBUFS;
 944                                 else
 945                                         /* only the initial fragment is
 946                                            time stamped */
 947                                         ipc->shtx.flags = 0;
 948                         }
 949                         if (skb == NULL)
 950                                 goto error;
 951
 952                         /*
 953                          *      Fill in the control structures
 954                          */
 955                         skb->ip_summed = csummode;
 956                         skb->csum = 0;
 957                         skb_reserve(skb, hh_len);
 958                         *skb_tx(skb) = ipc->shtx;
 959
 960                         /*
 961                          *      Find where to start putting bytes.
 962                          */
 963                         data = skb_put(skb, fraglen);
 964                         skb_set_network_header(skb, exthdrlen);
 965                         skb->transport_header = (skb->network_header +
 966                                                  fragheaderlen);
 967                         data += fragheaderlen;
 968
 969                         if (fraggap) {
 970                                 skb->csum = skb_copy_and_csum_bits(
 971                                         skb_prev, maxfraglen,
 972                                         data + transhdrlen, fraggap, 0);
 973                                 skb_prev->csum = csum_sub(skb_prev->csum,
 974                                                           skb->csum);
 975                                 data += fraggap;
 976                                 pskb_trim_unique(skb_prev, maxfraglen);
 977                         }
 978
 979                         copy = datalen - transhdrlen - fraggap;
 980                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 981                                 err = -EFAULT;
 982                                 kfree_skb(skb);
 983                                 goto error;
 984                         }
 985
 986                         offset += copy;
 987                         length -= datalen - fraggap;
 988                         transhdrlen = 0;
 989                         exthdrlen = 0;
 990                         csummode = CHECKSUM_NONE;
 991
 992                         /*
 993                          * Put the packet on the pending queue.
 994                          */
 995                         __skb_queue_tail(&sk->sk_write_queue, skb);
 996                         continue;
 997                 }
 998
 999                 if (copy > length)
1000                         copy = length;
1001
1002                 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
1003                         unsigned int off;
1004
1005                         off = skb->len;
1006                         if (getfrag(from, skb_put(skb, copy),
1007                                         offset, copy, off, skb) < 0) {
1008                                 __skb_trim(skb, off);
1009                                 err = -EFAULT;
1010                                 goto error;
1011                         }
1012                 } else {
1013                         int i = skb_shinfo(skb)->nr_frags;
1014                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1015                         struct page *page = sk->sk_sndmsg_page;
1016                         int off = sk->sk_sndmsg_off;
1017                         unsigned int left;
1018
1019                         if (page && (left = PAGE_SIZE - off) > 0) {
1020                                 if (copy >= left)
1021                                         copy = left;
1022                                 if (page != frag->page) {
1023                                         if (i == MAX_SKB_FRAGS) {
1024                                                 err = -EMSGSIZE;
1025                                                 goto error;
1026                                         }
1027                                         get_page(page);
1028                                         skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1029                                         frag = &skb_shinfo(skb)->frags[i];
1030                                 }
1031                         } else if (i < MAX_SKB_FRAGS) {
1032                                 if (copy > PAGE_SIZE)
1033                                         copy = PAGE_SIZE;
1034                                 page = alloc_pages(sk->sk_allocation, 0);
1035                                 if (page == NULL)  {
1036                                         err = -ENOMEM;
1037                                         goto error;
1038                                 }
1039                                 sk->sk_sndmsg_page = page;
1040                                 sk->sk_sndmsg_off = 0;
1041
1042                                 skb_fill_page_desc(skb, i, page, 0, 0);
1043                                 frag = &skb_shinfo(skb)->frags[i];
1044                         } else {
1045                                 err = -EMSGSIZE;
1046                                 goto error;
1047                         }
1048                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1049                                 err = -EFAULT;
1050                                 goto error;
1051                         }
1052                         sk->sk_sndmsg_off += copy;
1053                         frag->size += copy;
1054                         skb->len += copy;
1055                         skb->data_len += copy;
1056                         skb->truesize += copy;
1057                         atomic_add(copy, &sk->sk_wmem_alloc);
1058                 }
1059                 offset += copy;
1060                 length -= copy;
1061         }
1062
1063         return 0;
1064
1065 error:
1066         inet->cork.length -= length;
1067         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1068         return err;
1069 }
1070
1071 ssize_t ip_append_page(struct sock *sk, struct page *page,
1072                        int offset, size_t size, int flags)
1073 {
1074         struct inet_sock *inet = inet_sk(sk);
1075         struct sk_buff *skb;
1076         struct rtable *rt;
1077         struct ip_options *opt = NULL;
1078         int hh_len;
1079         int mtu;
1080         int len;
1081         int err;
1082         unsigned int maxfraglen, fragheaderlen, fraggap;
1083
1084         if (inet->hdrincl)
1085                 return -EPERM;
1086
1087         if (flags&MSG_PROBE)
1088                 return 0;
1089
1090         if (skb_queue_empty(&sk->sk_write_queue))
1091                 return -EINVAL;
1092
1093         rt = (struct rtable *)inet->cork.dst;
1094         if (inet->cork.flags & IPCORK_OPT)
1095                 opt = inet->cork.opt;
1096
1097         if (!(rt->u.dst.dev->features&NETIF_F_SG))
1098                 return -EOPNOTSUPP;
1099
1100         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1101         mtu = inet->cork.fragsize;
1102
1103         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1104         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1105
1106         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1107                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1108                 return -EMSGSIZE;
1109         }
1110
1111         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1112                 return -EINVAL;
1113
1114         inet->cork.length += size;
1115         if ((sk->sk_protocol == IPPROTO_UDP) &&
1116             (rt->u.dst.dev->features & NETIF_F_UFO)) {
1117                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1118                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1119         }
1120
1121
1122         while (size > 0) {
1123                 int i;
1124
1125                 if (skb_is_gso(skb))
1126                         len = size;
1127                 else {
1128
1129                         /* Check if the remaining data fits into current packet. */
1130                         len = mtu - skb->len;
1131                         if (len < size)
1132                                 len = maxfraglen - skb->len;
1133                 }
1134                 if (len <= 0) {
1135                         struct sk_buff *skb_prev;
1136                         int alloclen;
1137
1138                         skb_prev = skb;
1139                         fraggap = skb_prev->len - maxfraglen;
1140
1141                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1142                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1143                         if (unlikely(!skb)) {
1144                                 err = -ENOBUFS;
1145                                 goto error;
1146                         }
1147
1148                         /*
1149                          *      Fill in the control structures
1150                          */
1151                         skb->ip_summed = CHECKSUM_NONE;
1152                         skb->csum = 0;
1153                         skb_reserve(skb, hh_len);
1154
1155                         /*
1156                          *      Find where to start putting bytes.
1157                          */
1158                         skb_put(skb, fragheaderlen + fraggap);
1159                         skb_reset_network_header(skb);
1160                         skb->transport_header = (skb->network_header +
1161                                                  fragheaderlen);
1162                         if (fraggap) {
1163                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1164                                                                    maxfraglen,
1165                                                     skb_transport_header(skb),
1166                                                                    fraggap, 0);
1167                                 skb_prev->csum = csum_sub(skb_prev->csum,
1168                                                           skb->csum);
1169                                 pskb_trim_unique(skb_prev, maxfraglen);
1170                         }
1171
1172                         /*
1173                          * Put the packet on the pending queue.
1174                          */
1175                         __skb_queue_tail(&sk->sk_write_queue, skb);
1176                         continue;
1177                 }
1178
1179                 i = skb_shinfo(skb)->nr_frags;
1180                 if (len > size)
1181                         len = size;
1182                 if (skb_can_coalesce(skb, i, page, offset)) {
1183                         skb_shinfo(skb)->frags[i-1].size += len;
1184                 } else if (i < MAX_SKB_FRAGS) {
1185                         get_page(page);
1186                         skb_fill_page_desc(skb, i, page, offset, len);
1187                 } else {
1188                         err = -EMSGSIZE;
1189                         goto error;
1190                 }
1191
1192                 if (skb->ip_summed == CHECKSUM_NONE) {
1193                         __wsum csum;
1194                         csum = csum_page(page, offset, len);
1195                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1196                 }
1197
1198                 skb->len += len;
1199                 skb->data_len += len;
1200                 skb->truesize += len;
1201                 atomic_add(len, &sk->sk_wmem_alloc);
1202                 offset += len;
1203                 size -= len;
1204         }
1205         return 0;
1206
1207 error:
1208         inet->cork.length -= size;
1209         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1210         return err;
1211 }
1212
1213 static void ip_cork_release(struct inet_sock *inet)
1214 {
1215         inet->cork.flags &= ~IPCORK_OPT;
1216         kfree(inet->cork.opt);
1217         inet->cork.opt = NULL;
1218         dst_release(inet->cork.dst);
1219         inet->cork.dst = NULL;
1220 }
1221
1222 /*
1223  *      Combined all pending IP fragments on the socket as one IP datagram
1224  *      and push them out.
1225  */
1226 int ip_push_pending_frames(struct sock *sk)
1227 {
1228         struct sk_buff *skb, *tmp_skb;
1229         struct sk_buff **tail_skb;
1230         struct inet_sock *inet = inet_sk(sk);
1231         struct net *net = sock_net(sk);
1232         struct ip_options *opt = NULL;
1233         struct rtable *rt = (struct rtable *)inet->cork.dst;
1234         struct iphdr *iph;
1235         __be16 df = 0;
1236         __u8 ttl;
1237         int err = 0;
1238
1239         if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1240                 goto out;
1241         tail_skb = &(skb_shinfo(skb)->frag_list);
1242
1243         /* move skb->data to ip header from ext header */
1244         if (skb->data < skb_network_header(skb))
1245                 __skb_pull(skb, skb_network_offset(skb));
1246         while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1247                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1248                 *tail_skb = tmp_skb;
1249                 tail_skb = &(tmp_skb->next);
1250                 skb->len += tmp_skb->len;
1251                 skb->data_len += tmp_skb->len;
1252                 skb->truesize += tmp_skb->truesize;
1253                 tmp_skb->destructor = NULL;
1254                 tmp_skb->sk = NULL;
1255         }
1256
1257         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1258          * to fragment the frame generated here. No matter, what transforms
1259          * how transforms change size of the packet, it will come out.
1260          */
1261         if (inet->pmtudisc < IP_PMTUDISC_DO)
1262                 skb->local_df = 1;
1263
1264         /* DF bit is set when we want to see DF on outgoing frames.
1265          * If local_df is set too, we still allow to fragment this frame
1266          * locally. */
1267         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1268             (skb->len <= dst_mtu(&rt->u.dst) &&
1269              ip_dont_fragment(sk, &rt->u.dst)))
1270                 df = htons(IP_DF);
1271
1272         if (inet->cork.flags & IPCORK_OPT)
1273                 opt = inet->cork.opt;
1274
1275         if (rt->rt_type == RTN_MULTICAST)
1276                 ttl = inet->mc_ttl;
1277         else
1278                 ttl = ip_select_ttl(inet, &rt->u.dst);
1279
1280         iph = (struct iphdr *)skb->data;
1281         iph->version = 4;
1282         iph->ihl = 5;
1283         if (opt) {
1284                 iph->ihl += opt->optlen>>2;
1285                 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1286         }
1287         iph->tos = inet->tos;
1288         iph->frag_off = df;
1289         ip_select_ident(iph, &rt->u.dst, sk);
1290         iph->ttl = ttl;
1291         iph->protocol = sk->sk_protocol;
1292         iph->saddr = rt->rt_src;
1293         iph->daddr = rt->rt_dst;
1294
1295         skb->priority = sk->sk_priority;
1296         skb->mark = sk->sk_mark;
1297         /*
1298          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1299          * on dst refcount
1300          */
1301         inet->cork.dst = NULL;
1302         skb_dst_set(skb, &rt->u.dst);
1303
1304         if (iph->protocol == IPPROTO_ICMP)
1305                 icmp_out_count(net, ((struct icmphdr *)
1306                         skb_transport_header(skb))->type);
1307
1308         /* Netfilter gets whole the not fragmented skb. */
1309         err = ip_local_out(skb);
1310         if (err) {
1311                 if (err > 0)
1312                         err = net_xmit_errno(err);
1313                 if (err)
1314                         goto error;
1315         }
1316
1317 out:
1318         ip_cork_release(inet);
1319         return err;
1320
1321 error:
1322         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1323         goto out;
1324 }
1325
1326 /*
1327  *      Throw away all pending data on the socket.
1328  */
1329 void ip_flush_pending_frames(struct sock *sk)
1330 {
1331         struct sk_buff *skb;
1332
1333         while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1334                 kfree_skb(skb);
1335
1336         ip_cork_release(inet_sk(sk));
1337 }
1338
1339
1340 /*
1341  *      Fetch data from kernel space and fill in checksum if needed.
1342  */
1343 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1344                               int len, int odd, struct sk_buff *skb)
1345 {
1346         __wsum csum;
1347
1348         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1349         skb->csum = csum_block_add(skb->csum, csum, odd);
1350         return 0;
1351 }
1352
1353 /*
1354  *      Generic function to send a packet as reply to another packet.
1355  *      Used to send TCP resets so far. ICMP should use this function too.
1356  *
1357  *      Should run single threaded per socket because it uses the sock
1358  *      structure to pass arguments.
1359  */
1360 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1361                    unsigned int len)
1362 {
1363         struct inet_sock *inet = inet_sk(sk);
1364         struct {
1365                 struct ip_options       opt;
1366                 char                    data[40];
1367         } replyopts;
1368         struct ipcm_cookie ipc;
1369         __be32 daddr;
1370         struct rtable *rt = skb_rtable(skb);
1371
1372         if (ip_options_echo(&replyopts.opt, skb))
1373                 return;
1374
1375         daddr = ipc.addr = rt->rt_src;
1376         ipc.opt = NULL;
1377         ipc.shtx.flags = 0;
1378
1379         if (replyopts.opt.optlen) {
1380                 ipc.opt = &replyopts.opt;
1381
1382                 if (ipc.opt->srr)
1383                         daddr = replyopts.opt.faddr;
1384         }
1385
1386         {
1387                 struct flowi fl = { .oif = arg->bound_dev_if,
1388                                     .nl_u = { .ip4_u =
1389                                               { .daddr = daddr,
1390                                                 .saddr = rt->rt_spec_dst,
1391                                                 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1392                                     /* Not quite clean, but right. */
1393                                     .uli_u = { .ports =
1394                                                { .sport = tcp_hdr(skb)->dest,
1395                                                  .dport = tcp_hdr(skb)->source } },
1396                                     .proto = sk->sk_protocol,
1397                                     .flags = ip_reply_arg_flowi_flags(arg) };
1398                 security_skb_classify_flow(skb, &fl);
1399                 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1400                         return;
1401         }
1402
1403         /* And let IP do all the hard work.
1404
1405            This chunk is not reenterable, hence spinlock.
1406            Note that it uses the fact, that this function is called
1407            with locally disabled BH and that sk cannot be already spinlocked.
1408          */
1409         bh_lock_sock(sk);
1410         inet->tos = ip_hdr(skb)->tos;
1411         sk->sk_priority = skb->priority;
1412         sk->sk_protocol = ip_hdr(skb)->protocol;
1413         sk->sk_bound_dev_if = arg->bound_dev_if;
1414         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1415                        &ipc, &rt, MSG_DONTWAIT);
1416         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1417                 if (arg->csumoffset >= 0)
1418                         *((__sum16 *)skb_transport_header(skb) +
1419                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1420                                                                 arg->csum));
1421                 skb->ip_summed = CHECKSUM_NONE;
1422                 ip_push_pending_frames(sk);
1423         }
1424
1425         bh_unlock_sock(sk);
1426
1427         ip_rt_put(rt);
1428 }
1429
1430 void __init ip_init(void)
1431 {
1432         ip_rt_init();
1433         inet_initpeers();
1434
1435 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1436         igmp_mc_proc_init();
1437 #endif
1438 }
1439
1440 EXPORT_SYMBOL(ip_generic_getfrag);
1441 EXPORT_SYMBOL(ip_queue_xmit);
1442 EXPORT_SYMBOL(ip_send_check);