net/ipv4/ip_output.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              The Internet Protocol (IP) output module.
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Donald Becker, <becker@super.org>
  11  *              Alan Cox, <Alan.Cox@linux.org>
  12  *              Richard Underwood
  13  *              Stefan Becker, <stefanb@yello.ping.de>
  14  *              Jorge Cwik, <jorge@laser.satlink.net>
  15  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  16  *              Hirokazu Takahashi, <taka@valinux.co.jp>
  17  *
  18  *      See ip_input.c for original log
  19  *
  20  *      Fixes:
  21  *              Alan Cox        :       Missing nonblock feature in ip_build_xmit.
  22  *              Mike Kilburn    :       htons() missing in ip_build_xmit.
  23  *              Bradford Johnson:       Fix faulty handling of some frames when
  24  *                                      no route is found.
  25  *              Alexander Demenshin:    Missing sk/skb free in ip_queue_xmit
  26  *                                      (in case if packet not accepted by
  27  *                                      output firewall rules)
  28  *              Mike McLagan    :       Routing by source
  29  *              Alexey Kuznetsov:       use new route cache
  30  *              Andi Kleen:             Fix broken PMTU recovery and remove
  31  *                                      some redundant tests.
  32  *      Vitaly E. Lavrov        :       Transparent proxy revived after year coma.
  33  *              Andi Kleen      :       Replace ip_reply with ip_send_reply.
  34  *              Andi Kleen      :       Split fast and slow ip_build_xmit path
  35  *                                      for decreased register pressure on x86
  36  *                                      and more readibility.
  37  *              Marc Boucher    :       When call_out_firewall returns FW_QUEUE,
  38  *                                      silently drop skb instead of failing with -EPERM.
  39  *              Detlev Wengorz  :       Copy protocol for fragments.
  40  *              Hirokazu Takahashi:     HW checksumming for outgoing UDP
  41  *                                      datagrams.
  42  *              Hirokazu Takahashi:     sendfile() on UDP works now.
  43  */
  44
  45 #include <asm/uaccess.h>
  46 #include <asm/system.h>
  47 #include <linux/module.h>
  48 #include <linux/types.h>
  49 #include <linux/kernel.h>
  50 #include <linux/mm.h>
  51 #include <linux/string.h>
  52 #include <linux/errno.h>
  53 #include <linux/highmem.h>
  54 #include <linux/slab.h>
  55
  56 #include <linux/socket.h>
  57 #include <linux/sockios.h>
  58 #include <linux/in.h>
  59 #include <linux/inet.h>
  60 #include <linux/netdevice.h>
  61 #include <linux/etherdevice.h>
  62 #include <linux/proc_fs.h>
  63 #include <linux/stat.h>
  64 #include <linux/init.h>
  65
  66 #include <net/snmp.h>
  67 #include <net/ip.h>
  68 #include <net/protocol.h>
  69 #include <net/route.h>
  70 #include <net/xfrm.h>
  71 #include <linux/skbuff.h>
  72 #include <net/sock.h>
  73 #include <net/arp.h>
  74 #include <net/icmp.h>
  75 #include <net/checksum.h>
  76 #include <net/inetpeer.h>
  77 #include <linux/igmp.h>
  78 #include <linux/netfilter_ipv4.h>
  79 #include <linux/netfilter_bridge.h>
  80 #include <linux/mroute.h>
  81 #include <linux/netlink.h>
  82 #include <linux/tcp.h>
  83
  84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  85 EXPORT_SYMBOL(sysctl_ip_default_ttl);
  86
  87 /* Generate a checksum for an outgoing IP datagram. */
  88 __inline__ void ip_send_check(struct iphdr *iph)
  89 {
  90         iph->check = 0;
  91         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  92 }
  93 EXPORT_SYMBOL(ip_send_check);
  94
  95 int __ip_local_out(struct sk_buff *skb)
  96 {
  97         struct iphdr *iph = ip_hdr(skb);
  98
  99         iph->tot_len = htons(skb->len);
 100         ip_send_check(iph);
 101         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 102                        skb_dst(skb)->dev, dst_output);
 103 }
 104
 105 int ip_local_out(struct sk_buff *skb)
 106 {
 107         int err;
 108
 109         err = __ip_local_out(skb);
 110         if (likely(err == 1))
 111                 err = dst_output(skb);
 112
 113         return err;
 114 }
 115 EXPORT_SYMBOL_GPL(ip_local_out);
 116
 117 /* dev_loopback_xmit for use with netfilter. */
 118 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
 119 {
 120         skb_reset_mac_header(newskb);
 121         __skb_pull(newskb, skb_network_offset(newskb));
 122         newskb->pkt_type = PACKET_LOOPBACK;
 123         newskb->ip_summed = CHECKSUM_UNNECESSARY;
 124         WARN_ON(!skb_dst(newskb));
 125         netif_rx_ni(newskb);
 126         return 0;
 127 }
 128
 129 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 130 {
 131         int ttl = inet->uc_ttl;
 132
 133         if (ttl < 0)
 134                 ttl = ip4_dst_hoplimit(dst);
 135         return ttl;
 136 }
 137
 138 /*
 139  *              Add an ip header to a skbuff and send it out.
 140  *
 141  */
 142 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 143                           __be32 saddr, __be32 daddr, struct ip_options *opt)
 144 {
 145         struct inet_sock *inet = inet_sk(sk);
 146         struct rtable *rt = skb_rtable(skb);
 147         struct iphdr *iph;
 148
 149         /* Build the IP header. */
 150         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 151         skb_reset_network_header(skb);
 152         iph = ip_hdr(skb);
 153         iph->version  = 4;
 154         iph->ihl      = 5;
 155         iph->tos      = inet->tos;
 156         if (ip_dont_fragment(sk, &rt->dst))
 157                 iph->frag_off = htons(IP_DF);
 158         else
 159                 iph->frag_off = 0;
 160         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 161         iph->daddr    = rt->rt_dst;
 162         iph->saddr    = rt->rt_src;
 163         iph->protocol = sk->sk_protocol;
 164         ip_select_ident(iph, &rt->dst, sk);
 165
 166         if (opt && opt->optlen) {
 167                 iph->ihl += opt->optlen>>2;
 168                 ip_options_build(skb, opt, daddr, rt, 0);
 169         }
 170
 171         skb->priority = sk->sk_priority;
 172         skb->mark = sk->sk_mark;
 173
 174         /* Send it out. */
 175         return ip_local_out(skb);
 176 }
 177 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 178
 179 static inline int ip_finish_output2(struct sk_buff *skb)
 180 {
 181         struct dst_entry *dst = skb_dst(skb);
 182         struct rtable *rt = (struct rtable *)dst;
 183         struct net_device *dev = dst->dev;
 184         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 185
 186         if (rt->rt_type == RTN_MULTICAST) {
 187                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 188         } else if (rt->rt_type == RTN_BROADCAST)
 189                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 190
 191         /* Be paranoid, rather than too clever. */
 192         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 193                 struct sk_buff *skb2;
 194
 195                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 196                 if (skb2 == NULL) {
 197                         kfree_skb(skb);
 198                         return -ENOMEM;
 199                 }
 200                 if (skb->sk)
 201                         skb_set_owner_w(skb2, skb->sk);
 202                 kfree_skb(skb);
 203                 skb = skb2;
 204         }
 205
 206         if (dst->hh)
 207                 return neigh_hh_output(dst->hh, skb);
 208         else if (dst->neighbour)
 209                 return dst->neighbour->output(skb);
 210
 211         if (net_ratelimit())
 212                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 213         kfree_skb(skb);
 214         return -EINVAL;
 215 }
 216
 217 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 218 {
 219         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 220
 221         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 222                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 223 }
 224
 225 static int ip_finish_output(struct sk_buff *skb)
 226 {
 227 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 228         /* Policy lookup after SNAT yielded a new policy */
 229         if (skb_dst(skb)->xfrm != NULL) {
 230                 IPCB(skb)->flags |= IPSKB_REROUTED;
 231                 return dst_output(skb);
 232         }
 233 #endif
 234         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 235                 return ip_fragment(skb, ip_finish_output2);
 236         else
 237                 return ip_finish_output2(skb);
 238 }
 239
 240 int ip_mc_output(struct sk_buff *skb)
 241 {
 242         struct sock *sk = skb->sk;
 243         struct rtable *rt = skb_rtable(skb);
 244         struct net_device *dev = rt->dst.dev;
 245
 246         /*
 247          *      If the indicated interface is up and running, send the packet.
 248          */
 249         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 250
 251         skb->dev = dev;
 252         skb->protocol = htons(ETH_P_IP);
 253
 254         /*
 255          *      Multicasts are looped back for other local users
 256          */
 257
 258         if (rt->rt_flags&RTCF_MULTICAST) {
 259                 if (sk_mc_loop(sk)
 260 #ifdef CONFIG_IP_MROUTE
 261                 /* Small optimization: do not loopback not local frames,
 262                    which returned after forwarding; they will be  dropped
 263                    by ip_mr_input in any case.
 264                    Note, that local frames are looped back to be delivered
 265                    to local recipients.
 266
 267                    This check is duplicated in ip_mr_input at the moment.
 268                  */
 269                     &&
 270                     ((rt->rt_flags & RTCF_LOCAL) ||
 271                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 272 #endif
 273                    ) {
 274                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 275                         if (newskb)
 276                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 277                                         newskb, NULL, newskb->dev,
 278                                         ip_dev_loopback_xmit);
 279                 }
 280
 281                 /* Multicasts with ttl 0 must not go beyond the host */
 282
 283                 if (ip_hdr(skb)->ttl == 0) {
 284                         kfree_skb(skb);
 285                         return 0;
 286                 }
 287         }
 288
 289         if (rt->rt_flags&RTCF_BROADCAST) {
 290                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 291                 if (newskb)
 292                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 293                                 NULL, newskb->dev, ip_dev_loopback_xmit);
 294         }
 295
 296         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 297                             skb->dev, ip_finish_output,
 298                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 299 }
 300
 301 int ip_output(struct sk_buff *skb)
 302 {
 303         struct net_device *dev = skb_dst(skb)->dev;
 304
 305         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 306
 307         skb->dev = dev;
 308         skb->protocol = htons(ETH_P_IP);
 309
 310         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 311                             ip_finish_output,
 312                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 313 }
 314
 315 int ip_queue_xmit(struct sk_buff *skb)
 316 {
 317         struct sock *sk = skb->sk;
 318         struct inet_sock *inet = inet_sk(sk);
 319         struct ip_options *opt = inet->opt;
 320         struct rtable *rt;
 321         struct iphdr *iph;
 322         int res;
 323
 324         /* Skip all of this if the packet is already routed,
 325          * f.e. by something like SCTP.
 326          */
 327         rcu_read_lock();
 328         rt = skb_rtable(skb);
 329         if (rt != NULL)
 330                 goto packet_routed;
 331
 332         /* Make sure we can route this packet. */
 333         rt = (struct rtable *)__sk_dst_check(sk, 0);
 334         if (rt == NULL) {
 335                 __be32 daddr;
 336
 337                 /* Use correct destination address if we have options. */
 338                 daddr = inet->inet_daddr;
 339                 if(opt && opt->srr)
 340                         daddr = opt->faddr;
 341
 342                 {
 343                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
 344                                             .mark = sk->sk_mark,
 345                                             .fl4_dst = daddr,
 346                                             .fl4_src = inet->inet_saddr,
 347                                             .fl4_tos = RT_CONN_FLAGS(sk),
 348                                             .proto = sk->sk_protocol,
 349                                             .flags = inet_sk_flowi_flags(sk),
 350                                             .fl_ip_sport = inet->inet_sport,
 351                                             .fl_ip_dport = inet->inet_dport };
 352
 353                         /* If this fails, retransmit mechanism of transport layer will
 354                          * keep trying until route appears or the connection times
 355                          * itself out.
 356                          */
 357                         security_sk_classify_flow(sk, &fl);
 358                         rt = ip_route_output_flow(sock_net(sk), &fl, sk);
 359                         if (IS_ERR(rt))
 360                                 goto no_route;
 361                 }
 362                 sk_setup_caps(sk, &rt->dst);
 363         }
 364         skb_dst_set_noref(skb, &rt->dst);
 365
 366 packet_routed:
 367         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
 368                 goto no_route;
 369
 370         /* OK, we know where to send it, allocate and build IP header. */
 371         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 372         skb_reset_network_header(skb);
 373         iph = ip_hdr(skb);
 374         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 375         if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
 376                 iph->frag_off = htons(IP_DF);
 377         else
 378                 iph->frag_off = 0;
 379         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 380         iph->protocol = sk->sk_protocol;
 381         iph->saddr    = rt->rt_src;
 382         iph->daddr    = rt->rt_dst;
 383         /* Transport layer set skb->h.foo itself. */
 384
 385         if (opt && opt->optlen) {
 386                 iph->ihl += opt->optlen >> 2;
 387                 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
 388         }
 389
 390         ip_select_ident_more(iph, &rt->dst, sk,
 391                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 392
 393         skb->priority = sk->sk_priority;
 394         skb->mark = sk->sk_mark;
 395
 396         res = ip_local_out(skb);
 397         rcu_read_unlock();
 398         return res;
 399
 400 no_route:
 401         rcu_read_unlock();
 402         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 403         kfree_skb(skb);
 404         return -EHOSTUNREACH;
 405 }
 406 EXPORT_SYMBOL(ip_queue_xmit);
 407
 408
 409 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 410 {
 411         to->pkt_type = from->pkt_type;
 412         to->priority = from->priority;
 413         to->protocol = from->protocol;
 414         skb_dst_drop(to);
 415         skb_dst_copy(to, from);
 416         to->dev = from->dev;
 417         to->mark = from->mark;
 418
 419         /* Copy the flags to each fragment. */
 420         IPCB(to)->flags = IPCB(from)->flags;
 421
 422 #ifdef CONFIG_NET_SCHED
 423         to->tc_index = from->tc_index;
 424 #endif
 425         nf_copy(to, from);
 426 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 427     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 428         to->nf_trace = from->nf_trace;
 429 #endif
 430 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 431         to->ipvs_property = from->ipvs_property;
 432 #endif
 433         skb_copy_secmark(to, from);
 434 }
 435
 436 /*
 437  *      This IP datagram is too large to be sent in one piece.  Break it up into
 438  *      smaller pieces (each of size equal to IP header plus
 439  *      a block of the data of the original IP data part) that will yet fit in a
 440  *      single device frame, and queue such a frame for sending.
 441  */
 442
 443 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 444 {
 445         struct iphdr *iph;
 446         int ptr;
 447         struct net_device *dev;
 448         struct sk_buff *skb2;
 449         unsigned int mtu, hlen, left, len, ll_rs;
 450         int offset;
 451         __be16 not_last_frag;
 452         struct rtable *rt = skb_rtable(skb);
 453         int err = 0;
 454
 455         dev = rt->dst.dev;
 456
 457         /*
 458          *      Point into the IP datagram header.
 459          */
 460
 461         iph = ip_hdr(skb);
 462
 463         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 464                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 465                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 466                           htonl(ip_skb_dst_mtu(skb)));
 467                 kfree_skb(skb);
 468                 return -EMSGSIZE;
 469         }
 470
 471         /*
 472          *      Setup starting values.
 473          */
 474
 475         hlen = iph->ihl * 4;
 476         mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
 477 #ifdef CONFIG_BRIDGE_NETFILTER
 478         if (skb->nf_bridge)
 479                 mtu -= nf_bridge_mtu_reduction(skb);
 480 #endif
 481         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 482
 483         /* When frag_list is given, use it. First, check its validity:
 484          * some transformers could create wrong frag_list or break existing
 485          * one, it is not prohibited. In this case fall back to copying.
 486          *
 487          * LATER: this step can be merged to real generation of fragments,
 488          * we can switch to copy when see the first bad fragment.
 489          */
 490         if (skb_has_frag_list(skb)) {
 491                 struct sk_buff *frag, *frag2;
 492                 int first_len = skb_pagelen(skb);
 493
 494                 if (first_len - hlen > mtu ||
 495                     ((first_len - hlen) & 7) ||
 496                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
 497                     skb_cloned(skb))
 498                         goto slow_path;
 499
 500                 skb_walk_frags(skb, frag) {
 501                         /* Correct geometry. */
 502                         if (frag->len > mtu ||
 503                             ((frag->len & 7) && frag->next) ||
 504                             skb_headroom(frag) < hlen)
 505                                 goto slow_path_clean;
 506
 507                         /* Partially cloned skb? */
 508                         if (skb_shared(frag))
 509                                 goto slow_path_clean;
 510
 511                         BUG_ON(frag->sk);
 512                         if (skb->sk) {
 513                                 frag->sk = skb->sk;
 514                                 frag->destructor = sock_wfree;
 515                         }
 516                         skb->truesize -= frag->truesize;
 517                 }
 518
 519                 /* Everything is OK. Generate! */
 520
 521                 err = 0;
 522                 offset = 0;
 523                 frag = skb_shinfo(skb)->frag_list;
 524                 skb_frag_list_init(skb);
 525                 skb->data_len = first_len - skb_headlen(skb);
 526                 skb->len = first_len;
 527                 iph->tot_len = htons(first_len);
 528                 iph->frag_off = htons(IP_MF);
 529                 ip_send_check(iph);
 530
 531                 for (;;) {
 532                         /* Prepare header of the next frame,
 533                          * before previous one went down. */
 534                         if (frag) {
 535                                 frag->ip_summed = CHECKSUM_NONE;
 536                                 skb_reset_transport_header(frag);
 537                                 __skb_push(frag, hlen);
 538                                 skb_reset_network_header(frag);
 539                                 memcpy(skb_network_header(frag), iph, hlen);
 540                                 iph = ip_hdr(frag);
 541                                 iph->tot_len = htons(frag->len);
 542                                 ip_copy_metadata(frag, skb);
 543                                 if (offset == 0)
 544                                         ip_options_fragment(frag);
 545                                 offset += skb->len - hlen;
 546                                 iph->frag_off = htons(offset>>3);
 547                                 if (frag->next != NULL)
 548                                         iph->frag_off |= htons(IP_MF);
 549                                 /* Ready, complete checksum */
 550                                 ip_send_check(iph);
 551                         }
 552
 553                         err = output(skb);
 554
 555                         if (!err)
 556                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 557                         if (err || !frag)
 558                                 break;
 559
 560                         skb = frag;
 561                         frag = skb->next;
 562                         skb->next = NULL;
 563                 }
 564
 565                 if (err == 0) {
 566                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 567                         return 0;
 568                 }
 569
 570                 while (frag) {
 571                         skb = frag->next;
 572                         kfree_skb(frag);
 573                         frag = skb;
 574                 }
 575                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 576                 return err;
 577
 578 slow_path_clean:
 579                 skb_walk_frags(skb, frag2) {
 580                         if (frag2 == frag)
 581                                 break;
 582                         frag2->sk = NULL;
 583                         frag2->destructor = NULL;
 584                         skb->truesize += frag2->truesize;
 585                 }
 586         }
 587
 588 slow_path:
 589         left = skb->len - hlen;         /* Space per frame */
 590         ptr = hlen;             /* Where to start from */
 591
 592         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 593          * we need to make room for the encapsulating header
 594          */
 595         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 596
 597         /*
 598          *      Fragment the datagram.
 599          */
 600
 601         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 602         not_last_frag = iph->frag_off & htons(IP_MF);
 603
 604         /*
 605          *      Keep copying data until we run out.
 606          */
 607
 608         while (left > 0) {
 609                 len = left;
 610                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 611                 if (len > mtu)
 612                         len = mtu;
 613                 /* IF: we are not sending upto and including the packet end
 614                    then align the next start on an eight byte boundary */
 615                 if (len < left) {
 616                         len &= ~7;
 617                 }
 618                 /*
 619                  *      Allocate buffer.
 620                  */
 621
 622                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 623                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 624                         err = -ENOMEM;
 625                         goto fail;
 626                 }
 627
 628                 /*
 629                  *      Set up data on packet
 630                  */
 631
 632                 ip_copy_metadata(skb2, skb);
 633                 skb_reserve(skb2, ll_rs);
 634                 skb_put(skb2, len + hlen);
 635                 skb_reset_network_header(skb2);
 636                 skb2->transport_header = skb2->network_header + hlen;
 637
 638                 /*
 639                  *      Charge the memory for the fragment to any owner
 640                  *      it might possess
 641                  */
 642
 643                 if (skb->sk)
 644                         skb_set_owner_w(skb2, skb->sk);
 645
 646                 /*
 647                  *      Copy the packet header into the new buffer.
 648                  */
 649
 650                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 651
 652                 /*
 653                  *      Copy a block of the IP datagram.
 654                  */
 655                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 656                         BUG();
 657                 left -= len;
 658
 659                 /*
 660                  *      Fill in the new header fields.
 661                  */
 662                 iph = ip_hdr(skb2);
 663                 iph->frag_off = htons((offset >> 3));
 664
 665                 /* ANK: dirty, but effective trick. Upgrade options only if
 666                  * the segment to be fragmented was THE FIRST (otherwise,
 667                  * options are already fixed) and make it ONCE
 668                  * on the initial skb, so that all the following fragments
 669                  * will inherit fixed options.
 670                  */
 671                 if (offset == 0)
 672                         ip_options_fragment(skb);
 673
 674                 /*
 675                  *      Added AC : If we are fragmenting a fragment that's not the
 676                  *                 last fragment then keep MF on each bit
 677                  */
 678                 if (left > 0 || not_last_frag)
 679                         iph->frag_off |= htons(IP_MF);
 680                 ptr += len;
 681                 offset += len;
 682
 683                 /*
 684                  *      Put this fragment into the sending queue.
 685                  */
 686                 iph->tot_len = htons(len + hlen);
 687
 688                 ip_send_check(iph);
 689
 690                 err = output(skb2);
 691                 if (err)
 692                         goto fail;
 693
 694                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 695         }
 696         kfree_skb(skb);
 697         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 698         return err;
 699
 700 fail:
 701         kfree_skb(skb);
 702         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 703         return err;
 704 }
 705 EXPORT_SYMBOL(ip_fragment);
 706
 707 int
 708 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 709 {
 710         struct iovec *iov = from;
 711
 712         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 713                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 714                         return -EFAULT;
 715         } else {
 716                 __wsum csum = 0;
 717                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 718                         return -EFAULT;
 719                 skb->csum = csum_block_add(skb->csum, csum, odd);
 720         }
 721         return 0;
 722 }
 723 EXPORT_SYMBOL(ip_generic_getfrag);
 724
 725 static inline __wsum
 726 csum_page(struct page *page, int offset, int copy)
 727 {
 728         char *kaddr;
 729         __wsum csum;
 730         kaddr = kmap(page);
 731         csum = csum_partial(kaddr + offset, copy, 0);
 732         kunmap(page);
 733         return csum;
 734 }
 735
 736 static inline int ip_ufo_append_data(struct sock *sk,
 737                         struct sk_buff_head *queue,
 738                         int getfrag(void *from, char *to, int offset, int len,
 739                                int odd, struct sk_buff *skb),
 740                         void *from, int length, int hh_len, int fragheaderlen,
 741                         int transhdrlen, int mtu, unsigned int flags)
 742 {
 743         struct sk_buff *skb;
 744         int err;
 745
 746         /* There is support for UDP fragmentation offload by network
 747          * device, so create one single skb packet containing complete
 748          * udp datagram
 749          */
 750         if ((skb = skb_peek_tail(queue)) == NULL) {
 751                 skb = sock_alloc_send_skb(sk,
 752                         hh_len + fragheaderlen + transhdrlen + 20,
 753                         (flags & MSG_DONTWAIT), &err);
 754
 755                 if (skb == NULL)
 756                         return err;
 757
 758                 /* reserve space for Hardware header */
 759                 skb_reserve(skb, hh_len);
 760
 761                 /* create space for UDP/IP header */
 762                 skb_put(skb, fragheaderlen + transhdrlen);
 763
 764                 /* initialize network header pointer */
 765                 skb_reset_network_header(skb);
 766
 767                 /* initialize protocol header pointer */
 768                 skb->transport_header = skb->network_header + fragheaderlen;
 769
 770                 skb->ip_summed = CHECKSUM_PARTIAL;
 771                 skb->csum = 0;
 772
 773                 /* specify the length of each IP datagram fragment */
 774                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
 775                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 776                 __skb_queue_tail(queue, skb);
 777         }
 778
 779         return skb_append_datato_frags(sk, skb, getfrag, from,
 780                                        (length - transhdrlen));
 781 }
 782
 783 static int __ip_append_data(struct sock *sk, struct sk_buff_head *queue,
 784                             struct inet_cork *cork,
 785                             int getfrag(void *from, char *to, int offset,
 786                                         int len, int odd, struct sk_buff *skb),
 787                             void *from, int length, int transhdrlen,
 788                             unsigned int flags)
 789 {
 790         struct inet_sock *inet = inet_sk(sk);
 791         struct sk_buff *skb;
 792
 793         struct ip_options *opt = cork->opt;
 794         int hh_len;
 795         int exthdrlen;
 796         int mtu;
 797         int copy;
 798         int err;
 799         int offset = 0;
 800         unsigned int maxfraglen, fragheaderlen;
 801         int csummode = CHECKSUM_NONE;
 802         struct rtable *rt = (struct rtable *)cork->dst;
 803
 804         exthdrlen = transhdrlen ? rt->dst.header_len : 0;
 805         length += exthdrlen;
 806         transhdrlen += exthdrlen;
 807         mtu = cork->fragsize;
 808
 809         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 810
 811         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 812         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 813
 814         if (cork->length + length > 0xFFFF - fragheaderlen) {
 815                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
 816                                mtu-exthdrlen);
 817                 return -EMSGSIZE;
 818         }
 819
 820         /*
 821          * transhdrlen > 0 means that this is the first fragment and we wish
 822          * it won't be fragmented in the future.
 823          */
 824         if (transhdrlen &&
 825             length + fragheaderlen <= mtu &&
 826             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 827             !exthdrlen)
 828                 csummode = CHECKSUM_PARTIAL;
 829
 830         skb = skb_peek_tail(queue);
 831
 832         cork->length += length;
 833         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 834             (sk->sk_protocol == IPPROTO_UDP) &&
 835             (rt->dst.dev->features & NETIF_F_UFO)) {
 836                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 837                                          hh_len, fragheaderlen, transhdrlen,
 838                                          mtu, flags);
 839                 if (err)
 840                         goto error;
 841                 return 0;
 842         }
 843
 844         /* So, what's going on in the loop below?
 845          *
 846          * We use calculated fragment length to generate chained skb,
 847          * each of segments is IP fragment ready for sending to network after
 848          * adding appropriate IP header.
 849          */
 850
 851         if (!skb)
 852                 goto alloc_new_skb;
 853
 854         while (length > 0) {
 855                 /* Check if the remaining data fits into current packet. */
 856                 copy = mtu - skb->len;
 857                 if (copy < length)
 858                         copy = maxfraglen - skb->len;
 859                 if (copy <= 0) {
 860                         char *data;
 861                         unsigned int datalen;
 862                         unsigned int fraglen;
 863                         unsigned int fraggap;
 864                         unsigned int alloclen;
 865                         struct sk_buff *skb_prev;
 866 alloc_new_skb:
 867                         skb_prev = skb;
 868                         if (skb_prev)
 869                                 fraggap = skb_prev->len - maxfraglen;
 870                         else
 871                                 fraggap = 0;
 872
 873                         /*
 874                          * If remaining data exceeds the mtu,
 875                          * we know we need more fragment(s).
 876                          */
 877                         datalen = length + fraggap;
 878                         if (datalen > mtu - fragheaderlen)
 879                                 datalen = maxfraglen - fragheaderlen;
 880                         fraglen = datalen + fragheaderlen;
 881
 882                         if ((flags & MSG_MORE) &&
 883                             !(rt->dst.dev->features&NETIF_F_SG))
 884                                 alloclen = mtu;
 885                         else
 886                                 alloclen = fraglen;
 887
 888                         /* The last fragment gets additional space at tail.
 889                          * Note, with MSG_MORE we overallocate on fragments,
 890                          * because we have no idea what fragment will be
 891                          * the last.
 892                          */
 893                         if (datalen == length + fraggap) {
 894                                 alloclen += rt->dst.trailer_len;
 895                                 /* make sure mtu is not reached */
 896                                 if (datalen > mtu - fragheaderlen - rt->dst.trailer_len)
 897                                         datalen -= ALIGN(rt->dst.trailer_len, 8);
 898                         }
 899                         if (transhdrlen) {
 900                                 skb = sock_alloc_send_skb(sk,
 901                                                 alloclen + hh_len + 15,
 902                                                 (flags & MSG_DONTWAIT), &err);
 903                         } else {
 904                                 skb = NULL;
 905                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 906                                     2 * sk->sk_sndbuf)
 907                                         skb = sock_wmalloc(sk,
 908                                                            alloclen + hh_len + 15, 1,
 909                                                            sk->sk_allocation);
 910                                 if (unlikely(skb == NULL))
 911                                         err = -ENOBUFS;
 912                                 else
 913                                         /* only the initial fragment is
 914                                            time stamped */
 915                                         cork->tx_flags = 0;
 916                         }
 917                         if (skb == NULL)
 918                                 goto error;
 919
 920                         /*
 921                          *      Fill in the control structures
 922                          */
 923                         skb->ip_summed = csummode;
 924                         skb->csum = 0;
 925                         skb_reserve(skb, hh_len);
 926                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 927
 928                         /*
 929                          *      Find where to start putting bytes.
 930                          */
 931                         data = skb_put(skb, fraglen);
 932                         skb_set_network_header(skb, exthdrlen);
 933                         skb->transport_header = (skb->network_header +
 934                                                  fragheaderlen);
 935                         data += fragheaderlen;
 936
 937                         if (fraggap) {
 938                                 skb->csum = skb_copy_and_csum_bits(
 939                                         skb_prev, maxfraglen,
 940                                         data + transhdrlen, fraggap, 0);
 941                                 skb_prev->csum = csum_sub(skb_prev->csum,
 942                                                           skb->csum);
 943                                 data += fraggap;
 944                                 pskb_trim_unique(skb_prev, maxfraglen);
 945                         }
 946
 947                         copy = datalen - transhdrlen - fraggap;
 948                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 949                                 err = -EFAULT;
 950                                 kfree_skb(skb);
 951                                 goto error;
 952                         }
 953
 954                         offset += copy;
 955                         length -= datalen - fraggap;
 956                         transhdrlen = 0;
 957                         exthdrlen = 0;
 958                         csummode = CHECKSUM_NONE;
 959
 960                         /*
 961                          * Put the packet on the pending queue.
 962                          */
 963                         __skb_queue_tail(queue, skb);
 964                         continue;
 965                 }
 966
 967                 if (copy > length)
 968                         copy = length;
 969
 970                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
 971                         unsigned int off;
 972
 973                         off = skb->len;
 974                         if (getfrag(from, skb_put(skb, copy),
 975                                         offset, copy, off, skb) < 0) {
 976                                 __skb_trim(skb, off);
 977                                 err = -EFAULT;
 978                                 goto error;
 979                         }
 980                 } else {
 981                         int i = skb_shinfo(skb)->nr_frags;
 982                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
 983                         struct page *page = cork->page;
 984                         int off = cork->off;
 985                         unsigned int left;
 986
 987                         if (page && (left = PAGE_SIZE - off) > 0) {
 988                                 if (copy >= left)
 989                                         copy = left;
 990                                 if (page != frag->page) {
 991                                         if (i == MAX_SKB_FRAGS) {
 992                                                 err = -EMSGSIZE;
 993                                                 goto error;
 994                                         }
 995                                         get_page(page);
 996                                         skb_fill_page_desc(skb, i, page, off, 0);
 997                                         frag = &skb_shinfo(skb)->frags[i];
 998                                 }
 999                         } else if (i < MAX_SKB_FRAGS) {
1000                                 if (copy > PAGE_SIZE)
1001                                         copy = PAGE_SIZE;
1002                                 page = alloc_pages(sk->sk_allocation, 0);
1003                                 if (page == NULL)  {
1004                                         err = -ENOMEM;
1005                                         goto error;
1006                                 }
1007                                 cork->page = page;
1008                                 cork->off = 0;
1009
1010                                 skb_fill_page_desc(skb, i, page, 0, 0);
1011                                 frag = &skb_shinfo(skb)->frags[i];
1012                         } else {
1013                                 err = -EMSGSIZE;
1014                                 goto error;
1015                         }
1016                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1017                                 err = -EFAULT;
1018                                 goto error;
1019                         }
1020                         cork->off += copy;
1021                         frag->size += copy;
1022                         skb->len += copy;
1023                         skb->data_len += copy;
1024                         skb->truesize += copy;
1025                         atomic_add(copy, &sk->sk_wmem_alloc);
1026                 }
1027                 offset += copy;
1028                 length -= copy;
1029         }
1030
1031         return 0;
1032
1033 error:
1034         cork->length -= length;
1035         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1036         return err;
1037 }
1038
1039 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1040                          struct ipcm_cookie *ipc, struct rtable **rtp)
1041 {
1042         struct inet_sock *inet = inet_sk(sk);
1043         struct ip_options *opt;
1044         struct rtable *rt;
1045
1046         /*
1047          * setup for corking.
1048          */
1049         opt = ipc->opt;
1050         if (opt) {
1051                 if (cork->opt == NULL) {
1052                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1053                                             sk->sk_allocation);
1054                         if (unlikely(cork->opt == NULL))
1055                                 return -ENOBUFS;
1056                 }
1057                 memcpy(cork->opt, opt, sizeof(struct ip_options) + opt->optlen);
1058                 cork->flags |= IPCORK_OPT;
1059                 cork->addr = ipc->addr;
1060         }
1061         rt = *rtp;
1062         if (unlikely(!rt))
1063                 return -EFAULT;
1064         /*
1065          * We steal reference to this route, caller should not release it
1066          */
1067         *rtp = NULL;
1068         cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1069                          rt->dst.dev->mtu : dst_mtu(rt->dst.path);
1070         cork->dst = &rt->dst;
1071         cork->length = 0;
1072         cork->tx_flags = ipc->tx_flags;
1073         cork->page = NULL;
1074         cork->off = 0;
1075
1076         return 0;
1077 }
1078
1079 /*
1080  *      ip_append_data() and ip_append_page() can make one large IP datagram
1081  *      from many pieces of data. Each pieces will be holded on the socket
1082  *      until ip_push_pending_frames() is called. Each piece can be a page
1083  *      or non-page data.
1084  *
1085  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1086  *      this interface potentially.
1087  *
1088  *      LATER: length must be adjusted by pad at tail, when it is required.
1089  */
1090 int ip_append_data(struct sock *sk,
1091                    int getfrag(void *from, char *to, int offset, int len,
1092                                int odd, struct sk_buff *skb),
1093                    void *from, int length, int transhdrlen,
1094                    struct ipcm_cookie *ipc, struct rtable **rtp,
1095                    unsigned int flags)
1096 {
1097         struct inet_sock *inet = inet_sk(sk);
1098         int err;
1099
1100         if (flags&MSG_PROBE)
1101                 return 0;
1102
1103         if (skb_queue_empty(&sk->sk_write_queue)) {
1104                 err = ip_setup_cork(sk, &inet->cork, ipc, rtp);
1105                 if (err)
1106                         return err;
1107         } else {
1108                 transhdrlen = 0;
1109         }
1110
1111         return __ip_append_data(sk, &sk->sk_write_queue, &inet->cork, getfrag,
1112                                 from, length, transhdrlen, flags);
1113 }
1114
1115 ssize_t ip_append_page(struct sock *sk, struct page *page,
1116                        int offset, size_t size, int flags)
1117 {
1118         struct inet_sock *inet = inet_sk(sk);
1119         struct sk_buff *skb;
1120         struct rtable *rt;
1121         struct ip_options *opt = NULL;
1122         int hh_len;
1123         int mtu;
1124         int len;
1125         int err;
1126         unsigned int maxfraglen, fragheaderlen, fraggap;
1127
1128         if (inet->hdrincl)
1129                 return -EPERM;
1130
1131         if (flags&MSG_PROBE)
1132                 return 0;
1133
1134         if (skb_queue_empty(&sk->sk_write_queue))
1135                 return -EINVAL;
1136
1137         rt = (struct rtable *)inet->cork.dst;
1138         if (inet->cork.flags & IPCORK_OPT)
1139                 opt = inet->cork.opt;
1140
1141         if (!(rt->dst.dev->features&NETIF_F_SG))
1142                 return -EOPNOTSUPP;
1143
1144         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1145         mtu = inet->cork.fragsize;
1146
1147         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1148         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1149
1150         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1151                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1152                 return -EMSGSIZE;
1153         }
1154
1155         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1156                 return -EINVAL;
1157
1158         inet->cork.length += size;
1159         if ((size + skb->len > mtu) &&
1160             (sk->sk_protocol == IPPROTO_UDP) &&
1161             (rt->dst.dev->features & NETIF_F_UFO)) {
1162                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1163                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1164         }
1165
1166
1167         while (size > 0) {
1168                 int i;
1169
1170                 if (skb_is_gso(skb))
1171                         len = size;
1172                 else {
1173
1174                         /* Check if the remaining data fits into current packet. */
1175                         len = mtu - skb->len;
1176                         if (len < size)
1177                                 len = maxfraglen - skb->len;
1178                 }
1179                 if (len <= 0) {
1180                         struct sk_buff *skb_prev;
1181                         int alloclen;
1182
1183                         skb_prev = skb;
1184                         fraggap = skb_prev->len - maxfraglen;
1185
1186                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1187                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1188                         if (unlikely(!skb)) {
1189                                 err = -ENOBUFS;
1190                                 goto error;
1191                         }
1192
1193                         /*
1194                          *      Fill in the control structures
1195                          */
1196                         skb->ip_summed = CHECKSUM_NONE;
1197                         skb->csum = 0;
1198                         skb_reserve(skb, hh_len);
1199
1200                         /*
1201                          *      Find where to start putting bytes.
1202                          */
1203                         skb_put(skb, fragheaderlen + fraggap);
1204                         skb_reset_network_header(skb);
1205                         skb->transport_header = (skb->network_header +
1206                                                  fragheaderlen);
1207                         if (fraggap) {
1208                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1209                                                                    maxfraglen,
1210                                                     skb_transport_header(skb),
1211                                                                    fraggap, 0);
1212                                 skb_prev->csum = csum_sub(skb_prev->csum,
1213                                                           skb->csum);
1214                                 pskb_trim_unique(skb_prev, maxfraglen);
1215                         }
1216
1217                         /*
1218                          * Put the packet on the pending queue.
1219                          */
1220                         __skb_queue_tail(&sk->sk_write_queue, skb);
1221                         continue;
1222                 }
1223
1224                 i = skb_shinfo(skb)->nr_frags;
1225                 if (len > size)
1226                         len = size;
1227                 if (skb_can_coalesce(skb, i, page, offset)) {
1228                         skb_shinfo(skb)->frags[i-1].size += len;
1229                 } else if (i < MAX_SKB_FRAGS) {
1230                         get_page(page);
1231                         skb_fill_page_desc(skb, i, page, offset, len);
1232                 } else {
1233                         err = -EMSGSIZE;
1234                         goto error;
1235                 }
1236
1237                 if (skb->ip_summed == CHECKSUM_NONE) {
1238                         __wsum csum;
1239                         csum = csum_page(page, offset, len);
1240                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1241                 }
1242
1243                 skb->len += len;
1244                 skb->data_len += len;
1245                 skb->truesize += len;
1246                 atomic_add(len, &sk->sk_wmem_alloc);
1247                 offset += len;
1248                 size -= len;
1249         }
1250         return 0;
1251
1252 error:
1253         inet->cork.length -= size;
1254         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1255         return err;
1256 }
1257
1258 static void ip_cork_release(struct inet_cork *cork)
1259 {
1260         cork->flags &= ~IPCORK_OPT;
1261         kfree(cork->opt);
1262         cork->opt = NULL;
1263         dst_release(cork->dst);
1264         cork->dst = NULL;
1265 }
1266
1267 /*
1268  *      Combined all pending IP fragments on the socket as one IP datagram
1269  *      and push them out.
1270  */
1271 struct sk_buff *__ip_make_skb(struct sock *sk,
1272                               struct sk_buff_head *queue,
1273                               struct inet_cork *cork)
1274 {
1275         struct sk_buff *skb, *tmp_skb;
1276         struct sk_buff **tail_skb;
1277         struct inet_sock *inet = inet_sk(sk);
1278         struct net *net = sock_net(sk);
1279         struct ip_options *opt = NULL;
1280         struct rtable *rt = (struct rtable *)cork->dst;
1281         struct iphdr *iph;
1282         __be16 df = 0;
1283         __u8 ttl;
1284
1285         if ((skb = __skb_dequeue(queue)) == NULL)
1286                 goto out;
1287         tail_skb = &(skb_shinfo(skb)->frag_list);
1288
1289         /* move skb->data to ip header from ext header */
1290         if (skb->data < skb_network_header(skb))
1291                 __skb_pull(skb, skb_network_offset(skb));
1292         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1293                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1294                 *tail_skb = tmp_skb;
1295                 tail_skb = &(tmp_skb->next);
1296                 skb->len += tmp_skb->len;
1297                 skb->data_len += tmp_skb->len;
1298                 skb->truesize += tmp_skb->truesize;
1299                 tmp_skb->destructor = NULL;
1300                 tmp_skb->sk = NULL;
1301         }
1302
1303         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1304          * to fragment the frame generated here. No matter, what transforms
1305          * how transforms change size of the packet, it will come out.
1306          */
1307         if (inet->pmtudisc < IP_PMTUDISC_DO)
1308                 skb->local_df = 1;
1309
1310         /* DF bit is set when we want to see DF on outgoing frames.
1311          * If local_df is set too, we still allow to fragment this frame
1312          * locally. */
1313         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1314             (skb->len <= dst_mtu(&rt->dst) &&
1315              ip_dont_fragment(sk, &rt->dst)))
1316                 df = htons(IP_DF);
1317
1318         if (cork->flags & IPCORK_OPT)
1319                 opt = cork->opt;
1320
1321         if (rt->rt_type == RTN_MULTICAST)
1322                 ttl = inet->mc_ttl;
1323         else
1324                 ttl = ip_select_ttl(inet, &rt->dst);
1325
1326         iph = (struct iphdr *)skb->data;
1327         iph->version = 4;
1328         iph->ihl = 5;
1329         if (opt) {
1330                 iph->ihl += opt->optlen>>2;
1331                 ip_options_build(skb, opt, cork->addr, rt, 0);
1332         }
1333         iph->tos = inet->tos;
1334         iph->frag_off = df;
1335         ip_select_ident(iph, &rt->dst, sk);
1336         iph->ttl = ttl;
1337         iph->protocol = sk->sk_protocol;
1338         iph->saddr = rt->rt_src;
1339         iph->daddr = rt->rt_dst;
1340
1341         skb->priority = sk->sk_priority;
1342         skb->mark = sk->sk_mark;
1343         /*
1344          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1345          * on dst refcount
1346          */
1347         cork->dst = NULL;
1348         skb_dst_set(skb, &rt->dst);
1349
1350         if (iph->protocol == IPPROTO_ICMP)
1351                 icmp_out_count(net, ((struct icmphdr *)
1352                         skb_transport_header(skb))->type);
1353
1354         ip_cork_release(cork);
1355 out:
1356         return skb;
1357 }
1358
1359 int ip_send_skb(struct sk_buff *skb)
1360 {
1361         struct net *net = sock_net(skb->sk);
1362         int err;
1363
1364         err = ip_local_out(skb);
1365         if (err) {
1366                 if (err > 0)
1367                         err = net_xmit_errno(err);
1368                 if (err)
1369                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1370         }
1371
1372         return err;
1373 }
1374
1375 int ip_push_pending_frames(struct sock *sk)
1376 {
1377         struct sk_buff *skb;
1378
1379         skb = ip_finish_skb(sk);
1380         if (!skb)
1381                 return 0;
1382
1383         /* Netfilter gets whole the not fragmented skb. */
1384         return ip_send_skb(skb);
1385 }
1386
1387 /*
1388  *      Throw away all pending data on the socket.
1389  */
1390 static void __ip_flush_pending_frames(struct sock *sk,
1391                                       struct sk_buff_head *queue,
1392                                       struct inet_cork *cork)
1393 {
1394         struct sk_buff *skb;
1395
1396         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1397                 kfree_skb(skb);
1398
1399         ip_cork_release(cork);
1400 }
1401
1402 void ip_flush_pending_frames(struct sock *sk)
1403 {
1404         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork);
1405 }
1406
1407 struct sk_buff *ip_make_skb(struct sock *sk,
1408                             int getfrag(void *from, char *to, int offset,
1409                                         int len, int odd, struct sk_buff *skb),
1410                             void *from, int length, int transhdrlen,
1411                             struct ipcm_cookie *ipc, struct rtable **rtp,
1412                             unsigned int flags)
1413 {
1414         struct inet_cork cork = {};
1415         struct sk_buff_head queue;
1416         int err;
1417
1418         if (flags & MSG_PROBE)
1419                 return NULL;
1420
1421         __skb_queue_head_init(&queue);
1422
1423         err = ip_setup_cork(sk, &cork, ipc, rtp);
1424         if (err)
1425                 return ERR_PTR(err);
1426
1427         err = __ip_append_data(sk, &queue, &cork, getfrag,
1428                                from, length, transhdrlen, flags);
1429         if (err) {
1430                 __ip_flush_pending_frames(sk, &queue, &cork);
1431                 return ERR_PTR(err);
1432         }
1433
1434         return __ip_make_skb(sk, &queue, &cork);
1435 }
1436
1437 /*
1438  *      Fetch data from kernel space and fill in checksum if needed.
1439  */
1440 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1441                               int len, int odd, struct sk_buff *skb)
1442 {
1443         __wsum csum;
1444
1445         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1446         skb->csum = csum_block_add(skb->csum, csum, odd);
1447         return 0;
1448 }
1449
1450 /*
1451  *      Generic function to send a packet as reply to another packet.
1452  *      Used to send TCP resets so far. ICMP should use this function too.
1453  *
1454  *      Should run single threaded per socket because it uses the sock
1455  *      structure to pass arguments.
1456  */
1457 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1458                    unsigned int len)
1459 {
1460         struct inet_sock *inet = inet_sk(sk);
1461         struct {
1462                 struct ip_options       opt;
1463                 char                    data[40];
1464         } replyopts;
1465         struct ipcm_cookie ipc;
1466         __be32 daddr;
1467         struct rtable *rt = skb_rtable(skb);
1468
1469         if (ip_options_echo(&replyopts.opt, skb))
1470                 return;
1471
1472         daddr = ipc.addr = rt->rt_src;
1473         ipc.opt = NULL;
1474         ipc.tx_flags = 0;
1475
1476         if (replyopts.opt.optlen) {
1477                 ipc.opt = &replyopts.opt;
1478
1479                 if (ipc.opt->srr)
1480                         daddr = replyopts.opt.faddr;
1481         }
1482
1483         {
1484                 struct flowi fl = { .oif = arg->bound_dev_if,
1485                                     .fl4_dst = daddr,
1486                                     .fl4_src = rt->rt_spec_dst,
1487                                     .fl4_tos = RT_TOS(ip_hdr(skb)->tos),
1488                                     .fl_ip_sport = tcp_hdr(skb)->dest,
1489                                     .fl_ip_dport = tcp_hdr(skb)->source,
1490                                     .proto = sk->sk_protocol,
1491                                     .flags = ip_reply_arg_flowi_flags(arg) };
1492                 security_skb_classify_flow(skb, &fl);
1493                 rt = ip_route_output_key(sock_net(sk), &fl);
1494                 if (IS_ERR(rt))
1495                         return;
1496         }
1497
1498         /* And let IP do all the hard work.
1499
1500            This chunk is not reenterable, hence spinlock.
1501            Note that it uses the fact, that this function is called
1502            with locally disabled BH and that sk cannot be already spinlocked.
1503          */
1504         bh_lock_sock(sk);
1505         inet->tos = ip_hdr(skb)->tos;
1506         sk->sk_priority = skb->priority;
1507         sk->sk_protocol = ip_hdr(skb)->protocol;
1508         sk->sk_bound_dev_if = arg->bound_dev_if;
1509         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1510                        &ipc, &rt, MSG_DONTWAIT);
1511         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1512                 if (arg->csumoffset >= 0)
1513                         *((__sum16 *)skb_transport_header(skb) +
1514                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1515                                                                 arg->csum));
1516                 skb->ip_summed = CHECKSUM_NONE;
1517                 ip_push_pending_frames(sk);
1518         }
1519
1520         bh_unlock_sock(sk);
1521
1522         ip_rt_put(rt);
1523 }
1524
1525 void __init ip_init(void)
1526 {
1527         ip_rt_init();
1528         inet_initpeers();
1529
1530 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1531         igmp_mc_proc_init();
1532 #endif
1533 }