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