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  * Version:     $Id: ip_output.c,v 1.100 2002/02/01 22:01:03 davem Exp $
   9  *
  10  * Authors:     Ross Biro, <bir7@leland.Stanford.Edu>
  11  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  12  *              Donald Becker, <becker@super.org>
  13  *              Alan Cox, <Alan.Cox@linux.org>
  14  *              Richard Underwood
  15  *              Stefan Becker, <stefanb@yello.ping.de>
  16  *              Jorge Cwik, <jorge@laser.satlink.net>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Hirokazu Takahashi, <taka@valinux.co.jp>
  19  *
  20  *      See ip_input.c for original log
  21  *
  22  *      Fixes:
  23  *              Alan Cox        :       Missing nonblock feature in ip_build_xmit.
  24  *              Mike Kilburn    :       htons() missing in ip_build_xmit.
  25  *              Bradford Johnson:       Fix faulty handling of some frames when
  26  *                                      no route is found.
  27  *              Alexander Demenshin:    Missing sk/skb free in ip_queue_xmit
  28  *                                      (in case if packet not accepted by
  29  *                                      output firewall rules)
  30  *              Mike McLagan    :       Routing by source
  31  *              Alexey Kuznetsov:       use new route cache
  32  *              Andi Kleen:             Fix broken PMTU recovery and remove
  33  *                                      some redundant tests.
  34  *      Vitaly E. Lavrov        :       Transparent proxy revived after year coma.
  35  *              Andi Kleen      :       Replace ip_reply with ip_send_reply.
  36  *              Andi Kleen      :       Split fast and slow ip_build_xmit path
  37  *                                      for decreased register pressure on x86
  38  *                                      and more readibility.
  39  *              Marc Boucher    :       When call_out_firewall returns FW_QUEUE,
  40  *                                      silently drop skb instead of failing with -EPERM.
  41  *              Detlev Wengorz  :       Copy protocol for fragments.
  42  *              Hirokazu Takahashi:     HW checksumming for outgoing UDP
  43  *                                      datagrams.
  44  *              Hirokazu Takahashi:     sendfile() on UDP works now.
  45  */
  46
  47 #include <asm/uaccess.h>
  48 #include <asm/system.h>
  49 #include <linux/module.h>
  50 #include <linux/types.h>
  51 #include <linux/kernel.h>
  52 #include <linux/sched.h>
  53 #include <linux/mm.h>
  54 #include <linux/string.h>
  55 #include <linux/errno.h>
  56 #include <linux/config.h>
  57
  58 #include <linux/socket.h>
  59 #include <linux/sockios.h>
  60 #include <linux/in.h>
  61 #include <linux/inet.h>
  62 #include <linux/netdevice.h>
  63 #include <linux/etherdevice.h>
  64 #include <linux/proc_fs.h>
  65 #include <linux/stat.h>
  66 #include <linux/init.h>
  67
  68 #include <net/snmp.h>
  69 #include <net/ip.h>
  70 #include <net/protocol.h>
  71 #include <net/route.h>
  72 #include <net/tcp.h>
  73 #include <net/udp.h>
  74 #include <linux/skbuff.h>
  75 #include <net/sock.h>
  76 #include <net/arp.h>
  77 #include <net/icmp.h>
  78 #include <net/raw.h>
  79 #include <net/checksum.h>
  80 #include <net/inetpeer.h>
  81 #include <linux/igmp.h>
  82 #include <linux/netfilter_ipv4.h>
  83 #include <linux/netfilter_bridge.h>
  84 #include <linux/mroute.h>
  85 #include <linux/netlink.h>
  86
  87 /*
  88  *      Shall we try to damage output packets if routing dev changes?
  89  */
  90
  91 int sysctl_ip_dynaddr;
  92 int sysctl_ip_default_ttl = IPDEFTTL;
  93
  94 /* Generate a checksum for an outgoing IP datagram. */
  95 __inline__ void ip_send_check(struct iphdr *iph)
  96 {
  97         iph->check = 0;
  98         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  99 }
 100
 101 /* dev_loopback_xmit for use with netfilter. */
 102 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
 103 {
 104         newskb->mac.raw = newskb->data;
 105         __skb_pull(newskb, newskb->nh.raw - newskb->data);
 106         newskb->pkt_type = PACKET_LOOPBACK;
 107         newskb->ip_summed = CHECKSUM_UNNECESSARY;
 108         BUG_TRAP(newskb->dst);
 109
 110 #ifdef CONFIG_NETFILTER_DEBUG
 111         nf_debug_ip_loopback_xmit(newskb);
 112 #endif
 113         netif_rx(newskb);
 114         return 0;
 115 }
 116
 117 static inline int ip_select_ttl(struct inet_opt *inet, struct dst_entry *dst)
 118 {
 119         int ttl = inet->uc_ttl;
 120
 121         if (ttl < 0)
 122                 ttl = dst_metric(dst, RTAX_HOPLIMIT);
 123         return ttl;
 124 }
 125
 126 /*
 127  *              Add an ip header to a skbuff and send it out.
 128  *
 129  */
 130 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 131                           u32 saddr, u32 daddr, struct ip_options *opt)
 132 {
 133         struct inet_opt *inet = inet_sk(sk);
 134         struct rtable *rt = (struct rtable *)skb->dst;
 135         struct iphdr *iph;
 136
 137         /* Build the IP header. */
 138         if (opt)
 139                 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr) + opt->optlen);
 140         else
 141                 iph=(struct iphdr *)skb_push(skb,sizeof(struct iphdr));
 142
 143         iph->version  = 4;
 144         iph->ihl      = 5;
 145         iph->tos      = inet->tos;
 146         if (ip_dont_fragment(sk, &rt->u.dst))
 147                 iph->frag_off = htons(IP_DF);
 148         else
 149                 iph->frag_off = 0;
 150         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 151         iph->daddr    = rt->rt_dst;
 152         iph->saddr    = rt->rt_src;
 153         iph->protocol = sk->sk_protocol;
 154         iph->tot_len  = htons(skb->len);
 155         ip_select_ident(iph, &rt->u.dst, sk);
 156         skb->nh.iph   = iph;
 157
 158         if (opt && opt->optlen) {
 159                 iph->ihl += opt->optlen>>2;
 160                 ip_options_build(skb, opt, daddr, rt, 0);
 161         }
 162         ip_send_check(iph);
 163
 164         skb->priority = sk->sk_priority;
 165
 166         /* Send it out. */
 167         return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
 168                        dst_output);
 169 }
 170
 171 static inline int ip_finish_output2(struct sk_buff *skb)
 172 {
 173         struct dst_entry *dst = skb->dst;
 174         struct hh_cache *hh = dst->hh;
 175         struct net_device *dev = dst->dev;
 176         int hh_len = LL_RESERVED_SPACE(dev);
 177
 178         /* Be paranoid, rather than too clever. */
 179         if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
 180                 struct sk_buff *skb2;
 181
 182                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 183                 if (skb2 == NULL) {
 184                         kfree_skb(skb);
 185                         return -ENOMEM;
 186                 }
 187                 if (skb->sk)
 188                         skb_set_owner_w(skb2, skb->sk);
 189                 kfree_skb(skb);
 190                 skb = skb2;
 191         }
 192
 193 #ifdef CONFIG_NETFILTER_DEBUG
 194         nf_debug_ip_finish_output2(skb);
 195 #endif /*CONFIG_NETFILTER_DEBUG*/
 196
 197         if (hh) {
 198                 int hh_alen;
 199
 200                 read_lock_bh(&hh->hh_lock);
 201                 hh_alen = HH_DATA_ALIGN(hh->hh_len);
 202                 memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
 203                 read_unlock_bh(&hh->hh_lock);
 204                 skb_push(skb, hh->hh_len);
 205                 return hh->hh_output(skb);
 206         } else if (dst->neighbour)
 207                 return dst->neighbour->output(skb);
 208
 209         if (net_ratelimit())
 210                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 211         kfree_skb(skb);
 212         return -EINVAL;
 213 }
 214
 215 int ip_finish_output(struct sk_buff *skb)
 216 {
 217         struct net_device *dev = skb->dst->dev;
 218
 219         skb->dev = dev;
 220         skb->protocol = htons(ETH_P_IP);
 221
 222         return NF_HOOK(PF_INET, NF_IP_POST_ROUTING, skb, NULL, dev,
 223                        ip_finish_output2);
 224 }
 225
 226 int ip_mc_output(struct sk_buff **pskb)
 227 {
 228         struct sk_buff *skb = *pskb;
 229         struct sock *sk = skb->sk;
 230         struct rtable *rt = (struct rtable*)skb->dst;
 231         struct net_device *dev = rt->u.dst.dev;
 232
 233         /*
 234          *      If the indicated interface is up and running, send the packet.
 235          */
 236         IP_INC_STATS(OutRequests);
 237
 238         skb->dev = dev;
 239         skb->protocol = htons(ETH_P_IP);
 240
 241         /*
 242          *      Multicasts are looped back for other local users
 243          */
 244
 245         if (rt->rt_flags&RTCF_MULTICAST) {
 246                 if ((!sk || inet_sk(sk)->mc_loop)
 247 #ifdef CONFIG_IP_MROUTE
 248                 /* Small optimization: do not loopback not local frames,
 249                    which returned after forwarding; they will be  dropped
 250                    by ip_mr_input in any case.
 251                    Note, that local frames are looped back to be delivered
 252                    to local recipients.
 253
 254                    This check is duplicated in ip_mr_input at the moment.
 255                  */
 256                     && ((rt->rt_flags&RTCF_LOCAL) || !(IPCB(skb)->flags&IPSKB_FORWARDED))
 257 #endif
 258                 ) {
 259                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 260                         if (newskb)
 261                                 NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
 262                                         newskb->dev,
 263                                         ip_dev_loopback_xmit);
 264                 }
 265
 266                 /* Multicasts with ttl 0 must not go beyond the host */
 267
 268                 if (skb->nh.iph->ttl == 0) {
 269                         kfree_skb(skb);
 270                         return 0;
 271                 }
 272         }
 273
 274         if (rt->rt_flags&RTCF_BROADCAST) {
 275                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 276                 if (newskb)
 277                         NF_HOOK(PF_INET, NF_IP_POST_ROUTING, newskb, NULL,
 278                                 newskb->dev, ip_dev_loopback_xmit);
 279         }
 280
 281         if (skb->len > dst_pmtu(&rt->u.dst) || skb_shinfo(skb)->frag_list)
 282                 return ip_fragment(skb, ip_finish_output);
 283         else
 284                 return ip_finish_output(skb);
 285 }
 286
 287 int ip_output(struct sk_buff **pskb)
 288 {
 289         struct sk_buff *skb = *pskb;
 290
 291         IP_INC_STATS(OutRequests);
 292
 293         if ((skb->len > dst_pmtu(skb->dst) || skb_shinfo(skb)->frag_list) &&
 294             !skb_shinfo(skb)->tso_size)
 295                 return ip_fragment(skb, ip_finish_output);
 296         else
 297                 return ip_finish_output(skb);
 298 }
 299
 300 int ip_queue_xmit(struct sk_buff *skb, int ipfragok)
 301 {
 302         struct sock *sk = skb->sk;
 303         struct inet_opt *inet = inet_sk(sk);
 304         struct ip_options *opt = inet->opt;
 305         struct rtable *rt;
 306         struct iphdr *iph;
 307         u32 mtu;
 308
 309         /* Skip all of this if the packet is already routed,
 310          * f.e. by something like SCTP.
 311          */
 312         rt = (struct rtable *) skb->dst;
 313         if (rt != NULL)
 314                 goto packet_routed;
 315
 316         /* Make sure we can route this packet. */
 317         rt = (struct rtable *)__sk_dst_check(sk, 0);
 318         if (rt == NULL) {
 319                 u32 daddr;
 320
 321                 /* Use correct destination address if we have options. */
 322                 daddr = inet->daddr;
 323                 if(opt && opt->srr)
 324                         daddr = opt->faddr;
 325
 326                 {
 327                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
 328                                             .nl_u = { .ip4_u =
 329                                                       { .daddr = daddr,
 330                                                         .saddr = inet->saddr,
 331                                                         .tos = RT_CONN_FLAGS(sk) } },
 332                                             .proto = sk->sk_protocol,
 333                                             .uli_u = { .ports =
 334                                                        { .sport = inet->sport,
 335                                                          .dport = inet->dport } } };
 336
 337                         /* If this fails, retransmit mechanism of transport layer will
 338                          * keep trying until route appears or the connection times
 339                          * itself out.
 340                          */
 341                         if (ip_route_output_flow(&rt, &fl, sk, 0))
 342                                 goto no_route;
 343                 }
 344                 __sk_dst_set(sk, &rt->u.dst);
 345                 tcp_v4_setup_caps(sk, &rt->u.dst);
 346         }
 347         skb->dst = dst_clone(&rt->u.dst);
 348
 349 packet_routed:
 350         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
 351                 goto no_route;
 352
 353         /* OK, we know where to send it, allocate and build IP header. */
 354         iph = (struct iphdr *) skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 355         *((__u16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 356         iph->tot_len = htons(skb->len);
 357         if (ip_dont_fragment(sk, &rt->u.dst) && !ipfragok)
 358                 iph->frag_off = htons(IP_DF);
 359         else
 360                 iph->frag_off = 0;
 361         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 362         iph->protocol = sk->sk_protocol;
 363         iph->saddr    = rt->rt_src;
 364         iph->daddr    = rt->rt_dst;
 365         skb->nh.iph   = iph;
 366         /* Transport layer set skb->h.foo itself. */
 367
 368         if(opt && opt->optlen) {
 369                 iph->ihl += opt->optlen >> 2;
 370                 ip_options_build(skb, opt, inet->daddr, rt, 0);
 371         }
 372
 373         mtu = dst_pmtu(&rt->u.dst);
 374         if (skb->len > mtu && (sk->sk_route_caps & NETIF_F_TSO)) {
 375                 unsigned int hlen;
 376
 377                 /* Hack zone: all this must be done by TCP. */
 378                 hlen = ((skb->h.raw - skb->data) + (skb->h.th->doff << 2));
 379                 skb_shinfo(skb)->tso_size = mtu - hlen;
 380                 skb_shinfo(skb)->tso_segs =
 381                         (skb->len - hlen + skb_shinfo(skb)->tso_size - 1)/
 382                                 skb_shinfo(skb)->tso_size - 1;
 383         }
 384
 385         ip_select_ident_more(iph, &rt->u.dst, sk, skb_shinfo(skb)->tso_segs);
 386
 387         /* Add an IP checksum. */
 388         ip_send_check(iph);
 389
 390         skb->priority = sk->sk_priority;
 391
 392         return NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL, rt->u.dst.dev,
 393                        dst_output);
 394
 395 no_route:
 396         IP_INC_STATS(OutNoRoutes);
 397         kfree_skb(skb);
 398         return -EHOSTUNREACH;
 399 }
 400
 401
 402 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 403 {
 404         to->pkt_type = from->pkt_type;
 405         to->priority = from->priority;
 406         to->protocol = from->protocol;
 407         to->security = from->security;
 408         to->dst = dst_clone(from->dst);
 409         to->dev = from->dev;
 410
 411         /* Copy the flags to each fragment. */
 412         IPCB(to)->flags = IPCB(from)->flags;
 413
 414 #ifdef CONFIG_NET_SCHED
 415         to->tc_index = from->tc_index;
 416 #endif
 417 #ifdef CONFIG_NETFILTER
 418         to->nfmark = from->nfmark;
 419         to->nfcache = from->nfcache;
 420         /* Connection association is same as pre-frag packet */
 421         nf_conntrack_put(to->nfct);
 422         to->nfct = from->nfct;
 423         nf_conntrack_get(to->nfct);
 424 #ifdef CONFIG_BRIDGE_NETFILTER
 425         nf_bridge_put(to->nf_bridge);
 426         to->nf_bridge = from->nf_bridge;
 427         nf_bridge_get(to->nf_bridge);
 428 #endif
 429 #ifdef CONFIG_NETFILTER_DEBUG
 430         to->nf_debug = from->nf_debug;
 431 #endif
 432 #endif
 433 }
 434
 435 /*
 436  *      This IP datagram is too large to be sent in one piece.  Break it up into
 437  *      smaller pieces (each of size equal to IP header plus
 438  *      a block of the data of the original IP data part) that will yet fit in a
 439  *      single device frame, and queue such a frame for sending.
 440  */
 441
 442 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff*))
 443 {
 444         struct iphdr *iph;
 445         int raw = 0;
 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         int not_last_frag;
 452         struct rtable *rt = (struct rtable*)skb->dst;
 453         int err = 0;
 454
 455         dev = rt->u.dst.dev;
 456
 457         /*
 458          *      Point into the IP datagram header.
 459          */
 460
 461         iph = skb->nh.iph;
 462
 463         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 464                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 465                           htonl(dst_pmtu(&rt->u.dst)));
 466                 kfree_skb(skb);
 467                 return -EMSGSIZE;
 468         }
 469
 470         /*
 471          *      Setup starting values.
 472          */
 473
 474         hlen = iph->ihl * 4;
 475         mtu = dst_pmtu(&rt->u.dst) - hlen;      /* Size of data space */
 476
 477         /* When frag_list is given, use it. First, check its validity:
 478          * some transformers could create wrong frag_list or break existing
 479          * one, it is not prohibited. In this case fall back to copying.
 480          *
 481          * LATER: this step can be merged to real generation of fragments,
 482          * we can switch to copy when see the first bad fragment.
 483          */
 484         if (skb_shinfo(skb)->frag_list) {
 485                 struct sk_buff *frag;
 486                 int first_len = skb_pagelen(skb);
 487
 488                 if (first_len - hlen > mtu ||
 489                     ((first_len - hlen) & 7) ||
 490                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
 491                     skb_cloned(skb))
 492                         goto slow_path;
 493
 494                 for (frag = skb_shinfo(skb)->frag_list; frag; frag = frag->next) {
 495                         /* Correct geometry. */
 496                         if (frag->len > mtu ||
 497                             ((frag->len & 7) && frag->next) ||
 498                             skb_headroom(frag) < hlen)
 499                             goto slow_path;
 500
 501                         /* Correct socket ownership. */
 502                         if (frag->sk == NULL && skb->sk)
 503                                 goto slow_path;
 504
 505                         /* Partially cloned skb? */
 506                         if (skb_shared(frag))
 507                                 goto slow_path;
 508                 }
 509
 510                 /* Everything is OK. Generate! */
 511
 512                 err = 0;
 513                 offset = 0;
 514                 frag = skb_shinfo(skb)->frag_list;
 515                 skb_shinfo(skb)->frag_list = 0;
 516                 skb->data_len = first_len - skb_headlen(skb);
 517                 skb->len = first_len;
 518                 iph->tot_len = htons(first_len);
 519                 iph->frag_off |= htons(IP_MF);
 520                 ip_send_check(iph);
 521
 522                 for (;;) {
 523                         /* Prepare header of the next frame,
 524                          * before previous one went down. */
 525                         if (frag) {
 526                                 frag->h.raw = frag->data;
 527                                 frag->nh.raw = __skb_push(frag, hlen);
 528                                 memcpy(frag->nh.raw, iph, hlen);
 529                                 iph = frag->nh.iph;
 530                                 iph->tot_len = htons(frag->len);
 531                                 ip_copy_metadata(frag, skb);
 532                                 if (offset == 0)
 533                                         ip_options_fragment(frag);
 534                                 offset += skb->len - hlen;
 535                                 iph->frag_off = htons(offset>>3);
 536                                 if (frag->next != NULL)
 537                                         iph->frag_off |= htons(IP_MF);
 538                                 /* Ready, complete checksum */
 539                                 ip_send_check(iph);
 540                         }
 541
 542                         err = output(skb);
 543
 544                         if (err || !frag)
 545                                 break;
 546
 547                         skb = frag;
 548                         frag = skb->next;
 549                         skb->next = NULL;
 550                 }
 551
 552                 if (err == 0) {
 553                         IP_INC_STATS(FragOKs);
 554                         return 0;
 555                 }
 556
 557                 while (frag) {
 558                         skb = frag->next;
 559                         kfree_skb(frag);
 560                         frag = skb;
 561                 }
 562                 IP_INC_STATS(FragFails);
 563                 return err;
 564         }
 565
 566 slow_path:
 567         left = skb->len - hlen;         /* Space per frame */
 568         ptr = raw + hlen;               /* Where to start from */
 569
 570 #ifdef CONFIG_BRIDGE_NETFILTER
 571         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 572          * we need to make room for the encapsulating header */
 573         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, nf_bridge_pad(skb));
 574         mtu -= nf_bridge_pad(skb);
 575 #else
 576         ll_rs = LL_RESERVED_SPACE(rt->u.dst.dev);
 577 #endif
 578         /*
 579          *      Fragment the datagram.
 580          */
 581
 582         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 583         not_last_frag = iph->frag_off & htons(IP_MF);
 584
 585         /*
 586          *      Keep copying data until we run out.
 587          */
 588
 589         while(left > 0) {
 590                 len = left;
 591                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 592                 if (len > mtu)
 593                         len = mtu;
 594                 /* IF: we are not sending upto and including the packet end
 595                    then align the next start on an eight byte boundary */
 596                 if (len < left) {
 597                         len &= ~7;
 598                 }
 599                 /*
 600                  *      Allocate buffer.
 601                  */
 602
 603                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 604                         NETDEBUG(printk(KERN_INFO "IP: frag: no memory for new fragment!\n"));
 605                         err = -ENOMEM;
 606                         goto fail;
 607                 }
 608
 609                 /*
 610                  *      Set up data on packet
 611                  */
 612
 613                 ip_copy_metadata(skb2, skb);
 614                 skb_reserve(skb2, ll_rs);
 615                 skb_put(skb2, len + hlen);
 616                 skb2->nh.raw = skb2->data;
 617                 skb2->h.raw = skb2->data + hlen;
 618
 619                 /*
 620                  *      Charge the memory for the fragment to any owner
 621                  *      it might possess
 622                  */
 623
 624                 if (skb->sk)
 625                         skb_set_owner_w(skb2, skb->sk);
 626
 627                 /*
 628                  *      Copy the packet header into the new buffer.
 629                  */
 630
 631                 memcpy(skb2->nh.raw, skb->data, hlen);
 632
 633                 /*
 634                  *      Copy a block of the IP datagram.
 635                  */
 636                 if (skb_copy_bits(skb, ptr, skb2->h.raw, len))
 637                         BUG();
 638                 left -= len;
 639
 640                 /*
 641                  *      Fill in the new header fields.
 642                  */
 643                 iph = skb2->nh.iph;
 644                 iph->frag_off = htons((offset >> 3));
 645
 646                 /* ANK: dirty, but effective trick. Upgrade options only if
 647                  * the segment to be fragmented was THE FIRST (otherwise,
 648                  * options are already fixed) and make it ONCE
 649                  * on the initial skb, so that all the following fragments
 650                  * will inherit fixed options.
 651                  */
 652                 if (offset == 0)
 653                         ip_options_fragment(skb);
 654
 655                 /*
 656                  *      Added AC : If we are fragmenting a fragment that's not the
 657                  *                 last fragment then keep MF on each bit
 658                  */
 659                 if (left > 0 || not_last_frag)
 660                         iph->frag_off |= htons(IP_MF);
 661                 ptr += len;
 662                 offset += len;
 663
 664                 /*
 665                  *      Put this fragment into the sending queue.
 666                  */
 667
 668                 IP_INC_STATS(FragCreates);
 669
 670                 iph->tot_len = htons(len + hlen);
 671
 672                 ip_send_check(iph);
 673
 674                 err = output(skb2);
 675                 if (err)
 676                         goto fail;
 677         }
 678         kfree_skb(skb);
 679         IP_INC_STATS(FragOKs);
 680         return err;
 681
 682 fail:
 683         kfree_skb(skb);
 684         IP_INC_STATS(FragFails);
 685         return err;
 686 }
 687
 688 int
 689 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 690 {
 691         struct iovec *iov = from;
 692
 693         if (skb->ip_summed == CHECKSUM_HW) {
 694                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 695                         return -EFAULT;
 696         } else {
 697                 unsigned int csum = 0;
 698                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 699                         return -EFAULT;
 700                 skb->csum = csum_block_add(skb->csum, csum, odd);
 701         }
 702         return 0;
 703 }
 704
 705 static inline int
 706 skb_can_coalesce(struct sk_buff *skb, int i, struct page *page, int off)
 707 {
 708         if (i) {
 709                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
 710                 return page == frag->page &&
 711                         off == frag->page_offset+frag->size;
 712         }
 713         return 0;
 714 }
 715
 716 static inline unsigned int
 717 csum_page(struct page *page, int offset, int copy)
 718 {
 719         char *kaddr;
 720         unsigned int csum;
 721         kaddr = kmap(page);
 722         csum = csum_partial(kaddr + offset, copy, 0);
 723         kunmap(page);
 724         return csum;
 725 }
 726
 727 /*
 728  *      ip_append_data() and ip_append_page() can make one large IP datagram
 729  *      from many pieces of data. Each pieces will be holded on the socket
 730  *      until ip_push_pending_frames() is called. Eache pieces can be a page
 731  *      or non-page data.
 732  *
 733  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
 734  *      this interface potentially.
 735  *
 736  *      LATER: length must be adjusted by pad at tail, when it is required.
 737  */
 738 int ip_append_data(struct sock *sk,
 739                    int getfrag(void *from, char *to, int offset, int len,
 740                                int odd, struct sk_buff *skb),
 741                    void *from, int length, int transhdrlen,
 742                    struct ipcm_cookie *ipc, struct rtable *rt,
 743                    unsigned int flags)
 744 {
 745         struct inet_opt *inet = inet_sk(sk);
 746         struct sk_buff *skb;
 747
 748         struct ip_options *opt = NULL;
 749         int hh_len;
 750         int exthdrlen;
 751         int mtu;
 752         int copy;
 753         int err;
 754         int offset = 0;
 755         unsigned int maxfraglen, fragheaderlen;
 756         int csummode = CHECKSUM_NONE;
 757
 758         if (flags&MSG_PROBE)
 759                 return 0;
 760
 761         if (skb_queue_empty(&sk->sk_write_queue)) {
 762                 /*
 763                  * setup for corking.
 764                  */
 765                 opt = ipc->opt;
 766                 if (opt) {
 767                         if (inet->cork.opt == NULL) {
 768                                 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
 769                                 if (unlikely(inet->cork.opt == NULL))
 770                                         return -ENOBUFS;
 771                         }
 772                         memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
 773                         inet->cork.flags |= IPCORK_OPT;
 774                         inet->cork.addr = ipc->addr;
 775                 }
 776                 dst_hold(&rt->u.dst);
 777                 inet->cork.fragsize = mtu = dst_pmtu(&rt->u.dst);
 778                 inet->cork.rt = rt;
 779                 inet->cork.length = 0;
 780                 inet->sndmsg_page = NULL;
 781                 inet->sndmsg_off = 0;
 782                 if ((exthdrlen = rt->u.dst.header_len) != 0) {
 783                         length += exthdrlen;
 784                         transhdrlen += exthdrlen;
 785                 }
 786         } else {
 787                 rt = inet->cork.rt;
 788                 if (inet->cork.flags & IPCORK_OPT)
 789                         opt = inet->cork.opt;
 790
 791                 transhdrlen = 0;
 792                 exthdrlen = 0;
 793                 mtu = inet->cork.fragsize;
 794         }
 795         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
 796
 797         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 798         maxfraglen = ((mtu-fragheaderlen) & ~7) + fragheaderlen;
 799
 800         if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
 801                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu-exthdrlen);
 802                 return -EMSGSIZE;
 803         }
 804
 805         /*
 806          * transhdrlen > 0 means that this is the first fragment and we wish
 807          * it won't be fragmented in the future.
 808          */
 809         if (transhdrlen &&
 810             length + fragheaderlen <= maxfraglen &&
 811             rt->u.dst.dev->features&(NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM) &&
 812             !exthdrlen)
 813                 csummode = CHECKSUM_HW;
 814
 815         inet->cork.length += length;
 816
 817         /* So, what's going on in the loop below?
 818          *
 819          * We use calculated fragment length to generate chained skb,
 820          * each of segments is IP fragment ready for sending to network after
 821          * adding appropriate IP header.
 822          *
 823          * Mistake is:
 824          *
 825          *    If mtu-fragheaderlen is not 0 modulo 8, we generate additional
 826          *    small fragment of length (mtu-fragheaderlen)%8, even though
 827          *    it is not necessary. Not a big bug, but needs a fix.
 828          */
 829
 830         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
 831                 goto alloc_new_skb;
 832
 833         while (length > 0) {
 834                 if ((copy = maxfraglen - skb->len) <= 0) {
 835                         char *data;
 836                         unsigned int datalen;
 837                         unsigned int fraglen;
 838                         unsigned int alloclen;
 839                         BUG_TRAP(copy == 0);
 840
 841 alloc_new_skb:
 842                         datalen = maxfraglen - fragheaderlen;
 843                         if (datalen > length)
 844                                 datalen = length;
 845
 846                         fraglen = datalen + fragheaderlen;
 847                         if ((flags & MSG_MORE) &&
 848                             !(rt->u.dst.dev->features&NETIF_F_SG))
 849                                 alloclen = maxfraglen;
 850                         else
 851                                 alloclen = datalen + fragheaderlen;
 852
 853                         /* The last fragment gets additional space at tail.
 854                          * Note, with MSG_MORE we overallocate on fragments,
 855                          * because we have no idea what fragment will be
 856                          * the last.
 857                          */
 858                         if (datalen == length)
 859                                 alloclen += rt->u.dst.trailer_len;
 860
 861                         if (transhdrlen) {
 862                                 skb = sock_alloc_send_skb(sk,
 863                                                 alloclen + hh_len + 15,
 864                                                 (flags & MSG_DONTWAIT), &err);
 865                         } else {
 866                                 skb = NULL;
 867                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 868                                     2 * sk->sk_sndbuf)
 869                                         skb = sock_wmalloc(sk,
 870                                                            alloclen + hh_len + 15, 1,
 871                                                            sk->sk_allocation);
 872                                 if (unlikely(skb == NULL))
 873                                         err = -ENOBUFS;
 874                         }
 875                         if (skb == NULL)
 876                                 goto error;
 877
 878                         /*
 879                          *      Fill in the control structures
 880                          */
 881                         skb->ip_summed = csummode;
 882                         skb->csum = 0;
 883                         skb_reserve(skb, hh_len);
 884
 885                         /*
 886                          *      Find where to start putting bytes.
 887                          */
 888                         data = skb_put(skb, fraglen);
 889                         skb->nh.raw = data + exthdrlen;
 890                         data += fragheaderlen;
 891                         skb->h.raw = data + exthdrlen;
 892
 893                         copy = datalen - transhdrlen;
 894                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, 0, skb) < 0) {
 895                                 err = -EFAULT;
 896                                 kfree_skb(skb);
 897                                 goto error;
 898                         }
 899
 900                         offset += copy;
 901                         length -= datalen;
 902                         transhdrlen = 0;
 903                         exthdrlen = 0;
 904                         csummode = CHECKSUM_NONE;
 905
 906                         /*
 907                          * Put the packet on the pending queue.
 908                          */
 909                         __skb_queue_tail(&sk->sk_write_queue, skb);
 910                         continue;
 911                 }
 912
 913                 if (copy > length)
 914                         copy = length;
 915
 916                 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
 917                         unsigned int off;
 918
 919                         off = skb->len;
 920                         if (getfrag(from, skb_put(skb, copy),
 921                                         offset, copy, off, skb) < 0) {
 922                                 __skb_trim(skb, off);
 923                                 err = -EFAULT;
 924                                 goto error;
 925                         }
 926                 } else {
 927                         int i = skb_shinfo(skb)->nr_frags;
 928                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
 929                         struct page *page = inet->sndmsg_page;
 930                         int off = inet->sndmsg_off;
 931                         unsigned int left;
 932
 933                         if (page && (left = PAGE_SIZE - off) > 0) {
 934                                 if (copy >= left)
 935                                         copy = left;
 936                                 if (page != frag->page) {
 937                                         if (i == MAX_SKB_FRAGS) {
 938                                                 err = -EMSGSIZE;
 939                                                 goto error;
 940                                         }
 941                                         get_page(page);
 942                                         skb_fill_page_desc(skb, i, page, inet->sndmsg_off, 0);
 943                                         frag = &skb_shinfo(skb)->frags[i];
 944                                 }
 945                         } else if (i < MAX_SKB_FRAGS) {
 946                                 if (copy > PAGE_SIZE)
 947                                         copy = PAGE_SIZE;
 948                                 page = alloc_pages(sk->sk_allocation, 0);
 949                                 if (page == NULL)  {
 950                                         err = -ENOMEM;
 951                                         goto error;
 952                                 }
 953                                 inet->sndmsg_page = page;
 954                                 inet->sndmsg_off = 0;
 955
 956                                 skb_fill_page_desc(skb, i, page, 0, 0);
 957                                 frag = &skb_shinfo(skb)->frags[i];
 958                                 skb->truesize += PAGE_SIZE;
 959                                 atomic_add(PAGE_SIZE, &sk->sk_wmem_alloc);
 960                         } else {
 961                                 err = -EMSGSIZE;
 962                                 goto error;
 963                         }
 964                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
 965                                 err = -EFAULT;
 966                                 goto error;
 967                         }
 968                         inet->sndmsg_off += copy;
 969                         frag->size += copy;
 970                         skb->len += copy;
 971                         skb->data_len += copy;
 972                 }
 973                 offset += copy;
 974                 length -= copy;
 975         }
 976
 977         return 0;
 978
 979 error:
 980         inet->cork.length -= length;
 981         IP_INC_STATS(OutDiscards);
 982         return err;
 983 }
 984
 985 ssize_t ip_append_page(struct sock *sk, struct page *page,
 986                        int offset, size_t size, int flags)
 987 {
 988         struct inet_opt *inet = inet_sk(sk);
 989         struct sk_buff *skb;
 990         struct rtable *rt;
 991         struct ip_options *opt = NULL;
 992         int hh_len;
 993         int mtu;
 994         int len;
 995         int err;
 996         unsigned int maxfraglen, fragheaderlen;
 997
 998         if (inet->hdrincl)
 999                 return -EPERM;
1000
1001         if (flags&MSG_PROBE)
1002                 return 0;
1003
1004         if (skb_queue_empty(&sk->sk_write_queue))
1005                 return -EINVAL;
1006
1007         rt = inet->cork.rt;
1008         if (inet->cork.flags & IPCORK_OPT)
1009                 opt = inet->cork.opt;
1010
1011         if (!(rt->u.dst.dev->features&NETIF_F_SG))
1012                 return -EOPNOTSUPP;
1013
1014         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1015         mtu = inet->cork.fragsize;
1016
1017         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1018         maxfraglen = ((mtu-fragheaderlen) & ~7) + fragheaderlen;
1019
1020         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1021                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->dport, mtu);
1022                 return -EMSGSIZE;
1023         }
1024
1025         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1026                 return -EINVAL;
1027
1028         inet->cork.length += size;
1029
1030         while (size > 0) {
1031                 int i;
1032                 if ((len = maxfraglen - skb->len) <= 0) {
1033                         char *data;
1034                         struct iphdr *iph;
1035                         BUG_TRAP(len == 0);
1036
1037                         skb = sock_wmalloc(sk, fragheaderlen + hh_len + 15, 1,
1038                                            sk->sk_allocation);
1039                         if (unlikely(!skb)) {
1040                                 err = -ENOBUFS;
1041                                 goto error;
1042                         }
1043
1044                         /*
1045                          *      Fill in the control structures
1046                          */
1047                         skb->ip_summed = CHECKSUM_NONE;
1048                         skb->csum = 0;
1049                         skb_reserve(skb, hh_len);
1050
1051                         /*
1052                          *      Find where to start putting bytes.
1053                          */
1054                         data = skb_put(skb, fragheaderlen);
1055                         skb->nh.iph = iph = (struct iphdr *)data;
1056                         data += fragheaderlen;
1057                         skb->h.raw = data;
1058
1059                         /*
1060                          * Put the packet on the pending queue.
1061                          */
1062                         __skb_queue_tail(&sk->sk_write_queue, skb);
1063                         continue;
1064                 }
1065
1066                 i = skb_shinfo(skb)->nr_frags;
1067                 if (len > size)
1068                         len = size;
1069                 if (skb_can_coalesce(skb, i, page, offset)) {
1070                         skb_shinfo(skb)->frags[i-1].size += len;
1071                 } else if (i < MAX_SKB_FRAGS) {
1072                         get_page(page);
1073                         skb_fill_page_desc(skb, i, page, offset, len);
1074                 } else {
1075                         err = -EMSGSIZE;
1076                         goto error;
1077                 }
1078
1079                 if (skb->ip_summed == CHECKSUM_NONE) {
1080                         unsigned int csum;
1081                         csum = csum_page(page, offset, len);
1082                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1083                 }
1084
1085                 skb->len += len;
1086                 skb->data_len += len;
1087                 offset += len;
1088                 size -= len;
1089         }
1090         return 0;
1091
1092 error:
1093         inet->cork.length -= size;
1094         IP_INC_STATS(OutDiscards);
1095         return err;
1096 }
1097
1098 /*
1099  *      Combined all pending IP fragments on the socket as one IP datagram
1100  *      and push them out.
1101  */
1102 int ip_push_pending_frames(struct sock *sk)
1103 {
1104         struct sk_buff *skb, *tmp_skb;
1105         struct sk_buff **tail_skb;
1106         struct inet_opt *inet = inet_sk(sk);
1107         struct ip_options *opt = NULL;
1108         struct rtable *rt = inet->cork.rt;
1109         struct iphdr *iph;
1110         int df = 0;
1111         __u8 ttl;
1112         int err = 0;
1113
1114         if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1115                 goto out;
1116         tail_skb = &(skb_shinfo(skb)->frag_list);
1117
1118         /* move skb->data to ip header from ext header */
1119         if (skb->data < skb->nh.raw)
1120                 __skb_pull(skb, skb->nh.raw - skb->data);
1121         while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1122                 __skb_pull(tmp_skb, skb->h.raw - skb->nh.raw);
1123                 *tail_skb = tmp_skb;
1124                 tail_skb = &(tmp_skb->next);
1125                 skb->len += tmp_skb->len;
1126                 skb->data_len += tmp_skb->len;
1127 #if 0 /* Logically correct, but useless work, ip_fragment() will have to undo */
1128                 skb->truesize += tmp_skb->truesize;
1129                 __sock_put(tmp_skb->sk);
1130                 tmp_skb->destructor = NULL;
1131                 tmp_skb->sk = NULL;
1132 #endif
1133         }
1134
1135         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1136          * to fragment the frame generated here. No matter, what transforms
1137          * how transforms change size of the packet, it will come out.
1138          */
1139         if (inet->pmtudisc != IP_PMTUDISC_DO)
1140                 skb->local_df = 1;
1141
1142         /* DF bit is set when we want to see DF on outgoing frames.
1143          * If local_df is set too, we still allow to fragment this frame
1144          * locally. */
1145         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1146             (!skb_shinfo(skb)->frag_list && ip_dont_fragment(sk, &rt->u.dst)))
1147                 df = htons(IP_DF);
1148
1149         if (inet->cork.flags & IPCORK_OPT)
1150                 opt = inet->cork.opt;
1151
1152         if (rt->rt_type == RTN_MULTICAST)
1153                 ttl = inet->mc_ttl;
1154         else
1155                 ttl = ip_select_ttl(inet, &rt->u.dst);
1156
1157         iph = (struct iphdr *)skb->data;
1158         iph->version = 4;
1159         iph->ihl = 5;
1160         if (opt) {
1161                 iph->ihl += opt->optlen>>2;
1162                 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1163         }
1164         iph->tos = inet->tos;
1165         iph->tot_len = htons(skb->len);
1166         iph->frag_off = df;
1167         if (!df) {
1168                 __ip_select_ident(iph, &rt->u.dst, 0);
1169         } else {
1170                 iph->id = htons(inet->id++);
1171         }
1172         iph->ttl = ttl;
1173         iph->protocol = sk->sk_protocol;
1174         iph->saddr = rt->rt_src;
1175         iph->daddr = rt->rt_dst;
1176         ip_send_check(iph);
1177
1178         skb->priority = sk->sk_priority;
1179         skb->dst = dst_clone(&rt->u.dst);
1180
1181         /* Netfilter gets whole the not fragmented skb. */
1182         err = NF_HOOK(PF_INET, NF_IP_LOCAL_OUT, skb, NULL,
1183                       skb->dst->dev, dst_output);
1184         if (err) {
1185                 if (err > 0)
1186                         err = inet->recverr ? net_xmit_errno(err) : 0;
1187                 if (err)
1188                         goto error;
1189         }
1190
1191 out:
1192         inet->cork.flags &= ~IPCORK_OPT;
1193         if (inet->cork.opt) {
1194                 kfree(inet->cork.opt);
1195                 inet->cork.opt = NULL;
1196         }
1197         if (inet->cork.rt) {
1198                 ip_rt_put(inet->cork.rt);
1199                 inet->cork.rt = NULL;
1200         }
1201         return err;
1202
1203 error:
1204         IP_INC_STATS(OutDiscards);
1205         goto out;
1206 }
1207
1208 /*
1209  *      Throw away all pending data on the socket.
1210  */
1211 void ip_flush_pending_frames(struct sock *sk)
1212 {
1213         struct inet_opt *inet = inet_sk(sk);
1214         struct sk_buff *skb;
1215
1216         while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1217                 kfree_skb(skb);
1218
1219         inet->cork.flags &= ~IPCORK_OPT;
1220         if (inet->cork.opt) {
1221                 kfree(inet->cork.opt);
1222                 inet->cork.opt = NULL;
1223         }
1224         if (inet->cork.rt) {
1225                 ip_rt_put(inet->cork.rt);
1226                 inet->cork.rt = NULL;
1227         }
1228 }
1229
1230
1231 /*
1232  *      Fetch data from kernel space and fill in checksum if needed.
1233  */
1234 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1235                               int len, int odd, struct sk_buff *skb)
1236 {
1237         unsigned int csum;
1238
1239         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1240         skb->csum = csum_block_add(skb->csum, csum, odd);
1241         return 0;
1242 }
1243
1244 /*
1245  *      Generic function to send a packet as reply to another packet.
1246  *      Used to send TCP resets so far. ICMP should use this function too.
1247  *
1248  *      Should run single threaded per socket because it uses the sock
1249  *      structure to pass arguments.
1250  *
1251  *      LATER: switch from ip_build_xmit to ip_append_*
1252  */
1253 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1254                    unsigned int len)
1255 {
1256         struct inet_opt *inet = inet_sk(sk);
1257         struct {
1258                 struct ip_options       opt;
1259                 char                    data[40];
1260         } replyopts;
1261         struct ipcm_cookie ipc;
1262         u32 daddr;
1263         struct rtable *rt = (struct rtable*)skb->dst;
1264
1265         if (ip_options_echo(&replyopts.opt, skb))
1266                 return;
1267
1268         daddr = ipc.addr = rt->rt_src;
1269         ipc.opt = NULL;
1270
1271         if (replyopts.opt.optlen) {
1272                 ipc.opt = &replyopts.opt;
1273
1274                 if (ipc.opt->srr)
1275                         daddr = replyopts.opt.faddr;
1276         }
1277
1278         {
1279                 struct flowi fl = { .nl_u = { .ip4_u =
1280                                               { .daddr = daddr,
1281                                                 .saddr = rt->rt_spec_dst,
1282                                                 .tos = RT_TOS(skb->nh.iph->tos) } },
1283                                     /* Not quite clean, but right. */
1284                                     .uli_u = { .ports =
1285                                                { .sport = skb->h.th->dest,
1286                                                  .dport = skb->h.th->source } },
1287                                     .proto = sk->sk_protocol };
1288                 if (ip_route_output_key(&rt, &fl))
1289                         return;
1290         }
1291
1292         /* And let IP do all the hard work.
1293
1294            This chunk is not reenterable, hence spinlock.
1295            Note that it uses the fact, that this function is called
1296            with locally disabled BH and that sk cannot be already spinlocked.
1297          */
1298         bh_lock_sock(sk);
1299         inet->tos = skb->nh.iph->tos;
1300         sk->sk_priority = skb->priority;
1301         sk->sk_protocol = skb->nh.iph->protocol;
1302         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1303                        &ipc, rt, MSG_DONTWAIT);
1304         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1305                 if (arg->csumoffset >= 0)
1306                         *((u16 *)skb->h.raw + arg->csumoffset) = csum_fold(csum_add(skb->csum, arg->csum));
1307                 skb->ip_summed = CHECKSUM_NONE;
1308                 ip_push_pending_frames(sk);
1309         }
1310
1311         bh_unlock_sock(sk);
1312
1313         ip_rt_put(rt);
1314 }
1315
1316 /*
1317  *      IP protocol layer initialiser
1318  */
1319
1320 static struct packet_type ip_packet_type = {
1321         .type = __constant_htons(ETH_P_IP),
1322         .func = ip_rcv,
1323 };
1324
1325 /*
1326  *      IP registers the packet type and then calls the subprotocol initialisers
1327  */
1328
1329 void __init ip_init(void)
1330 {
1331         dev_add_pack(&ip_packet_type);
1332
1333         ip_rt_init();
1334         inet_initpeers();
1335
1336 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1337         igmp_mc_proc_init();
1338 #endif
1339 }
1340
1341 EXPORT_SYMBOL(ip_finish_output);
1342 EXPORT_SYMBOL(ip_fragment);
1343 EXPORT_SYMBOL(ip_generic_getfrag);
1344 EXPORT_SYMBOL(ip_queue_xmit);
1345 EXPORT_SYMBOL(ip_send_check);
1346
1347 #ifdef CONFIG_SYSCTL
1348 EXPORT_SYMBOL(sysctl_ip_default_ttl);
1349 #endif