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 <linux/module.h>
  47 #include <linux/types.h>
  48 #include <linux/kernel.h>
  49 #include <linux/mm.h>
  50 #include <linux/string.h>
  51 #include <linux/errno.h>
  52 #include <linux/highmem.h>
  53 #include <linux/slab.h>
  54
  55 #include <linux/socket.h>
  56 #include <linux/sockios.h>
  57 #include <linux/in.h>
  58 #include <linux/inet.h>
  59 #include <linux/netdevice.h>
  60 #include <linux/etherdevice.h>
  61 #include <linux/proc_fs.h>
  62 #include <linux/stat.h>
  63 #include <linux/init.h>
  64
  65 #include <net/snmp.h>
  66 #include <net/ip.h>
  67 #include <net/protocol.h>
  68 #include <net/route.h>
  69 #include <net/xfrm.h>
  70 #include <linux/skbuff.h>
  71 #include <net/sock.h>
  72 #include <net/arp.h>
  73 #include <net/icmp.h>
  74 #include <net/checksum.h>
  75 #include <net/inetpeer.h>
  76 #include <linux/igmp.h>
  77 #include <linux/netfilter_ipv4.h>
  78 #include <linux/netfilter_bridge.h>
  79 #include <linux/mroute.h>
  80 #include <linux/netlink.h>
  81 #include <linux/tcp.h>
  82
  83 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  84 EXPORT_SYMBOL(sysctl_ip_default_ttl);
  85
  86 /* Generate a checksum for an outgoing IP datagram. */
  87 __inline__ void ip_send_check(struct iphdr *iph)
  88 {
  89         iph->check = 0;
  90         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  91 }
  92 EXPORT_SYMBOL(ip_send_check);
  93
  94 int __ip_local_out(struct sk_buff *skb)
  95 {
  96         struct iphdr *iph = ip_hdr(skb);
  97
  98         iph->tot_len = htons(skb->len);
  99         ip_send_check(iph);
 100         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 101                        skb_dst(skb)->dev, dst_output);
 102 }
 103
 104 int ip_local_out(struct sk_buff *skb)
 105 {
 106         int err;
 107
 108         err = __ip_local_out(skb);
 109         if (likely(err == 1))
 110                 err = dst_output(skb);
 111
 112         return err;
 113 }
 114 EXPORT_SYMBOL_GPL(ip_local_out);
 115
 116 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 117 {
 118         int ttl = inet->uc_ttl;
 119
 120         if (ttl < 0)
 121                 ttl = ip4_dst_hoplimit(dst);
 122         return ttl;
 123 }
 124
 125 /*
 126  *              Add an ip header to a skbuff and send it out.
 127  *
 128  */
 129 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 130                           __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
 131 {
 132         struct inet_sock *inet = inet_sk(sk);
 133         struct rtable *rt = skb_rtable(skb);
 134         struct iphdr *iph;
 135
 136         /* Build the IP header. */
 137         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
 138         skb_reset_network_header(skb);
 139         iph = ip_hdr(skb);
 140         iph->version  = 4;
 141         iph->ihl      = 5;
 142         iph->tos      = inet->tos;
 143         if (ip_dont_fragment(sk, &rt->dst))
 144                 iph->frag_off = htons(IP_DF);
 145         else
 146                 iph->frag_off = 0;
 147         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 148         iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
 149         iph->saddr    = saddr;
 150         iph->protocol = sk->sk_protocol;
 151         ip_select_ident(iph, &rt->dst, sk);
 152
 153         if (opt && opt->opt.optlen) {
 154                 iph->ihl += opt->opt.optlen>>2;
 155                 ip_options_build(skb, &opt->opt, daddr, rt, 0);
 156         }
 157
 158         skb->priority = sk->sk_priority;
 159         skb->mark = sk->sk_mark;
 160
 161         /* Send it out. */
 162         return ip_local_out(skb);
 163 }
 164 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 165
 166 static inline int ip_finish_output2(struct sk_buff *skb)
 167 {
 168         struct dst_entry *dst = skb_dst(skb);
 169         struct rtable *rt = (struct rtable *)dst;
 170         struct net_device *dev = dst->dev;
 171         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 172         struct neighbour *neigh;
 173         u32 nexthop;
 174
 175         if (rt->rt_type == RTN_MULTICAST) {
 176                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 177         } else if (rt->rt_type == RTN_BROADCAST)
 178                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 179
 180         /* Be paranoid, rather than too clever. */
 181         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 182                 struct sk_buff *skb2;
 183
 184                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 185                 if (skb2 == NULL) {
 186                         kfree_skb(skb);
 187                         return -ENOMEM;
 188                 }
 189                 if (skb->sk)
 190                         skb_set_owner_w(skb2, skb->sk);
 191                 consume_skb(skb);
 192                 skb = skb2;
 193         }
 194
 195         rcu_read_lock_bh();
 196         nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
 197         neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
 198         if (unlikely(!neigh))
 199                 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
 200         if (!IS_ERR(neigh)) {
 201                 int res = dst_neigh_output(dst, neigh, skb);
 202
 203                 rcu_read_unlock_bh();
 204                 return res;
 205         }
 206         rcu_read_unlock_bh();
 207
 208         net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
 209                             __func__);
 210         kfree_skb(skb);
 211         return -EINVAL;
 212 }
 213
 214 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 215 {
 216         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 217
 218         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 219                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 220 }
 221
 222 static int ip_finish_output(struct sk_buff *skb)
 223 {
 224 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 225         /* Policy lookup after SNAT yielded a new policy */
 226         if (skb_dst(skb)->xfrm != NULL) {
 227                 IPCB(skb)->flags |= IPSKB_REROUTED;
 228                 return dst_output(skb);
 229         }
 230 #endif
 231         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 232                 return ip_fragment(skb, ip_finish_output2);
 233         else
 234                 return ip_finish_output2(skb);
 235 }
 236
 237 int ip_mc_output(struct sk_buff *skb)
 238 {
 239         struct sock *sk = skb->sk;
 240         struct rtable *rt = skb_rtable(skb);
 241         struct net_device *dev = rt->dst.dev;
 242
 243         /*
 244          *      If the indicated interface is up and running, send the packet.
 245          */
 246         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 247
 248         skb->dev = dev;
 249         skb->protocol = htons(ETH_P_IP);
 250
 251         /*
 252          *      Multicasts are looped back for other local users
 253          */
 254
 255         if (rt->rt_flags&RTCF_MULTICAST) {
 256                 if (sk_mc_loop(sk)
 257 #ifdef CONFIG_IP_MROUTE
 258                 /* Small optimization: do not loopback not local frames,
 259                    which returned after forwarding; they will be  dropped
 260                    by ip_mr_input in any case.
 261                    Note, that local frames are looped back to be delivered
 262                    to local recipients.
 263
 264                    This check is duplicated in ip_mr_input at the moment.
 265                  */
 266                     &&
 267                     ((rt->rt_flags & RTCF_LOCAL) ||
 268                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 269 #endif
 270                    ) {
 271                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 272                         if (newskb)
 273                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 274                                         newskb, NULL, newskb->dev,
 275                                         dev_loopback_xmit);
 276                 }
 277
 278                 /* Multicasts with ttl 0 must not go beyond the host */
 279
 280                 if (ip_hdr(skb)->ttl == 0) {
 281                         kfree_skb(skb);
 282                         return 0;
 283                 }
 284         }
 285
 286         if (rt->rt_flags&RTCF_BROADCAST) {
 287                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 288                 if (newskb)
 289                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 290                                 NULL, newskb->dev, dev_loopback_xmit);
 291         }
 292
 293         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 294                             skb->dev, ip_finish_output,
 295                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 296 }
 297
 298 int ip_output(struct sk_buff *skb)
 299 {
 300         struct net_device *dev = skb_dst(skb)->dev;
 301
 302         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 303
 304         skb->dev = dev;
 305         skb->protocol = htons(ETH_P_IP);
 306
 307         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 308                             ip_finish_output,
 309                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 310 }
 311
 312 /*
 313  * copy saddr and daddr, possibly using 64bit load/stores
 314  * Equivalent to :
 315  *   iph->saddr = fl4->saddr;
 316  *   iph->daddr = fl4->daddr;
 317  */
 318 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 319 {
 320         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 321                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 322         memcpy(&iph->saddr, &fl4->saddr,
 323                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 324 }
 325
 326 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
 327 {
 328         struct sock *sk = skb->sk;
 329         struct inet_sock *inet = inet_sk(sk);
 330         struct ip_options_rcu *inet_opt;
 331         struct flowi4 *fl4;
 332         struct rtable *rt;
 333         struct iphdr *iph;
 334         int res;
 335
 336         /* Skip all of this if the packet is already routed,
 337          * f.e. by something like SCTP.
 338          */
 339         rcu_read_lock();
 340         inet_opt = rcu_dereference(inet->inet_opt);
 341         fl4 = &fl->u.ip4;
 342         rt = skb_rtable(skb);
 343         if (rt != NULL)
 344                 goto packet_routed;
 345
 346         /* Make sure we can route this packet. */
 347         rt = (struct rtable *)__sk_dst_check(sk, 0);
 348         if (rt == NULL) {
 349                 __be32 daddr;
 350
 351                 /* Use correct destination address if we have options. */
 352                 daddr = inet->inet_daddr;
 353                 if (inet_opt && inet_opt->opt.srr)
 354                         daddr = inet_opt->opt.faddr;
 355
 356                 /* If this fails, retransmit mechanism of transport layer will
 357                  * keep trying until route appears or the connection times
 358                  * itself out.
 359                  */
 360                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 361                                            daddr, inet->inet_saddr,
 362                                            inet->inet_dport,
 363                                            inet->inet_sport,
 364                                            sk->sk_protocol,
 365                                            RT_CONN_FLAGS(sk),
 366                                            sk->sk_bound_dev_if);
 367                 if (IS_ERR(rt))
 368                         goto no_route;
 369                 sk_setup_caps(sk, &rt->dst);
 370         }
 371         skb_dst_set_noref(skb, &rt->dst);
 372
 373 packet_routed:
 374         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
 375                 goto no_route;
 376
 377         /* OK, we know where to send it, allocate and build IP header. */
 378         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 379         skb_reset_network_header(skb);
 380         iph = ip_hdr(skb);
 381         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 382         if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
 383                 iph->frag_off = htons(IP_DF);
 384         else
 385                 iph->frag_off = 0;
 386         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 387         iph->protocol = sk->sk_protocol;
 388         ip_copy_addrs(iph, fl4);
 389
 390         /* Transport layer set skb->h.foo itself. */
 391
 392         if (inet_opt && inet_opt->opt.optlen) {
 393                 iph->ihl += inet_opt->opt.optlen >> 2;
 394                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 395         }
 396
 397         ip_select_ident_more(iph, &rt->dst, sk,
 398                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 399
 400         skb->priority = sk->sk_priority;
 401         skb->mark = sk->sk_mark;
 402
 403         res = ip_local_out(skb);
 404         rcu_read_unlock();
 405         return res;
 406
 407 no_route:
 408         rcu_read_unlock();
 409         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 410         kfree_skb(skb);
 411         return -EHOSTUNREACH;
 412 }
 413 EXPORT_SYMBOL(ip_queue_xmit);
 414
 415
 416 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 417 {
 418         to->pkt_type = from->pkt_type;
 419         to->priority = from->priority;
 420         to->protocol = from->protocol;
 421         skb_dst_drop(to);
 422         skb_dst_copy(to, from);
 423         to->dev = from->dev;
 424         to->mark = from->mark;
 425
 426         /* Copy the flags to each fragment. */
 427         IPCB(to)->flags = IPCB(from)->flags;
 428
 429 #ifdef CONFIG_NET_SCHED
 430         to->tc_index = from->tc_index;
 431 #endif
 432         nf_copy(to, from);
 433 #if defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 434     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 435         to->nf_trace = from->nf_trace;
 436 #endif
 437 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 438         to->ipvs_property = from->ipvs_property;
 439 #endif
 440         skb_copy_secmark(to, from);
 441 }
 442
 443 /*
 444  *      This IP datagram is too large to be sent in one piece.  Break it up into
 445  *      smaller pieces (each of size equal to IP header plus
 446  *      a block of the data of the original IP data part) that will yet fit in a
 447  *      single device frame, and queue such a frame for sending.
 448  */
 449
 450 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 451 {
 452         struct iphdr *iph;
 453         int ptr;
 454         struct net_device *dev;
 455         struct sk_buff *skb2;
 456         unsigned int mtu, hlen, left, len, ll_rs;
 457         int offset;
 458         __be16 not_last_frag;
 459         struct rtable *rt = skb_rtable(skb);
 460         int err = 0;
 461
 462         dev = rt->dst.dev;
 463
 464         /*
 465          *      Point into the IP datagram header.
 466          */
 467
 468         iph = ip_hdr(skb);
 469
 470         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->local_df) ||
 471                      (IPCB(skb)->frag_max_size &&
 472                       IPCB(skb)->frag_max_size > dst_mtu(&rt->dst)))) {
 473                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 474                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 475                           htonl(ip_skb_dst_mtu(skb)));
 476                 kfree_skb(skb);
 477                 return -EMSGSIZE;
 478         }
 479
 480         /*
 481          *      Setup starting values.
 482          */
 483
 484         hlen = iph->ihl * 4;
 485         mtu = dst_mtu(&rt->dst) - hlen; /* Size of data space */
 486 #ifdef CONFIG_BRIDGE_NETFILTER
 487         if (skb->nf_bridge)
 488                 mtu -= nf_bridge_mtu_reduction(skb);
 489 #endif
 490         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 491
 492         /* When frag_list is given, use it. First, check its validity:
 493          * some transformers could create wrong frag_list or break existing
 494          * one, it is not prohibited. In this case fall back to copying.
 495          *
 496          * LATER: this step can be merged to real generation of fragments,
 497          * we can switch to copy when see the first bad fragment.
 498          */
 499         if (skb_has_frag_list(skb)) {
 500                 struct sk_buff *frag, *frag2;
 501                 int first_len = skb_pagelen(skb);
 502
 503                 if (first_len - hlen > mtu ||
 504                     ((first_len - hlen) & 7) ||
 505                     ip_is_fragment(iph) ||
 506                     skb_cloned(skb))
 507                         goto slow_path;
 508
 509                 skb_walk_frags(skb, frag) {
 510                         /* Correct geometry. */
 511                         if (frag->len > mtu ||
 512                             ((frag->len & 7) && frag->next) ||
 513                             skb_headroom(frag) < hlen)
 514                                 goto slow_path_clean;
 515
 516                         /* Partially cloned skb? */
 517                         if (skb_shared(frag))
 518                                 goto slow_path_clean;
 519
 520                         BUG_ON(frag->sk);
 521                         if (skb->sk) {
 522                                 frag->sk = skb->sk;
 523                                 frag->destructor = sock_wfree;
 524                         }
 525                         skb->truesize -= frag->truesize;
 526                 }
 527
 528                 /* Everything is OK. Generate! */
 529
 530                 err = 0;
 531                 offset = 0;
 532                 frag = skb_shinfo(skb)->frag_list;
 533                 skb_frag_list_init(skb);
 534                 skb->data_len = first_len - skb_headlen(skb);
 535                 skb->len = first_len;
 536                 iph->tot_len = htons(first_len);
 537                 iph->frag_off = htons(IP_MF);
 538                 ip_send_check(iph);
 539
 540                 for (;;) {
 541                         /* Prepare header of the next frame,
 542                          * before previous one went down. */
 543                         if (frag) {
 544                                 frag->ip_summed = CHECKSUM_NONE;
 545                                 skb_reset_transport_header(frag);
 546                                 __skb_push(frag, hlen);
 547                                 skb_reset_network_header(frag);
 548                                 memcpy(skb_network_header(frag), iph, hlen);
 549                                 iph = ip_hdr(frag);
 550                                 iph->tot_len = htons(frag->len);
 551                                 ip_copy_metadata(frag, skb);
 552                                 if (offset == 0)
 553                                         ip_options_fragment(frag);
 554                                 offset += skb->len - hlen;
 555                                 iph->frag_off = htons(offset>>3);
 556                                 if (frag->next != NULL)
 557                                         iph->frag_off |= htons(IP_MF);
 558                                 /* Ready, complete checksum */
 559                                 ip_send_check(iph);
 560                         }
 561
 562                         err = output(skb);
 563
 564                         if (!err)
 565                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 566                         if (err || !frag)
 567                                 break;
 568
 569                         skb = frag;
 570                         frag = skb->next;
 571                         skb->next = NULL;
 572                 }
 573
 574                 if (err == 0) {
 575                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 576                         return 0;
 577                 }
 578
 579                 while (frag) {
 580                         skb = frag->next;
 581                         kfree_skb(frag);
 582                         frag = skb;
 583                 }
 584                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 585                 return err;
 586
 587 slow_path_clean:
 588                 skb_walk_frags(skb, frag2) {
 589                         if (frag2 == frag)
 590                                 break;
 591                         frag2->sk = NULL;
 592                         frag2->destructor = NULL;
 593                         skb->truesize += frag2->truesize;
 594                 }
 595         }
 596
 597 slow_path:
 598         /* for offloaded checksums cleanup checksum before fragmentation */
 599         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 600                 goto fail;
 601         iph = ip_hdr(skb);
 602
 603         left = skb->len - hlen;         /* Space per frame */
 604         ptr = hlen;             /* Where to start from */
 605
 606         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 607          * we need to make room for the encapsulating header
 608          */
 609         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 610
 611         /*
 612          *      Fragment the datagram.
 613          */
 614
 615         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 616         not_last_frag = iph->frag_off & htons(IP_MF);
 617
 618         /*
 619          *      Keep copying data until we run out.
 620          */
 621
 622         while (left > 0) {
 623                 len = left;
 624                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 625                 if (len > mtu)
 626                         len = mtu;
 627                 /* IF: we are not sending up to and including the packet end
 628                    then align the next start on an eight byte boundary */
 629                 if (len < left) {
 630                         len &= ~7;
 631                 }
 632                 /*
 633                  *      Allocate buffer.
 634                  */
 635
 636                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 637                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 638                         err = -ENOMEM;
 639                         goto fail;
 640                 }
 641
 642                 /*
 643                  *      Set up data on packet
 644                  */
 645
 646                 ip_copy_metadata(skb2, skb);
 647                 skb_reserve(skb2, ll_rs);
 648                 skb_put(skb2, len + hlen);
 649                 skb_reset_network_header(skb2);
 650                 skb2->transport_header = skb2->network_header + hlen;
 651
 652                 /*
 653                  *      Charge the memory for the fragment to any owner
 654                  *      it might possess
 655                  */
 656
 657                 if (skb->sk)
 658                         skb_set_owner_w(skb2, skb->sk);
 659
 660                 /*
 661                  *      Copy the packet header into the new buffer.
 662                  */
 663
 664                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 665
 666                 /*
 667                  *      Copy a block of the IP datagram.
 668                  */
 669                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 670                         BUG();
 671                 left -= len;
 672
 673                 /*
 674                  *      Fill in the new header fields.
 675                  */
 676                 iph = ip_hdr(skb2);
 677                 iph->frag_off = htons((offset >> 3));
 678
 679                 /* ANK: dirty, but effective trick. Upgrade options only if
 680                  * the segment to be fragmented was THE FIRST (otherwise,
 681                  * options are already fixed) and make it ONCE
 682                  * on the initial skb, so that all the following fragments
 683                  * will inherit fixed options.
 684                  */
 685                 if (offset == 0)
 686                         ip_options_fragment(skb);
 687
 688                 /*
 689                  *      Added AC : If we are fragmenting a fragment that's not the
 690                  *                 last fragment then keep MF on each bit
 691                  */
 692                 if (left > 0 || not_last_frag)
 693                         iph->frag_off |= htons(IP_MF);
 694                 ptr += len;
 695                 offset += len;
 696
 697                 /*
 698                  *      Put this fragment into the sending queue.
 699                  */
 700                 iph->tot_len = htons(len + hlen);
 701
 702                 ip_send_check(iph);
 703
 704                 err = output(skb2);
 705                 if (err)
 706                         goto fail;
 707
 708                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 709         }
 710         consume_skb(skb);
 711         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 712         return err;
 713
 714 fail:
 715         kfree_skb(skb);
 716         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 717         return err;
 718 }
 719 EXPORT_SYMBOL(ip_fragment);
 720
 721 int
 722 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 723 {
 724         struct iovec *iov = from;
 725
 726         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 727                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 728                         return -EFAULT;
 729         } else {
 730                 __wsum csum = 0;
 731                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 732                         return -EFAULT;
 733                 skb->csum = csum_block_add(skb->csum, csum, odd);
 734         }
 735         return 0;
 736 }
 737 EXPORT_SYMBOL(ip_generic_getfrag);
 738
 739 static inline __wsum
 740 csum_page(struct page *page, int offset, int copy)
 741 {
 742         char *kaddr;
 743         __wsum csum;
 744         kaddr = kmap(page);
 745         csum = csum_partial(kaddr + offset, copy, 0);
 746         kunmap(page);
 747         return csum;
 748 }
 749
 750 static inline int ip_ufo_append_data(struct sock *sk,
 751                         struct sk_buff_head *queue,
 752                         int getfrag(void *from, char *to, int offset, int len,
 753                                int odd, struct sk_buff *skb),
 754                         void *from, int length, int hh_len, int fragheaderlen,
 755                         int transhdrlen, int maxfraglen, unsigned int flags)
 756 {
 757         struct sk_buff *skb;
 758         int err;
 759
 760         /* There is support for UDP fragmentation offload by network
 761          * device, so create one single skb packet containing complete
 762          * udp datagram
 763          */
 764         if ((skb = skb_peek_tail(queue)) == NULL) {
 765                 skb = sock_alloc_send_skb(sk,
 766                         hh_len + fragheaderlen + transhdrlen + 20,
 767                         (flags & MSG_DONTWAIT), &err);
 768
 769                 if (skb == NULL)
 770                         return err;
 771
 772                 /* reserve space for Hardware header */
 773                 skb_reserve(skb, hh_len);
 774
 775                 /* create space for UDP/IP header */
 776                 skb_put(skb, fragheaderlen + transhdrlen);
 777
 778                 /* initialize network header pointer */
 779                 skb_reset_network_header(skb);
 780
 781                 /* initialize protocol header pointer */
 782                 skb->transport_header = skb->network_header + fragheaderlen;
 783
 784                 skb->ip_summed = CHECKSUM_PARTIAL;
 785                 skb->csum = 0;
 786
 787                 /* specify the length of each IP datagram fragment */
 788                 skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 789                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 790                 __skb_queue_tail(queue, skb);
 791         }
 792
 793         return skb_append_datato_frags(sk, skb, getfrag, from,
 794                                        (length - transhdrlen));
 795 }
 796
 797 static int __ip_append_data(struct sock *sk,
 798                             struct flowi4 *fl4,
 799                             struct sk_buff_head *queue,
 800                             struct inet_cork *cork,
 801                             struct page_frag *pfrag,
 802                             int getfrag(void *from, char *to, int offset,
 803                                         int len, int odd, struct sk_buff *skb),
 804                             void *from, int length, int transhdrlen,
 805                             unsigned int flags)
 806 {
 807         struct inet_sock *inet = inet_sk(sk);
 808         struct sk_buff *skb;
 809
 810         struct ip_options *opt = cork->opt;
 811         int hh_len;
 812         int exthdrlen;
 813         int mtu;
 814         int copy;
 815         int err;
 816         int offset = 0;
 817         unsigned int maxfraglen, fragheaderlen;
 818         int csummode = CHECKSUM_NONE;
 819         struct rtable *rt = (struct rtable *)cork->dst;
 820
 821         skb = skb_peek_tail(queue);
 822
 823         exthdrlen = !skb ? rt->dst.header_len : 0;
 824         mtu = cork->fragsize;
 825
 826         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 827
 828         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 829         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 830
 831         if (cork->length + length > 0xFFFF - fragheaderlen) {
 832                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 833                                mtu-exthdrlen);
 834                 return -EMSGSIZE;
 835         }
 836
 837         /*
 838          * transhdrlen > 0 means that this is the first fragment and we wish
 839          * it won't be fragmented in the future.
 840          */
 841         if (transhdrlen &&
 842             length + fragheaderlen <= mtu &&
 843             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 844             !exthdrlen)
 845                 csummode = CHECKSUM_PARTIAL;
 846
 847         cork->length += length;
 848         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 849             (sk->sk_protocol == IPPROTO_UDP) &&
 850             (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
 851                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 852                                          hh_len, fragheaderlen, transhdrlen,
 853                                          maxfraglen, flags);
 854                 if (err)
 855                         goto error;
 856                 return 0;
 857         }
 858
 859         /* So, what's going on in the loop below?
 860          *
 861          * We use calculated fragment length to generate chained skb,
 862          * each of segments is IP fragment ready for sending to network after
 863          * adding appropriate IP header.
 864          */
 865
 866         if (!skb)
 867                 goto alloc_new_skb;
 868
 869         while (length > 0) {
 870                 /* Check if the remaining data fits into current packet. */
 871                 copy = mtu - skb->len;
 872                 if (copy < length)
 873                         copy = maxfraglen - skb->len;
 874                 if (copy <= 0) {
 875                         char *data;
 876                         unsigned int datalen;
 877                         unsigned int fraglen;
 878                         unsigned int fraggap;
 879                         unsigned int alloclen;
 880                         struct sk_buff *skb_prev;
 881 alloc_new_skb:
 882                         skb_prev = skb;
 883                         if (skb_prev)
 884                                 fraggap = skb_prev->len - maxfraglen;
 885                         else
 886                                 fraggap = 0;
 887
 888                         /*
 889                          * If remaining data exceeds the mtu,
 890                          * we know we need more fragment(s).
 891                          */
 892                         datalen = length + fraggap;
 893                         if (datalen > mtu - fragheaderlen)
 894                                 datalen = maxfraglen - fragheaderlen;
 895                         fraglen = datalen + fragheaderlen;
 896
 897                         if ((flags & MSG_MORE) &&
 898                             !(rt->dst.dev->features&NETIF_F_SG))
 899                                 alloclen = mtu;
 900                         else
 901                                 alloclen = fraglen;
 902
 903                         alloclen += exthdrlen;
 904
 905                         /* The last fragment gets additional space at tail.
 906                          * Note, with MSG_MORE we overallocate on fragments,
 907                          * because we have no idea what fragment will be
 908                          * the last.
 909                          */
 910                         if (datalen == length + fraggap)
 911                                 alloclen += rt->dst.trailer_len;
 912
 913                         if (transhdrlen) {
 914                                 skb = sock_alloc_send_skb(sk,
 915                                                 alloclen + hh_len + 15,
 916                                                 (flags & MSG_DONTWAIT), &err);
 917                         } else {
 918                                 skb = NULL;
 919                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 920                                     2 * sk->sk_sndbuf)
 921                                         skb = sock_wmalloc(sk,
 922                                                            alloclen + hh_len + 15, 1,
 923                                                            sk->sk_allocation);
 924                                 if (unlikely(skb == NULL))
 925                                         err = -ENOBUFS;
 926                                 else
 927                                         /* only the initial fragment is
 928                                            time stamped */
 929                                         cork->tx_flags = 0;
 930                         }
 931                         if (skb == NULL)
 932                                 goto error;
 933
 934                         /*
 935                          *      Fill in the control structures
 936                          */
 937                         skb->ip_summed = csummode;
 938                         skb->csum = 0;
 939                         skb_reserve(skb, hh_len);
 940                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 941
 942                         /*
 943                          *      Find where to start putting bytes.
 944                          */
 945                         data = skb_put(skb, fraglen + exthdrlen);
 946                         skb_set_network_header(skb, exthdrlen);
 947                         skb->transport_header = (skb->network_header +
 948                                                  fragheaderlen);
 949                         data += fragheaderlen + exthdrlen;
 950
 951                         if (fraggap) {
 952                                 skb->csum = skb_copy_and_csum_bits(
 953                                         skb_prev, maxfraglen,
 954                                         data + transhdrlen, fraggap, 0);
 955                                 skb_prev->csum = csum_sub(skb_prev->csum,
 956                                                           skb->csum);
 957                                 data += fraggap;
 958                                 pskb_trim_unique(skb_prev, maxfraglen);
 959                         }
 960
 961                         copy = datalen - transhdrlen - fraggap;
 962                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 963                                 err = -EFAULT;
 964                                 kfree_skb(skb);
 965                                 goto error;
 966                         }
 967
 968                         offset += copy;
 969                         length -= datalen - fraggap;
 970                         transhdrlen = 0;
 971                         exthdrlen = 0;
 972                         csummode = CHECKSUM_NONE;
 973
 974                         /*
 975                          * Put the packet on the pending queue.
 976                          */
 977                         __skb_queue_tail(queue, skb);
 978                         continue;
 979                 }
 980
 981                 if (copy > length)
 982                         copy = length;
 983
 984                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
 985                         unsigned int off;
 986
 987                         off = skb->len;
 988                         if (getfrag(from, skb_put(skb, copy),
 989                                         offset, copy, off, skb) < 0) {
 990                                 __skb_trim(skb, off);
 991                                 err = -EFAULT;
 992                                 goto error;
 993                         }
 994                 } else {
 995                         int i = skb_shinfo(skb)->nr_frags;
 996
 997                         err = -ENOMEM;
 998                         if (!sk_page_frag_refill(sk, pfrag))
 999                                 goto error;
1000
1001                         if (!skb_can_coalesce(skb, i, pfrag->page,
1002                                               pfrag->offset)) {
1003                                 err = -EMSGSIZE;
1004                                 if (i == MAX_SKB_FRAGS)
1005                                         goto error;
1006
1007                                 __skb_fill_page_desc(skb, i, pfrag->page,
1008                                                      pfrag->offset, 0);
1009                                 skb_shinfo(skb)->nr_frags = ++i;
1010                                 get_page(pfrag->page);
1011                         }
1012                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1013                         if (getfrag(from,
1014                                     page_address(pfrag->page) + pfrag->offset,
1015                                     offset, copy, skb->len, skb) < 0)
1016                                 goto error_efault;
1017
1018                         pfrag->offset += copy;
1019                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1020                         skb->len += copy;
1021                         skb->data_len += copy;
1022                         skb->truesize += copy;
1023                         atomic_add(copy, &sk->sk_wmem_alloc);
1024                 }
1025                 offset += copy;
1026                 length -= copy;
1027         }
1028
1029         return 0;
1030
1031 error_efault:
1032         err = -EFAULT;
1033 error:
1034         cork->length -= length;
1035         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1036         return err;
1037 }
1038
1039 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1040                          struct ipcm_cookie *ipc, struct rtable **rtp)
1041 {
1042         struct inet_sock *inet = inet_sk(sk);
1043         struct ip_options_rcu *opt;
1044         struct rtable *rt;
1045
1046         /*
1047          * setup for corking.
1048          */
1049         opt = ipc->opt;
1050         if (opt) {
1051                 if (cork->opt == NULL) {
1052                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1053                                             sk->sk_allocation);
1054                         if (unlikely(cork->opt == NULL))
1055                                 return -ENOBUFS;
1056                 }
1057                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1058                 cork->flags |= IPCORK_OPT;
1059                 cork->addr = ipc->addr;
1060         }
1061         rt = *rtp;
1062         if (unlikely(!rt))
1063                 return -EFAULT;
1064         /*
1065          * We steal reference to this route, caller should not release it
1066          */
1067         *rtp = NULL;
1068         cork->fragsize = inet->pmtudisc == IP_PMTUDISC_PROBE ?
1069                          rt->dst.dev->mtu : dst_mtu(&rt->dst);
1070         cork->dst = &rt->dst;
1071         cork->length = 0;
1072         cork->tx_flags = ipc->tx_flags;
1073
1074         return 0;
1075 }
1076
1077 /*
1078  *      ip_append_data() and ip_append_page() can make one large IP datagram
1079  *      from many pieces of data. Each pieces will be holded on the socket
1080  *      until ip_push_pending_frames() is called. Each piece can be a page
1081  *      or non-page data.
1082  *
1083  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1084  *      this interface potentially.
1085  *
1086  *      LATER: length must be adjusted by pad at tail, when it is required.
1087  */
1088 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1089                    int getfrag(void *from, char *to, int offset, int len,
1090                                int odd, struct sk_buff *skb),
1091                    void *from, int length, int transhdrlen,
1092                    struct ipcm_cookie *ipc, struct rtable **rtp,
1093                    unsigned int flags)
1094 {
1095         struct inet_sock *inet = inet_sk(sk);
1096         int err;
1097
1098         if (flags&MSG_PROBE)
1099                 return 0;
1100
1101         if (skb_queue_empty(&sk->sk_write_queue)) {
1102                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1103                 if (err)
1104                         return err;
1105         } else {
1106                 transhdrlen = 0;
1107         }
1108
1109         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1110                                 sk_page_frag(sk), getfrag,
1111                                 from, length, transhdrlen, flags);
1112 }
1113
1114 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1115                        int offset, size_t size, int flags)
1116 {
1117         struct inet_sock *inet = inet_sk(sk);
1118         struct sk_buff *skb;
1119         struct rtable *rt;
1120         struct ip_options *opt = NULL;
1121         struct inet_cork *cork;
1122         int hh_len;
1123         int mtu;
1124         int len;
1125         int err;
1126         unsigned int maxfraglen, fragheaderlen, fraggap;
1127
1128         if (inet->hdrincl)
1129                 return -EPERM;
1130
1131         if (flags&MSG_PROBE)
1132                 return 0;
1133
1134         if (skb_queue_empty(&sk->sk_write_queue))
1135                 return -EINVAL;
1136
1137         cork = &inet->cork.base;
1138         rt = (struct rtable *)cork->dst;
1139         if (cork->flags & IPCORK_OPT)
1140                 opt = cork->opt;
1141
1142         if (!(rt->dst.dev->features&NETIF_F_SG))
1143                 return -EOPNOTSUPP;
1144
1145         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1146         mtu = cork->fragsize;
1147
1148         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1149         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1150
1151         if (cork->length + size > 0xFFFF - fragheaderlen) {
1152                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport, mtu);
1153                 return -EMSGSIZE;
1154         }
1155
1156         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1157                 return -EINVAL;
1158
1159         cork->length += size;
1160         if ((size + skb->len > mtu) &&
1161             (sk->sk_protocol == IPPROTO_UDP) &&
1162             (rt->dst.dev->features & NETIF_F_UFO)) {
1163                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1164                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1165         }
1166
1167
1168         while (size > 0) {
1169                 int i;
1170
1171                 if (skb_is_gso(skb))
1172                         len = size;
1173                 else {
1174
1175                         /* Check if the remaining data fits into current packet. */
1176                         len = mtu - skb->len;
1177                         if (len < size)
1178                                 len = maxfraglen - skb->len;
1179                 }
1180                 if (len <= 0) {
1181                         struct sk_buff *skb_prev;
1182                         int alloclen;
1183
1184                         skb_prev = skb;
1185                         fraggap = skb_prev->len - maxfraglen;
1186
1187                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1188                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1189                         if (unlikely(!skb)) {
1190                                 err = -ENOBUFS;
1191                                 goto error;
1192                         }
1193
1194                         /*
1195                          *      Fill in the control structures
1196                          */
1197                         skb->ip_summed = CHECKSUM_NONE;
1198                         skb->csum = 0;
1199                         skb_reserve(skb, hh_len);
1200
1201                         /*
1202                          *      Find where to start putting bytes.
1203                          */
1204                         skb_put(skb, fragheaderlen + fraggap);
1205                         skb_reset_network_header(skb);
1206                         skb->transport_header = (skb->network_header +
1207                                                  fragheaderlen);
1208                         if (fraggap) {
1209                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1210                                                                    maxfraglen,
1211                                                     skb_transport_header(skb),
1212                                                                    fraggap, 0);
1213                                 skb_prev->csum = csum_sub(skb_prev->csum,
1214                                                           skb->csum);
1215                                 pskb_trim_unique(skb_prev, maxfraglen);
1216                         }
1217
1218                         /*
1219                          * Put the packet on the pending queue.
1220                          */
1221                         __skb_queue_tail(&sk->sk_write_queue, skb);
1222                         continue;
1223                 }
1224
1225                 i = skb_shinfo(skb)->nr_frags;
1226                 if (len > size)
1227                         len = size;
1228                 if (skb_can_coalesce(skb, i, page, offset)) {
1229                         skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1230                 } else if (i < MAX_SKB_FRAGS) {
1231                         get_page(page);
1232                         skb_fill_page_desc(skb, i, page, offset, len);
1233                 } else {
1234                         err = -EMSGSIZE;
1235                         goto error;
1236                 }
1237
1238                 if (skb->ip_summed == CHECKSUM_NONE) {
1239                         __wsum csum;
1240                         csum = csum_page(page, offset, len);
1241                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1242                 }
1243
1244                 skb->len += len;
1245                 skb->data_len += len;
1246                 skb->truesize += len;
1247                 atomic_add(len, &sk->sk_wmem_alloc);
1248                 offset += len;
1249                 size -= len;
1250         }
1251         return 0;
1252
1253 error:
1254         cork->length -= size;
1255         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1256         return err;
1257 }
1258
1259 static void ip_cork_release(struct inet_cork *cork)
1260 {
1261         cork->flags &= ~IPCORK_OPT;
1262         kfree(cork->opt);
1263         cork->opt = NULL;
1264         dst_release(cork->dst);
1265         cork->dst = NULL;
1266 }
1267
1268 /*
1269  *      Combined all pending IP fragments on the socket as one IP datagram
1270  *      and push them out.
1271  */
1272 struct sk_buff *__ip_make_skb(struct sock *sk,
1273                               struct flowi4 *fl4,
1274                               struct sk_buff_head *queue,
1275                               struct inet_cork *cork)
1276 {
1277         struct sk_buff *skb, *tmp_skb;
1278         struct sk_buff **tail_skb;
1279         struct inet_sock *inet = inet_sk(sk);
1280         struct net *net = sock_net(sk);
1281         struct ip_options *opt = NULL;
1282         struct rtable *rt = (struct rtable *)cork->dst;
1283         struct iphdr *iph;
1284         __be16 df = 0;
1285         __u8 ttl;
1286
1287         if ((skb = __skb_dequeue(queue)) == NULL)
1288                 goto out;
1289         tail_skb = &(skb_shinfo(skb)->frag_list);
1290
1291         /* move skb->data to ip header from ext header */
1292         if (skb->data < skb_network_header(skb))
1293                 __skb_pull(skb, skb_network_offset(skb));
1294         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1295                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1296                 *tail_skb = tmp_skb;
1297                 tail_skb = &(tmp_skb->next);
1298                 skb->len += tmp_skb->len;
1299                 skb->data_len += tmp_skb->len;
1300                 skb->truesize += tmp_skb->truesize;
1301                 tmp_skb->destructor = NULL;
1302                 tmp_skb->sk = NULL;
1303         }
1304
1305         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1306          * to fragment the frame generated here. No matter, what transforms
1307          * how transforms change size of the packet, it will come out.
1308          */
1309         if (inet->pmtudisc < IP_PMTUDISC_DO)
1310                 skb->local_df = 1;
1311
1312         /* DF bit is set when we want to see DF on outgoing frames.
1313          * If local_df is set too, we still allow to fragment this frame
1314          * locally. */
1315         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1316             (skb->len <= dst_mtu(&rt->dst) &&
1317              ip_dont_fragment(sk, &rt->dst)))
1318                 df = htons(IP_DF);
1319
1320         if (cork->flags & IPCORK_OPT)
1321                 opt = cork->opt;
1322
1323         if (rt->rt_type == RTN_MULTICAST)
1324                 ttl = inet->mc_ttl;
1325         else
1326                 ttl = ip_select_ttl(inet, &rt->dst);
1327
1328         iph = (struct iphdr *)skb->data;
1329         iph->version = 4;
1330         iph->ihl = 5;
1331         iph->tos = inet->tos;
1332         iph->frag_off = df;
1333         iph->ttl = ttl;
1334         iph->protocol = sk->sk_protocol;
1335         ip_copy_addrs(iph, fl4);
1336         ip_select_ident(iph, &rt->dst, sk);
1337
1338         if (opt) {
1339                 iph->ihl += opt->optlen>>2;
1340                 ip_options_build(skb, opt, cork->addr, rt, 0);
1341         }
1342
1343         skb->priority = sk->sk_priority;
1344         skb->mark = sk->sk_mark;
1345         /*
1346          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1347          * on dst refcount
1348          */
1349         cork->dst = NULL;
1350         skb_dst_set(skb, &rt->dst);
1351
1352         if (iph->protocol == IPPROTO_ICMP)
1353                 icmp_out_count(net, ((struct icmphdr *)
1354                         skb_transport_header(skb))->type);
1355
1356         ip_cork_release(cork);
1357 out:
1358         return skb;
1359 }
1360
1361 int ip_send_skb(struct net *net, struct sk_buff *skb)
1362 {
1363         int err;
1364
1365         err = ip_local_out(skb);
1366         if (err) {
1367                 if (err > 0)
1368                         err = net_xmit_errno(err);
1369                 if (err)
1370                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1371         }
1372
1373         return err;
1374 }
1375
1376 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1377 {
1378         struct sk_buff *skb;
1379
1380         skb = ip_finish_skb(sk, fl4);
1381         if (!skb)
1382                 return 0;
1383
1384         /* Netfilter gets whole the not fragmented skb. */
1385         return ip_send_skb(sock_net(sk), skb);
1386 }
1387
1388 /*
1389  *      Throw away all pending data on the socket.
1390  */
1391 static void __ip_flush_pending_frames(struct sock *sk,
1392                                       struct sk_buff_head *queue,
1393                                       struct inet_cork *cork)
1394 {
1395         struct sk_buff *skb;
1396
1397         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1398                 kfree_skb(skb);
1399
1400         ip_cork_release(cork);
1401 }
1402
1403 void ip_flush_pending_frames(struct sock *sk)
1404 {
1405         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1406 }
1407
1408 struct sk_buff *ip_make_skb(struct sock *sk,
1409                             struct flowi4 *fl4,
1410                             int getfrag(void *from, char *to, int offset,
1411                                         int len, int odd, struct sk_buff *skb),
1412                             void *from, int length, int transhdrlen,
1413                             struct ipcm_cookie *ipc, struct rtable **rtp,
1414                             unsigned int flags)
1415 {
1416         struct inet_cork cork;
1417         struct sk_buff_head queue;
1418         int err;
1419
1420         if (flags & MSG_PROBE)
1421                 return NULL;
1422
1423         __skb_queue_head_init(&queue);
1424
1425         cork.flags = 0;
1426         cork.addr = 0;
1427         cork.opt = NULL;
1428         err = ip_setup_cork(sk, &cork, ipc, rtp);
1429         if (err)
1430                 return ERR_PTR(err);
1431
1432         err = __ip_append_data(sk, fl4, &queue, &cork,
1433                                &current->task_frag, getfrag,
1434                                from, length, transhdrlen, flags);
1435         if (err) {
1436                 __ip_flush_pending_frames(sk, &queue, &cork);
1437                 return ERR_PTR(err);
1438         }
1439
1440         return __ip_make_skb(sk, fl4, &queue, &cork);
1441 }
1442
1443 /*
1444  *      Fetch data from kernel space and fill in checksum if needed.
1445  */
1446 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1447                               int len, int odd, struct sk_buff *skb)
1448 {
1449         __wsum csum;
1450
1451         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1452         skb->csum = csum_block_add(skb->csum, csum, odd);
1453         return 0;
1454 }
1455
1456 /*
1457  *      Generic function to send a packet as reply to another packet.
1458  *      Used to send some TCP resets/acks so far.
1459  *
1460  *      Use a fake percpu inet socket to avoid false sharing and contention.
1461  */
1462 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1463         .sk = {
1464                 .__sk_common = {
1465                         .skc_refcnt = ATOMIC_INIT(1),
1466                 },
1467                 .sk_wmem_alloc  = ATOMIC_INIT(1),
1468                 .sk_allocation  = GFP_ATOMIC,
1469                 .sk_flags       = (1UL << SOCK_USE_WRITE_QUEUE),
1470         },
1471         .pmtudisc       = IP_PMTUDISC_WANT,
1472         .uc_ttl         = -1,
1473 };
1474
1475 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1476                            __be32 saddr, const struct ip_reply_arg *arg,
1477                            unsigned int len)
1478 {
1479         struct ip_options_data replyopts;
1480         struct ipcm_cookie ipc;
1481         struct flowi4 fl4;
1482         struct rtable *rt = skb_rtable(skb);
1483         struct sk_buff *nskb;
1484         struct sock *sk;
1485         struct inet_sock *inet;
1486
1487         if (ip_options_echo(&replyopts.opt.opt, skb))
1488                 return;
1489
1490         ipc.addr = daddr;
1491         ipc.opt = NULL;
1492         ipc.tx_flags = 0;
1493
1494         if (replyopts.opt.opt.optlen) {
1495                 ipc.opt = &replyopts.opt;
1496
1497                 if (replyopts.opt.opt.srr)
1498                         daddr = replyopts.opt.opt.faddr;
1499         }
1500
1501         flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1502                            RT_TOS(arg->tos),
1503                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1504                            ip_reply_arg_flowi_flags(arg),
1505                            daddr, saddr,
1506                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1507         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1508         rt = ip_route_output_key(net, &fl4);
1509         if (IS_ERR(rt))
1510                 return;
1511
1512         inet = &get_cpu_var(unicast_sock);
1513
1514         inet->tos = arg->tos;
1515         sk = &inet->sk;
1516         sk->sk_priority = skb->priority;
1517         sk->sk_protocol = ip_hdr(skb)->protocol;
1518         sk->sk_bound_dev_if = arg->bound_dev_if;
1519         sock_net_set(sk, net);
1520         __skb_queue_head_init(&sk->sk_write_queue);
1521         sk->sk_sndbuf = sysctl_wmem_default;
1522         ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1523                        &ipc, &rt, MSG_DONTWAIT);
1524         nskb = skb_peek(&sk->sk_write_queue);
1525         if (nskb) {
1526                 if (arg->csumoffset >= 0)
1527                         *((__sum16 *)skb_transport_header(nskb) +
1528                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1529                                                                 arg->csum));
1530                 nskb->ip_summed = CHECKSUM_NONE;
1531                 skb_orphan(nskb);
1532                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1533                 ip_push_pending_frames(sk, &fl4);
1534         }
1535
1536         put_cpu_var(unicast_sock);
1537
1538         ip_rt_put(rt);
1539 }
1540
1541 void __init ip_init(void)
1542 {
1543         ip_rt_init();
1544         inet_initpeers();
1545
1546 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1547         igmp_mc_proc_init();
1548 #endif
1549 }