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 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(skb, &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 int ip_finish_output(struct sk_buff *skb)
 215 {
 216 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 217         /* Policy lookup after SNAT yielded a new policy */
 218         if (skb_dst(skb)->xfrm != NULL) {
 219                 IPCB(skb)->flags |= IPSKB_REROUTED;
 220                 return dst_output(skb);
 221         }
 222 #endif
 223         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 224                 return ip_fragment(skb, ip_finish_output2);
 225         else
 226                 return ip_finish_output2(skb);
 227 }
 228
 229 int ip_mc_output(struct sk_buff *skb)
 230 {
 231         struct sock *sk = skb->sk;
 232         struct rtable *rt = skb_rtable(skb);
 233         struct net_device *dev = rt->dst.dev;
 234
 235         /*
 236          *      If the indicated interface is up and running, send the packet.
 237          */
 238         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 239
 240         skb->dev = dev;
 241         skb->protocol = htons(ETH_P_IP);
 242
 243         /*
 244          *      Multicasts are looped back for other local users
 245          */
 246
 247         if (rt->rt_flags&RTCF_MULTICAST) {
 248                 if (sk_mc_loop(sk)
 249 #ifdef CONFIG_IP_MROUTE
 250                 /* Small optimization: do not loopback not local frames,
 251                    which returned after forwarding; they will be  dropped
 252                    by ip_mr_input in any case.
 253                    Note, that local frames are looped back to be delivered
 254                    to local recipients.
 255
 256                    This check is duplicated in ip_mr_input at the moment.
 257                  */
 258                     &&
 259                     ((rt->rt_flags & RTCF_LOCAL) ||
 260                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 261 #endif
 262                    ) {
 263                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 264                         if (newskb)
 265                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 266                                         newskb, NULL, newskb->dev,
 267                                         dev_loopback_xmit);
 268                 }
 269
 270                 /* Multicasts with ttl 0 must not go beyond the host */
 271
 272                 if (ip_hdr(skb)->ttl == 0) {
 273                         kfree_skb(skb);
 274                         return 0;
 275                 }
 276         }
 277
 278         if (rt->rt_flags&RTCF_BROADCAST) {
 279                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 280                 if (newskb)
 281                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 282                                 NULL, newskb->dev, dev_loopback_xmit);
 283         }
 284
 285         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 286                             skb->dev, ip_finish_output,
 287                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 288 }
 289
 290 int ip_output(struct sk_buff *skb)
 291 {
 292         struct net_device *dev = skb_dst(skb)->dev;
 293
 294         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 295
 296         skb->dev = dev;
 297         skb->protocol = htons(ETH_P_IP);
 298
 299         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 300                             ip_finish_output,
 301                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 302 }
 303
 304 /*
 305  * copy saddr and daddr, possibly using 64bit load/stores
 306  * Equivalent to :
 307  *   iph->saddr = fl4->saddr;
 308  *   iph->daddr = fl4->daddr;
 309  */
 310 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 311 {
 312         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 313                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 314         memcpy(&iph->saddr, &fl4->saddr,
 315                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 316 }
 317
 318 /* Note: skb->sk can be different from sk, in case of tunnels */
 319 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
 320 {
 321         struct inet_sock *inet = inet_sk(sk);
 322         struct ip_options_rcu *inet_opt;
 323         struct flowi4 *fl4;
 324         struct rtable *rt;
 325         struct iphdr *iph;
 326         int res;
 327
 328         /* Skip all of this if the packet is already routed,
 329          * f.e. by something like SCTP.
 330          */
 331         rcu_read_lock();
 332         inet_opt = rcu_dereference(inet->inet_opt);
 333         fl4 = &fl->u.ip4;
 334         rt = skb_rtable(skb);
 335         if (rt != NULL)
 336                 goto packet_routed;
 337
 338         /* Make sure we can route this packet. */
 339         rt = (struct rtable *)__sk_dst_check(sk, 0);
 340         if (rt == NULL) {
 341                 __be32 daddr;
 342
 343                 /* Use correct destination address if we have options. */
 344                 daddr = inet->inet_daddr;
 345                 if (inet_opt && inet_opt->opt.srr)
 346                         daddr = inet_opt->opt.faddr;
 347
 348                 /* If this fails, retransmit mechanism of transport layer will
 349                  * keep trying until route appears or the connection times
 350                  * itself out.
 351                  */
 352                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 353                                            daddr, inet->inet_saddr,
 354                                            inet->inet_dport,
 355                                            inet->inet_sport,
 356                                            sk->sk_protocol,
 357                                            RT_CONN_FLAGS(sk),
 358                                            sk->sk_bound_dev_if);
 359                 if (IS_ERR(rt))
 360                         goto no_route;
 361                 sk_setup_caps(sk, &rt->dst);
 362         }
 363         skb_dst_set_noref(skb, &rt->dst);
 364
 365 packet_routed:
 366         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
 367                 goto no_route;
 368
 369         /* OK, we know where to send it, allocate and build IP header. */
 370         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 371         skb_reset_network_header(skb);
 372         iph = ip_hdr(skb);
 373         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 374         if (ip_dont_fragment(sk, &rt->dst) && !skb->local_df)
 375                 iph->frag_off = htons(IP_DF);
 376         else
 377                 iph->frag_off = 0;
 378         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 379         iph->protocol = sk->sk_protocol;
 380         ip_copy_addrs(iph, fl4);
 381
 382         /* Transport layer set skb->h.foo itself. */
 383
 384         if (inet_opt && inet_opt->opt.optlen) {
 385                 iph->ihl += inet_opt->opt.optlen >> 2;
 386                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 387         }
 388
 389         ip_select_ident_more(skb, &rt->dst, sk,
 390                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 391
 392         /* TODO : should we use skb->sk here instead of sk ? */
 393         skb->priority = sk->sk_priority;
 394         skb->mark = sk->sk_mark;
 395
 396         res = ip_local_out(skb);
 397         rcu_read_unlock();
 398         return res;
 399
 400 no_route:
 401         rcu_read_unlock();
 402         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 403         kfree_skb(skb);
 404         return -EHOSTUNREACH;
 405 }
 406 EXPORT_SYMBOL(ip_queue_xmit);
 407
 408
 409 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 410 {
 411         to->pkt_type = from->pkt_type;
 412         to->priority = from->priority;
 413         to->protocol = from->protocol;
 414         skb_dst_drop(to);
 415         skb_dst_copy(to, from);
 416         to->dev = from->dev;
 417         to->mark = from->mark;
 418
 419         /* Copy the flags to each fragment. */
 420         IPCB(to)->flags = IPCB(from)->flags;
 421
 422 #ifdef CONFIG_NET_SCHED
 423         to->tc_index = from->tc_index;
 424 #endif
 425         nf_copy(to, from);
 426 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 427         to->ipvs_property = from->ipvs_property;
 428 #endif
 429         skb_copy_secmark(to, from);
 430 }
 431
 432 /*
 433  *      This IP datagram is too large to be sent in one piece.  Break it up into
 434  *      smaller pieces (each of size equal to IP header plus
 435  *      a block of the data of the original IP data part) that will yet fit in a
 436  *      single device frame, and queue such a frame for sending.
 437  */
 438
 439 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 440 {
 441         struct iphdr *iph;
 442         int ptr;
 443         struct net_device *dev;
 444         struct sk_buff *skb2;
 445         unsigned int mtu, hlen, left, len, ll_rs;
 446         int offset;
 447         __be16 not_last_frag;
 448         struct rtable *rt = skb_rtable(skb);
 449         int err = 0;
 450
 451         dev = rt->dst.dev;
 452
 453         /*
 454          *      Point into the IP datagram header.
 455          */
 456
 457         iph = ip_hdr(skb);
 458
 459         mtu = ip_skb_dst_mtu(skb);
 460         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->local_df) ||
 461                      (IPCB(skb)->frag_max_size &&
 462                       IPCB(skb)->frag_max_size > mtu))) {
 463                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 464                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 465                           htonl(mtu));
 466                 kfree_skb(skb);
 467                 return -EMSGSIZE;
 468         }
 469
 470         /*
 471          *      Setup starting values.
 472          */
 473
 474         hlen = iph->ihl * 4;
 475         mtu = mtu - hlen;       /* Size of data space */
 476 #ifdef CONFIG_BRIDGE_NETFILTER
 477         if (skb->nf_bridge)
 478                 mtu -= nf_bridge_mtu_reduction(skb);
 479 #endif
 480         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 481
 482         /* When frag_list is given, use it. First, check its validity:
 483          * some transformers could create wrong frag_list or break existing
 484          * one, it is not prohibited. In this case fall back to copying.
 485          *
 486          * LATER: this step can be merged to real generation of fragments,
 487          * we can switch to copy when see the first bad fragment.
 488          */
 489         if (skb_has_frag_list(skb)) {
 490                 struct sk_buff *frag, *frag2;
 491                 int first_len = skb_pagelen(skb);
 492
 493                 if (first_len - hlen > mtu ||
 494                     ((first_len - hlen) & 7) ||
 495                     ip_is_fragment(iph) ||
 496                     skb_cloned(skb))
 497                         goto slow_path;
 498
 499                 skb_walk_frags(skb, frag) {
 500                         /* Correct geometry. */
 501                         if (frag->len > mtu ||
 502                             ((frag->len & 7) && frag->next) ||
 503                             skb_headroom(frag) < hlen)
 504                                 goto slow_path_clean;
 505
 506                         /* Partially cloned skb? */
 507                         if (skb_shared(frag))
 508                                 goto slow_path_clean;
 509
 510                         BUG_ON(frag->sk);
 511                         if (skb->sk) {
 512                                 frag->sk = skb->sk;
 513                                 frag->destructor = sock_wfree;
 514                         }
 515                         skb->truesize -= frag->truesize;
 516                 }
 517
 518                 /* Everything is OK. Generate! */
 519
 520                 err = 0;
 521                 offset = 0;
 522                 frag = skb_shinfo(skb)->frag_list;
 523                 skb_frag_list_init(skb);
 524                 skb->data_len = first_len - skb_headlen(skb);
 525                 skb->len = first_len;
 526                 iph->tot_len = htons(first_len);
 527                 iph->frag_off = htons(IP_MF);
 528                 ip_send_check(iph);
 529
 530                 for (;;) {
 531                         /* Prepare header of the next frame,
 532                          * before previous one went down. */
 533                         if (frag) {
 534                                 frag->ip_summed = CHECKSUM_NONE;
 535                                 skb_reset_transport_header(frag);
 536                                 __skb_push(frag, hlen);
 537                                 skb_reset_network_header(frag);
 538                                 memcpy(skb_network_header(frag), iph, hlen);
 539                                 iph = ip_hdr(frag);
 540                                 iph->tot_len = htons(frag->len);
 541                                 ip_copy_metadata(frag, skb);
 542                                 if (offset == 0)
 543                                         ip_options_fragment(frag);
 544                                 offset += skb->len - hlen;
 545                                 iph->frag_off = htons(offset>>3);
 546                                 if (frag->next != NULL)
 547                                         iph->frag_off |= htons(IP_MF);
 548                                 /* Ready, complete checksum */
 549                                 ip_send_check(iph);
 550                         }
 551
 552                         err = output(skb);
 553
 554                         if (!err)
 555                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 556                         if (err || !frag)
 557                                 break;
 558
 559                         skb = frag;
 560                         frag = skb->next;
 561                         skb->next = NULL;
 562                 }
 563
 564                 if (err == 0) {
 565                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 566                         return 0;
 567                 }
 568
 569                 while (frag) {
 570                         skb = frag->next;
 571                         kfree_skb(frag);
 572                         frag = skb;
 573                 }
 574                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 575                 return err;
 576
 577 slow_path_clean:
 578                 skb_walk_frags(skb, frag2) {
 579                         if (frag2 == frag)
 580                                 break;
 581                         frag2->sk = NULL;
 582                         frag2->destructor = NULL;
 583                         skb->truesize += frag2->truesize;
 584                 }
 585         }
 586
 587 slow_path:
 588         /* for offloaded checksums cleanup checksum before fragmentation */
 589         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 590                 goto fail;
 591         iph = ip_hdr(skb);
 592
 593         left = skb->len - hlen;         /* Space per frame */
 594         ptr = hlen;             /* Where to start from */
 595
 596         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 597          * we need to make room for the encapsulating header
 598          */
 599         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 600
 601         /*
 602          *      Fragment the datagram.
 603          */
 604
 605         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 606         not_last_frag = iph->frag_off & htons(IP_MF);
 607
 608         /*
 609          *      Keep copying data until we run out.
 610          */
 611
 612         while (left > 0) {
 613                 len = left;
 614                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 615                 if (len > mtu)
 616                         len = mtu;
 617                 /* IF: we are not sending up to and including the packet end
 618                    then align the next start on an eight byte boundary */
 619                 if (len < left) {
 620                         len &= ~7;
 621                 }
 622                 /*
 623                  *      Allocate buffer.
 624                  */
 625
 626                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 627                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 628                         err = -ENOMEM;
 629                         goto fail;
 630                 }
 631
 632                 /*
 633                  *      Set up data on packet
 634                  */
 635
 636                 ip_copy_metadata(skb2, skb);
 637                 skb_reserve(skb2, ll_rs);
 638                 skb_put(skb2, len + hlen);
 639                 skb_reset_network_header(skb2);
 640                 skb2->transport_header = skb2->network_header + hlen;
 641
 642                 /*
 643                  *      Charge the memory for the fragment to any owner
 644                  *      it might possess
 645                  */
 646
 647                 if (skb->sk)
 648                         skb_set_owner_w(skb2, skb->sk);
 649
 650                 /*
 651                  *      Copy the packet header into the new buffer.
 652                  */
 653
 654                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 655
 656                 /*
 657                  *      Copy a block of the IP datagram.
 658                  */
 659                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 660                         BUG();
 661                 left -= len;
 662
 663                 /*
 664                  *      Fill in the new header fields.
 665                  */
 666                 iph = ip_hdr(skb2);
 667                 iph->frag_off = htons((offset >> 3));
 668
 669                 /* ANK: dirty, but effective trick. Upgrade options only if
 670                  * the segment to be fragmented was THE FIRST (otherwise,
 671                  * options are already fixed) and make it ONCE
 672                  * on the initial skb, so that all the following fragments
 673                  * will inherit fixed options.
 674                  */
 675                 if (offset == 0)
 676                         ip_options_fragment(skb);
 677
 678                 /*
 679                  *      Added AC : If we are fragmenting a fragment that's not the
 680                  *                 last fragment then keep MF on each bit
 681                  */
 682                 if (left > 0 || not_last_frag)
 683                         iph->frag_off |= htons(IP_MF);
 684                 ptr += len;
 685                 offset += len;
 686
 687                 /*
 688                  *      Put this fragment into the sending queue.
 689                  */
 690                 iph->tot_len = htons(len + hlen);
 691
 692                 ip_send_check(iph);
 693
 694                 err = output(skb2);
 695                 if (err)
 696                         goto fail;
 697
 698                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 699         }
 700         consume_skb(skb);
 701         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 702         return err;
 703
 704 fail:
 705         kfree_skb(skb);
 706         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 707         return err;
 708 }
 709 EXPORT_SYMBOL(ip_fragment);
 710
 711 int
 712 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 713 {
 714         struct iovec *iov = from;
 715
 716         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 717                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 718                         return -EFAULT;
 719         } else {
 720                 __wsum csum = 0;
 721                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 722                         return -EFAULT;
 723                 skb->csum = csum_block_add(skb->csum, csum, odd);
 724         }
 725         return 0;
 726 }
 727 EXPORT_SYMBOL(ip_generic_getfrag);
 728
 729 static inline __wsum
 730 csum_page(struct page *page, int offset, int copy)
 731 {
 732         char *kaddr;
 733         __wsum csum;
 734         kaddr = kmap(page);
 735         csum = csum_partial(kaddr + offset, copy, 0);
 736         kunmap(page);
 737         return csum;
 738 }
 739
 740 static inline int ip_ufo_append_data(struct sock *sk,
 741                         struct sk_buff_head *queue,
 742                         int getfrag(void *from, char *to, int offset, int len,
 743                                int odd, struct sk_buff *skb),
 744                         void *from, int length, int hh_len, int fragheaderlen,
 745                         int transhdrlen, int maxfraglen, unsigned int flags)
 746 {
 747         struct sk_buff *skb;
 748         int err;
 749
 750         /* There is support for UDP fragmentation offload by network
 751          * device, so create one single skb packet containing complete
 752          * udp datagram
 753          */
 754         if ((skb = skb_peek_tail(queue)) == NULL) {
 755                 skb = sock_alloc_send_skb(sk,
 756                         hh_len + fragheaderlen + transhdrlen + 20,
 757                         (flags & MSG_DONTWAIT), &err);
 758
 759                 if (skb == NULL)
 760                         return err;
 761
 762                 /* reserve space for Hardware header */
 763                 skb_reserve(skb, hh_len);
 764
 765                 /* create space for UDP/IP header */
 766                 skb_put(skb, fragheaderlen + transhdrlen);
 767
 768                 /* initialize network header pointer */
 769                 skb_reset_network_header(skb);
 770
 771                 /* initialize protocol header pointer */
 772                 skb->transport_header = skb->network_header + fragheaderlen;
 773
 774                 skb->csum = 0;
 775
 776
 777                 __skb_queue_tail(queue, skb);
 778         } else if (skb_is_gso(skb)) {
 779                 goto append;
 780         }
 781
 782         skb->ip_summed = CHECKSUM_PARTIAL;
 783         /* specify the length of each IP datagram fragment */
 784         skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 785         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 786
 787 append:
 788         return skb_append_datato_frags(sk, skb, getfrag, from,
 789                                        (length - transhdrlen));
 790 }
 791
 792 static int __ip_append_data(struct sock *sk,
 793                             struct flowi4 *fl4,
 794                             struct sk_buff_head *queue,
 795                             struct inet_cork *cork,
 796                             struct page_frag *pfrag,
 797                             int getfrag(void *from, char *to, int offset,
 798                                         int len, int odd, struct sk_buff *skb),
 799                             void *from, int length, int transhdrlen,
 800                             unsigned int flags)
 801 {
 802         struct inet_sock *inet = inet_sk(sk);
 803         struct sk_buff *skb;
 804
 805         struct ip_options *opt = cork->opt;
 806         int hh_len;
 807         int exthdrlen;
 808         int mtu;
 809         int copy;
 810         int err;
 811         int offset = 0;
 812         unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
 813         int csummode = CHECKSUM_NONE;
 814         struct rtable *rt = (struct rtable *)cork->dst;
 815
 816         skb = skb_peek_tail(queue);
 817
 818         exthdrlen = !skb ? rt->dst.header_len : 0;
 819         mtu = cork->fragsize;
 820
 821         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 822
 823         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 824         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 825         maxnonfragsize = ip_sk_local_df(sk) ? 0xFFFF : mtu;
 826
 827         if (cork->length + length > maxnonfragsize - fragheaderlen) {
 828                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 829                                mtu - (opt ? opt->optlen : 0));
 830                 return -EMSGSIZE;
 831         }
 832
 833         /*
 834          * transhdrlen > 0 means that this is the first fragment and we wish
 835          * it won't be fragmented in the future.
 836          */
 837         if (transhdrlen &&
 838             length + fragheaderlen <= mtu &&
 839             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 840             !exthdrlen)
 841                 csummode = CHECKSUM_PARTIAL;
 842
 843         cork->length += length;
 844         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 845             (sk->sk_protocol == IPPROTO_UDP) &&
 846             (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len) {
 847                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 848                                          hh_len, fragheaderlen, transhdrlen,
 849                                          maxfraglen, flags);
 850                 if (err)
 851                         goto error;
 852                 return 0;
 853         }
 854
 855         /* So, what's going on in the loop below?
 856          *
 857          * We use calculated fragment length to generate chained skb,
 858          * each of segments is IP fragment ready for sending to network after
 859          * adding appropriate IP header.
 860          */
 861
 862         if (!skb)
 863                 goto alloc_new_skb;
 864
 865         while (length > 0) {
 866                 /* Check if the remaining data fits into current packet. */
 867                 copy = mtu - skb->len;
 868                 if (copy < length)
 869                         copy = maxfraglen - skb->len;
 870                 if (copy <= 0) {
 871                         char *data;
 872                         unsigned int datalen;
 873                         unsigned int fraglen;
 874                         unsigned int fraggap;
 875                         unsigned int alloclen;
 876                         struct sk_buff *skb_prev;
 877 alloc_new_skb:
 878                         skb_prev = skb;
 879                         if (skb_prev)
 880                                 fraggap = skb_prev->len - maxfraglen;
 881                         else
 882                                 fraggap = 0;
 883
 884                         /*
 885                          * If remaining data exceeds the mtu,
 886                          * we know we need more fragment(s).
 887                          */
 888                         datalen = length + fraggap;
 889                         if (datalen > mtu - fragheaderlen)
 890                                 datalen = maxfraglen - fragheaderlen;
 891                         fraglen = datalen + fragheaderlen;
 892
 893                         if ((flags & MSG_MORE) &&
 894                             !(rt->dst.dev->features&NETIF_F_SG))
 895                                 alloclen = mtu;
 896                         else
 897                                 alloclen = fraglen;
 898
 899                         alloclen += exthdrlen;
 900
 901                         /* The last fragment gets additional space at tail.
 902                          * Note, with MSG_MORE we overallocate on fragments,
 903                          * because we have no idea what fragment will be
 904                          * the last.
 905                          */
 906                         if (datalen == length + fraggap)
 907                                 alloclen += rt->dst.trailer_len;
 908
 909                         if (transhdrlen) {
 910                                 skb = sock_alloc_send_skb(sk,
 911                                                 alloclen + hh_len + 15,
 912                                                 (flags & MSG_DONTWAIT), &err);
 913                         } else {
 914                                 skb = NULL;
 915                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 916                                     2 * sk->sk_sndbuf)
 917                                         skb = sock_wmalloc(sk,
 918                                                            alloclen + hh_len + 15, 1,
 919                                                            sk->sk_allocation);
 920                                 if (unlikely(skb == NULL))
 921                                         err = -ENOBUFS;
 922                                 else
 923                                         /* only the initial fragment is
 924                                            time stamped */
 925                                         cork->tx_flags = 0;
 926                         }
 927                         if (skb == NULL)
 928                                 goto error;
 929
 930                         /*
 931                          *      Fill in the control structures
 932                          */
 933                         skb->ip_summed = csummode;
 934                         skb->csum = 0;
 935                         skb_reserve(skb, hh_len);
 936                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 937
 938                         /*
 939                          *      Find where to start putting bytes.
 940                          */
 941                         data = skb_put(skb, fraglen + exthdrlen);
 942                         skb_set_network_header(skb, exthdrlen);
 943                         skb->transport_header = (skb->network_header +
 944                                                  fragheaderlen);
 945                         data += fragheaderlen + exthdrlen;
 946
 947                         if (fraggap) {
 948                                 skb->csum = skb_copy_and_csum_bits(
 949                                         skb_prev, maxfraglen,
 950                                         data + transhdrlen, fraggap, 0);
 951                                 skb_prev->csum = csum_sub(skb_prev->csum,
 952                                                           skb->csum);
 953                                 data += fraggap;
 954                                 pskb_trim_unique(skb_prev, maxfraglen);
 955                         }
 956
 957                         copy = datalen - transhdrlen - fraggap;
 958                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 959                                 err = -EFAULT;
 960                                 kfree_skb(skb);
 961                                 goto error;
 962                         }
 963
 964                         offset += copy;
 965                         length -= datalen - fraggap;
 966                         transhdrlen = 0;
 967                         exthdrlen = 0;
 968                         csummode = CHECKSUM_NONE;
 969
 970                         /*
 971                          * Put the packet on the pending queue.
 972                          */
 973                         __skb_queue_tail(queue, skb);
 974                         continue;
 975                 }
 976
 977                 if (copy > length)
 978                         copy = length;
 979
 980                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
 981                         unsigned int off;
 982
 983                         off = skb->len;
 984                         if (getfrag(from, skb_put(skb, copy),
 985                                         offset, copy, off, skb) < 0) {
 986                                 __skb_trim(skb, off);
 987                                 err = -EFAULT;
 988                                 goto error;
 989                         }
 990                 } else {
 991                         int i = skb_shinfo(skb)->nr_frags;
 992
 993                         err = -ENOMEM;
 994                         if (!sk_page_frag_refill(sk, pfrag))
 995                                 goto error;
 996
 997                         if (!skb_can_coalesce(skb, i, pfrag->page,
 998                                               pfrag->offset)) {
 999                                 err = -EMSGSIZE;
1000                                 if (i == MAX_SKB_FRAGS)
1001                                         goto error;
1002
1003                                 __skb_fill_page_desc(skb, i, pfrag->page,
1004                                                      pfrag->offset, 0);
1005                                 skb_shinfo(skb)->nr_frags = ++i;
1006                                 get_page(pfrag->page);
1007                         }
1008                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1009                         if (getfrag(from,
1010                                     page_address(pfrag->page) + pfrag->offset,
1011                                     offset, copy, skb->len, skb) < 0)
1012                                 goto error_efault;
1013
1014                         pfrag->offset += copy;
1015                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1016                         skb->len += copy;
1017                         skb->data_len += copy;
1018                         skb->truesize += copy;
1019                         atomic_add(copy, &sk->sk_wmem_alloc);
1020                 }
1021                 offset += copy;
1022                 length -= copy;
1023         }
1024
1025         return 0;
1026
1027 error_efault:
1028         err = -EFAULT;
1029 error:
1030         cork->length -= length;
1031         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1032         return err;
1033 }
1034
1035 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1036                          struct ipcm_cookie *ipc, struct rtable **rtp)
1037 {
1038         struct ip_options_rcu *opt;
1039         struct rtable *rt;
1040
1041         /*
1042          * setup for corking.
1043          */
1044         opt = ipc->opt;
1045         if (opt) {
1046                 if (cork->opt == NULL) {
1047                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1048                                             sk->sk_allocation);
1049                         if (unlikely(cork->opt == NULL))
1050                                 return -ENOBUFS;
1051                 }
1052                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1053                 cork->flags |= IPCORK_OPT;
1054                 cork->addr = ipc->addr;
1055         }
1056         rt = *rtp;
1057         if (unlikely(!rt))
1058                 return -EFAULT;
1059         /*
1060          * We steal reference to this route, caller should not release it
1061          */
1062         *rtp = NULL;
1063         cork->fragsize = ip_sk_use_pmtu(sk) ?
1064                          dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1065         cork->dst = &rt->dst;
1066         cork->length = 0;
1067         cork->ttl = ipc->ttl;
1068         cork->tos = ipc->tos;
1069         cork->priority = ipc->priority;
1070         cork->tx_flags = ipc->tx_flags;
1071
1072         return 0;
1073 }
1074
1075 /*
1076  *      ip_append_data() and ip_append_page() can make one large IP datagram
1077  *      from many pieces of data. Each pieces will be holded on the socket
1078  *      until ip_push_pending_frames() is called. Each piece can be a page
1079  *      or non-page data.
1080  *
1081  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1082  *      this interface potentially.
1083  *
1084  *      LATER: length must be adjusted by pad at tail, when it is required.
1085  */
1086 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1087                    int getfrag(void *from, char *to, int offset, int len,
1088                                int odd, struct sk_buff *skb),
1089                    void *from, int length, int transhdrlen,
1090                    struct ipcm_cookie *ipc, struct rtable **rtp,
1091                    unsigned int flags)
1092 {
1093         struct inet_sock *inet = inet_sk(sk);
1094         int err;
1095
1096         if (flags&MSG_PROBE)
1097                 return 0;
1098
1099         if (skb_queue_empty(&sk->sk_write_queue)) {
1100                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1101                 if (err)
1102                         return err;
1103         } else {
1104                 transhdrlen = 0;
1105         }
1106
1107         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1108                                 sk_page_frag(sk), getfrag,
1109                                 from, length, transhdrlen, flags);
1110 }
1111
1112 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1113                        int offset, size_t size, int flags)
1114 {
1115         struct inet_sock *inet = inet_sk(sk);
1116         struct sk_buff *skb;
1117         struct rtable *rt;
1118         struct ip_options *opt = NULL;
1119         struct inet_cork *cork;
1120         int hh_len;
1121         int mtu;
1122         int len;
1123         int err;
1124         unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1125
1126         if (inet->hdrincl)
1127                 return -EPERM;
1128
1129         if (flags&MSG_PROBE)
1130                 return 0;
1131
1132         if (skb_queue_empty(&sk->sk_write_queue))
1133                 return -EINVAL;
1134
1135         cork = &inet->cork.base;
1136         rt = (struct rtable *)cork->dst;
1137         if (cork->flags & IPCORK_OPT)
1138                 opt = cork->opt;
1139
1140         if (!(rt->dst.dev->features&NETIF_F_SG))
1141                 return -EOPNOTSUPP;
1142
1143         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1144         mtu = cork->fragsize;
1145
1146         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1147         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1148         maxnonfragsize = ip_sk_local_df(sk) ? 0xFFFF : mtu;
1149
1150         if (cork->length + size > maxnonfragsize - fragheaderlen) {
1151                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1152                                mtu - (opt ? opt->optlen : 0));
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         skb->local_df = ip_sk_local_df(sk);
1310
1311         /* DF bit is set when we want to see DF on outgoing frames.
1312          * If local_df is set too, we still allow to fragment this frame
1313          * locally. */
1314         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1315             inet->pmtudisc == IP_PMTUDISC_PROBE ||
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 (cork->ttl != 0)
1324                 ttl = cork->ttl;
1325         else if (rt->rt_type == RTN_MULTICAST)
1326                 ttl = inet->mc_ttl;
1327         else
1328                 ttl = ip_select_ttl(inet, &rt->dst);
1329
1330         iph = ip_hdr(skb);
1331         iph->version = 4;
1332         iph->ihl = 5;
1333         iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1334         iph->frag_off = df;
1335         iph->ttl = ttl;
1336         iph->protocol = sk->sk_protocol;
1337         ip_copy_addrs(iph, fl4);
1338         ip_select_ident(skb, &rt->dst, sk);
1339
1340         if (opt) {
1341                 iph->ihl += opt->optlen>>2;
1342                 ip_options_build(skb, opt, cork->addr, rt, 0);
1343         }
1344
1345         skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1346         skb->mark = sk->sk_mark;
1347         /*
1348          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1349          * on dst refcount
1350          */
1351         cork->dst = NULL;
1352         skb_dst_set(skb, &rt->dst);
1353
1354         if (iph->protocol == IPPROTO_ICMP)
1355                 icmp_out_count(net, ((struct icmphdr *)
1356                         skb_transport_header(skb))->type);
1357
1358         ip_cork_release(cork);
1359 out:
1360         return skb;
1361 }
1362
1363 int ip_send_skb(struct net *net, struct sk_buff *skb)
1364 {
1365         int err;
1366
1367         err = ip_local_out(skb);
1368         if (err) {
1369                 if (err > 0)
1370                         err = net_xmit_errno(err);
1371                 if (err)
1372                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1373         }
1374
1375         return err;
1376 }
1377
1378 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1379 {
1380         struct sk_buff *skb;
1381
1382         skb = ip_finish_skb(sk, fl4);
1383         if (!skb)
1384                 return 0;
1385
1386         /* Netfilter gets whole the not fragmented skb. */
1387         return ip_send_skb(sock_net(sk), skb);
1388 }
1389
1390 /*
1391  *      Throw away all pending data on the socket.
1392  */
1393 static void __ip_flush_pending_frames(struct sock *sk,
1394                                       struct sk_buff_head *queue,
1395                                       struct inet_cork *cork)
1396 {
1397         struct sk_buff *skb;
1398
1399         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1400                 kfree_skb(skb);
1401
1402         ip_cork_release(cork);
1403 }
1404
1405 void ip_flush_pending_frames(struct sock *sk)
1406 {
1407         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1408 }
1409
1410 struct sk_buff *ip_make_skb(struct sock *sk,
1411                             struct flowi4 *fl4,
1412                             int getfrag(void *from, char *to, int offset,
1413                                         int len, int odd, struct sk_buff *skb),
1414                             void *from, int length, int transhdrlen,
1415                             struct ipcm_cookie *ipc, struct rtable **rtp,
1416                             unsigned int flags)
1417 {
1418         struct inet_cork cork;
1419         struct sk_buff_head queue;
1420         int err;
1421
1422         if (flags & MSG_PROBE)
1423                 return NULL;
1424
1425         __skb_queue_head_init(&queue);
1426
1427         cork.flags = 0;
1428         cork.addr = 0;
1429         cork.opt = NULL;
1430         err = ip_setup_cork(sk, &cork, ipc, rtp);
1431         if (err)
1432                 return ERR_PTR(err);
1433
1434         err = __ip_append_data(sk, fl4, &queue, &cork,
1435                                &current->task_frag, getfrag,
1436                                from, length, transhdrlen, flags);
1437         if (err) {
1438                 __ip_flush_pending_frames(sk, &queue, &cork);
1439                 return ERR_PTR(err);
1440         }
1441
1442         return __ip_make_skb(sk, fl4, &queue, &cork);
1443 }
1444
1445 /*
1446  *      Fetch data from kernel space and fill in checksum if needed.
1447  */
1448 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1449                               int len, int odd, struct sk_buff *skb)
1450 {
1451         __wsum csum;
1452
1453         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1454         skb->csum = csum_block_add(skb->csum, csum, odd);
1455         return 0;
1456 }
1457
1458 /*
1459  *      Generic function to send a packet as reply to another packet.
1460  *      Used to send some TCP resets/acks so far.
1461  *
1462  *      Use a fake percpu inet socket to avoid false sharing and contention.
1463  */
1464 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1465         .sk = {
1466                 .__sk_common = {
1467                         .skc_refcnt = ATOMIC_INIT(1),
1468                 },
1469                 .sk_wmem_alloc  = ATOMIC_INIT(1),
1470                 .sk_allocation  = GFP_ATOMIC,
1471                 .sk_flags       = (1UL << SOCK_USE_WRITE_QUEUE),
1472         },
1473         .pmtudisc       = IP_PMTUDISC_WANT,
1474         .uc_ttl         = -1,
1475 };
1476
1477 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1478                            __be32 saddr, const struct ip_reply_arg *arg,
1479                            unsigned int len)
1480 {
1481         struct ip_options_data replyopts;
1482         struct ipcm_cookie ipc;
1483         struct flowi4 fl4;
1484         struct rtable *rt = skb_rtable(skb);
1485         struct sk_buff *nskb;
1486         struct sock *sk;
1487         struct inet_sock *inet;
1488
1489         if (ip_options_echo(&replyopts.opt.opt, skb))
1490                 return;
1491
1492         ipc.addr = daddr;
1493         ipc.opt = NULL;
1494         ipc.tx_flags = 0;
1495         ipc.ttl = 0;
1496         ipc.tos = -1;
1497
1498         if (replyopts.opt.opt.optlen) {
1499                 ipc.opt = &replyopts.opt;
1500
1501                 if (replyopts.opt.opt.srr)
1502                         daddr = replyopts.opt.opt.faddr;
1503         }
1504
1505         flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1506                            RT_TOS(arg->tos),
1507                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1508                            ip_reply_arg_flowi_flags(arg),
1509                            daddr, saddr,
1510                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1511         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1512         rt = ip_route_output_key(net, &fl4);
1513         if (IS_ERR(rt))
1514                 return;
1515
1516         inet = &get_cpu_var(unicast_sock);
1517
1518         inet->tos = arg->tos;
1519         sk = &inet->sk;
1520         sk->sk_priority = skb->priority;
1521         sk->sk_protocol = ip_hdr(skb)->protocol;
1522         sk->sk_bound_dev_if = arg->bound_dev_if;
1523         sock_net_set(sk, net);
1524         __skb_queue_head_init(&sk->sk_write_queue);
1525         sk->sk_sndbuf = sysctl_wmem_default;
1526         ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1527                        &ipc, &rt, MSG_DONTWAIT);
1528         nskb = skb_peek(&sk->sk_write_queue);
1529         if (nskb) {
1530                 if (arg->csumoffset >= 0)
1531                         *((__sum16 *)skb_transport_header(nskb) +
1532                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1533                                                                 arg->csum));
1534                 nskb->ip_summed = CHECKSUM_NONE;
1535                 skb_orphan(nskb);
1536                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1537                 ip_push_pending_frames(sk, &fl4);
1538         }
1539
1540         put_cpu_var(unicast_sock);
1541
1542         ip_rt_put(rt);
1543 }
1544
1545 void __init ip_init(void)
1546 {
1547         ip_rt_init();
1548         inet_initpeers();
1549
1550 #if defined(CONFIG_IP_MULTICAST)
1551         igmp_mc_init();
1552 #endif
1553 }