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 int ip_queue_xmit(struct sk_buff *skb, struct flowi *fl)
 319 {
 320         struct sock *sk = skb->sk;
 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         skb->priority = sk->sk_priority;
 393         skb->mark = sk->sk_mark;
 394
 395         res = ip_local_out(skb);
 396         rcu_read_unlock();
 397         return res;
 398
 399 no_route:
 400         rcu_read_unlock();
 401         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 402         kfree_skb(skb);
 403         return -EHOSTUNREACH;
 404 }
 405 EXPORT_SYMBOL(ip_queue_xmit);
 406
 407
 408 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 409 {
 410         to->pkt_type = from->pkt_type;
 411         to->priority = from->priority;
 412         to->protocol = from->protocol;
 413         skb_dst_drop(to);
 414         skb_dst_copy(to, from);
 415         to->dev = from->dev;
 416         to->mark = from->mark;
 417
 418         /* Copy the flags to each fragment. */
 419         IPCB(to)->flags = IPCB(from)->flags;
 420
 421 #ifdef CONFIG_NET_SCHED
 422         to->tc_index = from->tc_index;
 423 #endif
 424         nf_copy(to, from);
 425 #if IS_ENABLED(CONFIG_NETFILTER_XT_TARGET_TRACE)
 426         to->nf_trace = from->nf_trace;
 427 #endif
 428 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 429         to->ipvs_property = from->ipvs_property;
 430 #endif
 431         skb_copy_secmark(to, from);
 432 }
 433
 434 /*
 435  *      This IP datagram is too large to be sent in one piece.  Break it up into
 436  *      smaller pieces (each of size equal to IP header plus
 437  *      a block of the data of the original IP data part) that will yet fit in a
 438  *      single device frame, and queue such a frame for sending.
 439  */
 440
 441 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 442 {
 443         struct iphdr *iph;
 444         int ptr;
 445         struct net_device *dev;
 446         struct sk_buff *skb2;
 447         unsigned int mtu, hlen, left, len, ll_rs;
 448         int offset;
 449         __be16 not_last_frag;
 450         struct rtable *rt = skb_rtable(skb);
 451         int err = 0;
 452         bool forwarding = IPCB(skb)->flags & IPSKB_FORWARDED;
 453
 454         dev = rt->dst.dev;
 455
 456         /*
 457          *      Point into the IP datagram header.
 458          */
 459
 460         iph = ip_hdr(skb);
 461
 462         mtu = ip_dst_mtu_maybe_forward(&rt->dst, forwarding);
 463         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->local_df) ||
 464                      (IPCB(skb)->frag_max_size &&
 465                       IPCB(skb)->frag_max_size > mtu))) {
 466                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 467                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 468                           htonl(mtu));
 469                 kfree_skb(skb);
 470                 return -EMSGSIZE;
 471         }
 472
 473         /*
 474          *      Setup starting values.
 475          */
 476
 477         hlen = iph->ihl * 4;
 478         mtu = mtu - hlen;       /* Size of data space */
 479 #ifdef CONFIG_BRIDGE_NETFILTER
 480         if (skb->nf_bridge)
 481                 mtu -= nf_bridge_mtu_reduction(skb);
 482 #endif
 483         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 484
 485         /* When frag_list is given, use it. First, check its validity:
 486          * some transformers could create wrong frag_list or break existing
 487          * one, it is not prohibited. In this case fall back to copying.
 488          *
 489          * LATER: this step can be merged to real generation of fragments,
 490          * we can switch to copy when see the first bad fragment.
 491          */
 492         if (skb_has_frag_list(skb)) {
 493                 struct sk_buff *frag, *frag2;
 494                 int first_len = skb_pagelen(skb);
 495
 496                 if (first_len - hlen > mtu ||
 497                     ((first_len - hlen) & 7) ||
 498                     ip_is_fragment(iph) ||
 499                     skb_cloned(skb))
 500                         goto slow_path;
 501
 502                 skb_walk_frags(skb, frag) {
 503                         /* Correct geometry. */
 504                         if (frag->len > mtu ||
 505                             ((frag->len & 7) && frag->next) ||
 506                             skb_headroom(frag) < hlen)
 507                                 goto slow_path_clean;
 508
 509                         /* Partially cloned skb? */
 510                         if (skb_shared(frag))
 511                                 goto slow_path_clean;
 512
 513                         BUG_ON(frag->sk);
 514                         if (skb->sk) {
 515                                 frag->sk = skb->sk;
 516                                 frag->destructor = sock_wfree;
 517                         }
 518                         skb->truesize -= frag->truesize;
 519                 }
 520
 521                 /* Everything is OK. Generate! */
 522
 523                 err = 0;
 524                 offset = 0;
 525                 frag = skb_shinfo(skb)->frag_list;
 526                 skb_frag_list_init(skb);
 527                 skb->data_len = first_len - skb_headlen(skb);
 528                 skb->len = first_len;
 529                 iph->tot_len = htons(first_len);
 530                 iph->frag_off = htons(IP_MF);
 531                 ip_send_check(iph);
 532
 533                 for (;;) {
 534                         /* Prepare header of the next frame,
 535                          * before previous one went down. */
 536                         if (frag) {
 537                                 frag->ip_summed = CHECKSUM_NONE;
 538                                 skb_reset_transport_header(frag);
 539                                 __skb_push(frag, hlen);
 540                                 skb_reset_network_header(frag);
 541                                 memcpy(skb_network_header(frag), iph, hlen);
 542                                 iph = ip_hdr(frag);
 543                                 iph->tot_len = htons(frag->len);
 544                                 ip_copy_metadata(frag, skb);
 545                                 if (offset == 0)
 546                                         ip_options_fragment(frag);
 547                                 offset += skb->len - hlen;
 548                                 iph->frag_off = htons(offset>>3);
 549                                 if (frag->next != NULL)
 550                                         iph->frag_off |= htons(IP_MF);
 551                                 /* Ready, complete checksum */
 552                                 ip_send_check(iph);
 553                         }
 554
 555                         err = output(skb);
 556
 557                         if (!err)
 558                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 559                         if (err || !frag)
 560                                 break;
 561
 562                         skb = frag;
 563                         frag = skb->next;
 564                         skb->next = NULL;
 565                 }
 566
 567                 if (err == 0) {
 568                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 569                         return 0;
 570                 }
 571
 572                 while (frag) {
 573                         skb = frag->next;
 574                         kfree_skb(frag);
 575                         frag = skb;
 576                 }
 577                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 578                 return err;
 579
 580 slow_path_clean:
 581                 skb_walk_frags(skb, frag2) {
 582                         if (frag2 == frag)
 583                                 break;
 584                         frag2->sk = NULL;
 585                         frag2->destructor = NULL;
 586                         skb->truesize += frag2->truesize;
 587                 }
 588         }
 589
 590 slow_path:
 591         /* for offloaded checksums cleanup checksum before fragmentation */
 592         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 593                 goto fail;
 594         iph = ip_hdr(skb);
 595
 596         left = skb->len - hlen;         /* Space per frame */
 597         ptr = hlen;             /* Where to start from */
 598
 599         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 600          * we need to make room for the encapsulating header
 601          */
 602         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 603
 604         /*
 605          *      Fragment the datagram.
 606          */
 607
 608         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 609         not_last_frag = iph->frag_off & htons(IP_MF);
 610
 611         /*
 612          *      Keep copying data until we run out.
 613          */
 614
 615         while (left > 0) {
 616                 len = left;
 617                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 618                 if (len > mtu)
 619                         len = mtu;
 620                 /* IF: we are not sending up to and including the packet end
 621                    then align the next start on an eight byte boundary */
 622                 if (len < left) {
 623                         len &= ~7;
 624                 }
 625                 /*
 626                  *      Allocate buffer.
 627                  */
 628
 629                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 630                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 631                         err = -ENOMEM;
 632                         goto fail;
 633                 }
 634
 635                 /*
 636                  *      Set up data on packet
 637                  */
 638
 639                 ip_copy_metadata(skb2, skb);
 640                 skb_reserve(skb2, ll_rs);
 641                 skb_put(skb2, len + hlen);
 642                 skb_reset_network_header(skb2);
 643                 skb2->transport_header = skb2->network_header + hlen;
 644
 645                 /*
 646                  *      Charge the memory for the fragment to any owner
 647                  *      it might possess
 648                  */
 649
 650                 if (skb->sk)
 651                         skb_set_owner_w(skb2, skb->sk);
 652
 653                 /*
 654                  *      Copy the packet header into the new buffer.
 655                  */
 656
 657                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 658
 659                 /*
 660                  *      Copy a block of the IP datagram.
 661                  */
 662                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 663                         BUG();
 664                 left -= len;
 665
 666                 /*
 667                  *      Fill in the new header fields.
 668                  */
 669                 iph = ip_hdr(skb2);
 670                 iph->frag_off = htons((offset >> 3));
 671
 672                 /* ANK: dirty, but effective trick. Upgrade options only if
 673                  * the segment to be fragmented was THE FIRST (otherwise,
 674                  * options are already fixed) and make it ONCE
 675                  * on the initial skb, so that all the following fragments
 676                  * will inherit fixed options.
 677                  */
 678                 if (offset == 0)
 679                         ip_options_fragment(skb);
 680
 681                 /*
 682                  *      Added AC : If we are fragmenting a fragment that's not the
 683                  *                 last fragment then keep MF on each bit
 684                  */
 685                 if (left > 0 || not_last_frag)
 686                         iph->frag_off |= htons(IP_MF);
 687                 ptr += len;
 688                 offset += len;
 689
 690                 /*
 691                  *      Put this fragment into the sending queue.
 692                  */
 693                 iph->tot_len = htons(len + hlen);
 694
 695                 ip_send_check(iph);
 696
 697                 err = output(skb2);
 698                 if (err)
 699                         goto fail;
 700
 701                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 702         }
 703         consume_skb(skb);
 704         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 705         return err;
 706
 707 fail:
 708         kfree_skb(skb);
 709         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 710         return err;
 711 }
 712 EXPORT_SYMBOL(ip_fragment);
 713
 714 int
 715 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 716 {
 717         struct iovec *iov = from;
 718
 719         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 720                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 721                         return -EFAULT;
 722         } else {
 723                 __wsum csum = 0;
 724                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 725                         return -EFAULT;
 726                 skb->csum = csum_block_add(skb->csum, csum, odd);
 727         }
 728         return 0;
 729 }
 730 EXPORT_SYMBOL(ip_generic_getfrag);
 731
 732 static inline __wsum
 733 csum_page(struct page *page, int offset, int copy)
 734 {
 735         char *kaddr;
 736         __wsum csum;
 737         kaddr = kmap(page);
 738         csum = csum_partial(kaddr + offset, copy, 0);
 739         kunmap(page);
 740         return csum;
 741 }
 742
 743 static inline int ip_ufo_append_data(struct sock *sk,
 744                         struct sk_buff_head *queue,
 745                         int getfrag(void *from, char *to, int offset, int len,
 746                                int odd, struct sk_buff *skb),
 747                         void *from, int length, int hh_len, int fragheaderlen,
 748                         int transhdrlen, int maxfraglen, unsigned int flags)
 749 {
 750         struct sk_buff *skb;
 751         int err;
 752
 753         /* There is support for UDP fragmentation offload by network
 754          * device, so create one single skb packet containing complete
 755          * udp datagram
 756          */
 757         if ((skb = skb_peek_tail(queue)) == NULL) {
 758                 skb = sock_alloc_send_skb(sk,
 759                         hh_len + fragheaderlen + transhdrlen + 20,
 760                         (flags & MSG_DONTWAIT), &err);
 761
 762                 if (skb == NULL)
 763                         return err;
 764
 765                 /* reserve space for Hardware header */
 766                 skb_reserve(skb, hh_len);
 767
 768                 /* create space for UDP/IP header */
 769                 skb_put(skb, fragheaderlen + transhdrlen);
 770
 771                 /* initialize network header pointer */
 772                 skb_reset_network_header(skb);
 773
 774                 /* initialize protocol header pointer */
 775                 skb->transport_header = skb->network_header + fragheaderlen;
 776
 777                 skb->csum = 0;
 778
 779
 780                 __skb_queue_tail(queue, skb);
 781         } else if (skb_is_gso(skb)) {
 782                 goto append;
 783         }
 784
 785         skb->ip_summed = CHECKSUM_PARTIAL;
 786         /* specify the length of each IP datagram fragment */
 787         skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 788         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 789
 790 append:
 791         return skb_append_datato_frags(sk, skb, getfrag, from,
 792                                        (length - transhdrlen));
 793 }
 794
 795 static int __ip_append_data(struct sock *sk,
 796                             struct flowi4 *fl4,
 797                             struct sk_buff_head *queue,
 798                             struct inet_cork *cork,
 799                             struct page_frag *pfrag,
 800                             int getfrag(void *from, char *to, int offset,
 801                                         int len, int odd, struct sk_buff *skb),
 802                             void *from, int length, int transhdrlen,
 803                             unsigned int flags)
 804 {
 805         struct inet_sock *inet = inet_sk(sk);
 806         struct sk_buff *skb;
 807
 808         struct ip_options *opt = cork->opt;
 809         int hh_len;
 810         int exthdrlen;
 811         int mtu;
 812         int copy;
 813         int err;
 814         int offset = 0;
 815         unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
 816         int csummode = CHECKSUM_NONE;
 817         struct rtable *rt = (struct rtable *)cork->dst;
 818
 819         skb = skb_peek_tail(queue);
 820
 821         exthdrlen = !skb ? rt->dst.header_len : 0;
 822         mtu = cork->fragsize;
 823
 824         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 825
 826         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 827         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 828         maxnonfragsize = (inet->pmtudisc >= IP_PMTUDISC_DO) ?
 829                          mtu : 0xFFFF;
 830
 831         if (cork->length + length > maxnonfragsize - fragheaderlen) {
 832                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 833                                mtu - (opt ? opt->optlen : 0));
 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 ip_options_rcu *opt;
1043         struct rtable *rt;
1044
1045         /*
1046          * setup for corking.
1047          */
1048         opt = ipc->opt;
1049         if (opt) {
1050                 if (cork->opt == NULL) {
1051                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1052                                             sk->sk_allocation);
1053                         if (unlikely(cork->opt == NULL))
1054                                 return -ENOBUFS;
1055                 }
1056                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1057                 cork->flags |= IPCORK_OPT;
1058                 cork->addr = ipc->addr;
1059         }
1060         rt = *rtp;
1061         if (unlikely(!rt))
1062                 return -EFAULT;
1063         /*
1064          * We steal reference to this route, caller should not release it
1065          */
1066         *rtp = NULL;
1067         cork->fragsize = ip_sk_use_pmtu(sk) ?
1068                          dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1069         cork->dst = &rt->dst;
1070         cork->length = 0;
1071         cork->ttl = ipc->ttl;
1072         cork->tos = ipc->tos;
1073         cork->priority = ipc->priority;
1074         cork->tx_flags = ipc->tx_flags;
1075
1076         return 0;
1077 }
1078
1079 /*
1080  *      ip_append_data() and ip_append_page() can make one large IP datagram
1081  *      from many pieces of data. Each pieces will be holded on the socket
1082  *      until ip_push_pending_frames() is called. Each piece can be a page
1083  *      or non-page data.
1084  *
1085  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1086  *      this interface potentially.
1087  *
1088  *      LATER: length must be adjusted by pad at tail, when it is required.
1089  */
1090 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1091                    int getfrag(void *from, char *to, int offset, int len,
1092                                int odd, struct sk_buff *skb),
1093                    void *from, int length, int transhdrlen,
1094                    struct ipcm_cookie *ipc, struct rtable **rtp,
1095                    unsigned int flags)
1096 {
1097         struct inet_sock *inet = inet_sk(sk);
1098         int err;
1099
1100         if (flags&MSG_PROBE)
1101                 return 0;
1102
1103         if (skb_queue_empty(&sk->sk_write_queue)) {
1104                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1105                 if (err)
1106                         return err;
1107         } else {
1108                 transhdrlen = 0;
1109         }
1110
1111         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1112                                 sk_page_frag(sk), getfrag,
1113                                 from, length, transhdrlen, flags);
1114 }
1115
1116 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1117                        int offset, size_t size, int flags)
1118 {
1119         struct inet_sock *inet = inet_sk(sk);
1120         struct sk_buff *skb;
1121         struct rtable *rt;
1122         struct ip_options *opt = NULL;
1123         struct inet_cork *cork;
1124         int hh_len;
1125         int mtu;
1126         int len;
1127         int err;
1128         unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1129
1130         if (inet->hdrincl)
1131                 return -EPERM;
1132
1133         if (flags&MSG_PROBE)
1134                 return 0;
1135
1136         if (skb_queue_empty(&sk->sk_write_queue))
1137                 return -EINVAL;
1138
1139         cork = &inet->cork.base;
1140         rt = (struct rtable *)cork->dst;
1141         if (cork->flags & IPCORK_OPT)
1142                 opt = cork->opt;
1143
1144         if (!(rt->dst.dev->features&NETIF_F_SG))
1145                 return -EOPNOTSUPP;
1146
1147         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1148         mtu = cork->fragsize;
1149
1150         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1151         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1152         maxnonfragsize = (inet->pmtudisc >= IP_PMTUDISC_DO) ?
1153                          mtu : 0xFFFF;
1154
1155         if (cork->length + size > maxnonfragsize - fragheaderlen) {
1156                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1157                                mtu - (opt ? opt->optlen : 0));
1158                 return -EMSGSIZE;
1159         }
1160
1161         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1162                 return -EINVAL;
1163
1164         cork->length += size;
1165         if ((size + skb->len > mtu) &&
1166             (sk->sk_protocol == IPPROTO_UDP) &&
1167             (rt->dst.dev->features & NETIF_F_UFO)) {
1168                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1169                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1170         }
1171
1172
1173         while (size > 0) {
1174                 int i;
1175
1176                 if (skb_is_gso(skb))
1177                         len = size;
1178                 else {
1179
1180                         /* Check if the remaining data fits into current packet. */
1181                         len = mtu - skb->len;
1182                         if (len < size)
1183                                 len = maxfraglen - skb->len;
1184                 }
1185                 if (len <= 0) {
1186                         struct sk_buff *skb_prev;
1187                         int alloclen;
1188
1189                         skb_prev = skb;
1190                         fraggap = skb_prev->len - maxfraglen;
1191
1192                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1193                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1194                         if (unlikely(!skb)) {
1195                                 err = -ENOBUFS;
1196                                 goto error;
1197                         }
1198
1199                         /*
1200                          *      Fill in the control structures
1201                          */
1202                         skb->ip_summed = CHECKSUM_NONE;
1203                         skb->csum = 0;
1204                         skb_reserve(skb, hh_len);
1205
1206                         /*
1207                          *      Find where to start putting bytes.
1208                          */
1209                         skb_put(skb, fragheaderlen + fraggap);
1210                         skb_reset_network_header(skb);
1211                         skb->transport_header = (skb->network_header +
1212                                                  fragheaderlen);
1213                         if (fraggap) {
1214                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1215                                                                    maxfraglen,
1216                                                     skb_transport_header(skb),
1217                                                                    fraggap, 0);
1218                                 skb_prev->csum = csum_sub(skb_prev->csum,
1219                                                           skb->csum);
1220                                 pskb_trim_unique(skb_prev, maxfraglen);
1221                         }
1222
1223                         /*
1224                          * Put the packet on the pending queue.
1225                          */
1226                         __skb_queue_tail(&sk->sk_write_queue, skb);
1227                         continue;
1228                 }
1229
1230                 i = skb_shinfo(skb)->nr_frags;
1231                 if (len > size)
1232                         len = size;
1233                 if (skb_can_coalesce(skb, i, page, offset)) {
1234                         skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1235                 } else if (i < MAX_SKB_FRAGS) {
1236                         get_page(page);
1237                         skb_fill_page_desc(skb, i, page, offset, len);
1238                 } else {
1239                         err = -EMSGSIZE;
1240                         goto error;
1241                 }
1242
1243                 if (skb->ip_summed == CHECKSUM_NONE) {
1244                         __wsum csum;
1245                         csum = csum_page(page, offset, len);
1246                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1247                 }
1248
1249                 skb->len += len;
1250                 skb->data_len += len;
1251                 skb->truesize += len;
1252                 atomic_add(len, &sk->sk_wmem_alloc);
1253                 offset += len;
1254                 size -= len;
1255         }
1256         return 0;
1257
1258 error:
1259         cork->length -= size;
1260         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1261         return err;
1262 }
1263
1264 static void ip_cork_release(struct inet_cork *cork)
1265 {
1266         cork->flags &= ~IPCORK_OPT;
1267         kfree(cork->opt);
1268         cork->opt = NULL;
1269         dst_release(cork->dst);
1270         cork->dst = NULL;
1271 }
1272
1273 /*
1274  *      Combined all pending IP fragments on the socket as one IP datagram
1275  *      and push them out.
1276  */
1277 struct sk_buff *__ip_make_skb(struct sock *sk,
1278                               struct flowi4 *fl4,
1279                               struct sk_buff_head *queue,
1280                               struct inet_cork *cork)
1281 {
1282         struct sk_buff *skb, *tmp_skb;
1283         struct sk_buff **tail_skb;
1284         struct inet_sock *inet = inet_sk(sk);
1285         struct net *net = sock_net(sk);
1286         struct ip_options *opt = NULL;
1287         struct rtable *rt = (struct rtable *)cork->dst;
1288         struct iphdr *iph;
1289         __be16 df = 0;
1290         __u8 ttl;
1291
1292         if ((skb = __skb_dequeue(queue)) == NULL)
1293                 goto out;
1294         tail_skb = &(skb_shinfo(skb)->frag_list);
1295
1296         /* move skb->data to ip header from ext header */
1297         if (skb->data < skb_network_header(skb))
1298                 __skb_pull(skb, skb_network_offset(skb));
1299         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1300                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1301                 *tail_skb = tmp_skb;
1302                 tail_skb = &(tmp_skb->next);
1303                 skb->len += tmp_skb->len;
1304                 skb->data_len += tmp_skb->len;
1305                 skb->truesize += tmp_skb->truesize;
1306                 tmp_skb->destructor = NULL;
1307                 tmp_skb->sk = NULL;
1308         }
1309
1310         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1311          * to fragment the frame generated here. No matter, what transforms
1312          * how transforms change size of the packet, it will come out.
1313          */
1314         if (inet->pmtudisc < IP_PMTUDISC_DO)
1315                 skb->local_df = 1;
1316
1317         /* DF bit is set when we want to see DF on outgoing frames.
1318          * If local_df is set too, we still allow to fragment this frame
1319          * locally. */
1320         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1321             inet->pmtudisc == IP_PMTUDISC_PROBE ||
1322             (skb->len <= dst_mtu(&rt->dst) &&
1323              ip_dont_fragment(sk, &rt->dst)))
1324                 df = htons(IP_DF);
1325
1326         if (cork->flags & IPCORK_OPT)
1327                 opt = cork->opt;
1328
1329         if (cork->ttl != 0)
1330                 ttl = cork->ttl;
1331         else if (rt->rt_type == RTN_MULTICAST)
1332                 ttl = inet->mc_ttl;
1333         else
1334                 ttl = ip_select_ttl(inet, &rt->dst);
1335
1336         iph = ip_hdr(skb);
1337         iph->version = 4;
1338         iph->ihl = 5;
1339         iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1340         iph->frag_off = df;
1341         iph->ttl = ttl;
1342         iph->protocol = sk->sk_protocol;
1343         ip_copy_addrs(iph, fl4);
1344         ip_select_ident(skb, &rt->dst, sk);
1345
1346         if (opt) {
1347                 iph->ihl += opt->optlen>>2;
1348                 ip_options_build(skb, opt, cork->addr, rt, 0);
1349         }
1350
1351         skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1352         skb->mark = sk->sk_mark;
1353         /*
1354          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1355          * on dst refcount
1356          */
1357         cork->dst = NULL;
1358         skb_dst_set(skb, &rt->dst);
1359
1360         if (iph->protocol == IPPROTO_ICMP)
1361                 icmp_out_count(net, ((struct icmphdr *)
1362                         skb_transport_header(skb))->type);
1363
1364         ip_cork_release(cork);
1365 out:
1366         return skb;
1367 }
1368
1369 int ip_send_skb(struct net *net, struct sk_buff *skb)
1370 {
1371         int err;
1372
1373         err = ip_local_out(skb);
1374         if (err) {
1375                 if (err > 0)
1376                         err = net_xmit_errno(err);
1377                 if (err)
1378                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1379         }
1380
1381         return err;
1382 }
1383
1384 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1385 {
1386         struct sk_buff *skb;
1387
1388         skb = ip_finish_skb(sk, fl4);
1389         if (!skb)
1390                 return 0;
1391
1392         /* Netfilter gets whole the not fragmented skb. */
1393         return ip_send_skb(sock_net(sk), skb);
1394 }
1395
1396 /*
1397  *      Throw away all pending data on the socket.
1398  */
1399 static void __ip_flush_pending_frames(struct sock *sk,
1400                                       struct sk_buff_head *queue,
1401                                       struct inet_cork *cork)
1402 {
1403         struct sk_buff *skb;
1404
1405         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1406                 kfree_skb(skb);
1407
1408         ip_cork_release(cork);
1409 }
1410
1411 void ip_flush_pending_frames(struct sock *sk)
1412 {
1413         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1414 }
1415
1416 struct sk_buff *ip_make_skb(struct sock *sk,
1417                             struct flowi4 *fl4,
1418                             int getfrag(void *from, char *to, int offset,
1419                                         int len, int odd, struct sk_buff *skb),
1420                             void *from, int length, int transhdrlen,
1421                             struct ipcm_cookie *ipc, struct rtable **rtp,
1422                             unsigned int flags)
1423 {
1424         struct inet_cork cork;
1425         struct sk_buff_head queue;
1426         int err;
1427
1428         if (flags & MSG_PROBE)
1429                 return NULL;
1430
1431         __skb_queue_head_init(&queue);
1432
1433         cork.flags = 0;
1434         cork.addr = 0;
1435         cork.opt = NULL;
1436         err = ip_setup_cork(sk, &cork, ipc, rtp);
1437         if (err)
1438                 return ERR_PTR(err);
1439
1440         err = __ip_append_data(sk, fl4, &queue, &cork,
1441                                &current->task_frag, getfrag,
1442                                from, length, transhdrlen, flags);
1443         if (err) {
1444                 __ip_flush_pending_frames(sk, &queue, &cork);
1445                 return ERR_PTR(err);
1446         }
1447
1448         return __ip_make_skb(sk, fl4, &queue, &cork);
1449 }
1450
1451 /*
1452  *      Fetch data from kernel space and fill in checksum if needed.
1453  */
1454 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1455                               int len, int odd, struct sk_buff *skb)
1456 {
1457         __wsum csum;
1458
1459         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1460         skb->csum = csum_block_add(skb->csum, csum, odd);
1461         return 0;
1462 }
1463
1464 /*
1465  *      Generic function to send a packet as reply to another packet.
1466  *      Used to send some TCP resets/acks so far.
1467  *
1468  *      Use a fake percpu inet socket to avoid false sharing and contention.
1469  */
1470 static DEFINE_PER_CPU(struct inet_sock, unicast_sock) = {
1471         .sk = {
1472                 .__sk_common = {
1473                         .skc_refcnt = ATOMIC_INIT(1),
1474                 },
1475                 .sk_wmem_alloc  = ATOMIC_INIT(1),
1476                 .sk_allocation  = GFP_ATOMIC,
1477                 .sk_flags       = (1UL << SOCK_USE_WRITE_QUEUE),
1478         },
1479         .pmtudisc       = IP_PMTUDISC_WANT,
1480         .uc_ttl         = -1,
1481 };
1482
1483 void ip_send_unicast_reply(struct net *net, struct sk_buff *skb, __be32 daddr,
1484                            __be32 saddr, const struct ip_reply_arg *arg,
1485                            unsigned int len)
1486 {
1487         struct ip_options_data replyopts;
1488         struct ipcm_cookie ipc;
1489         struct flowi4 fl4;
1490         struct rtable *rt = skb_rtable(skb);
1491         struct sk_buff *nskb;
1492         struct sock *sk;
1493         struct inet_sock *inet;
1494
1495         if (ip_options_echo(&replyopts.opt.opt, skb))
1496                 return;
1497
1498         ipc.addr = daddr;
1499         ipc.opt = NULL;
1500         ipc.tx_flags = 0;
1501         ipc.ttl = 0;
1502         ipc.tos = -1;
1503
1504         if (replyopts.opt.opt.optlen) {
1505                 ipc.opt = &replyopts.opt;
1506
1507                 if (replyopts.opt.opt.srr)
1508                         daddr = replyopts.opt.opt.faddr;
1509         }
1510
1511         flowi4_init_output(&fl4, arg->bound_dev_if, 0,
1512                            RT_TOS(arg->tos),
1513                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1514                            ip_reply_arg_flowi_flags(arg),
1515                            daddr, saddr,
1516                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1517         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1518         rt = ip_route_output_key(net, &fl4);
1519         if (IS_ERR(rt))
1520                 return;
1521
1522         inet = &get_cpu_var(unicast_sock);
1523
1524         inet->tos = arg->tos;
1525         sk = &inet->sk;
1526         sk->sk_priority = skb->priority;
1527         sk->sk_protocol = ip_hdr(skb)->protocol;
1528         sk->sk_bound_dev_if = arg->bound_dev_if;
1529         sock_net_set(sk, net);
1530         __skb_queue_head_init(&sk->sk_write_queue);
1531         sk->sk_sndbuf = sysctl_wmem_default;
1532         ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1533                        &ipc, &rt, MSG_DONTWAIT);
1534         nskb = skb_peek(&sk->sk_write_queue);
1535         if (nskb) {
1536                 if (arg->csumoffset >= 0)
1537                         *((__sum16 *)skb_transport_header(nskb) +
1538                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1539                                                                 arg->csum));
1540                 nskb->ip_summed = CHECKSUM_NONE;
1541                 skb_orphan(nskb);
1542                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1543                 ip_push_pending_frames(sk, &fl4);
1544         }
1545
1546         put_cpu_var(unicast_sock);
1547
1548         ip_rt_put(rt);
1549 }
1550
1551 void __init ip_init(void)
1552 {
1553         ip_rt_init();
1554         inet_initpeers();
1555
1556 #if defined(CONFIG_IP_MULTICAST)
1557         igmp_mc_init();
1558 #endif
1559 }