net/ipv4/ip_output.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              The Internet Protocol (IP) output module.
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Donald Becker, <becker@super.org>
  11  *              Alan Cox, <Alan.Cox@linux.org>
  12  *              Richard Underwood
  13  *              Stefan Becker, <stefanb@yello.ping.de>
  14  *              Jorge Cwik, <jorge@laser.satlink.net>
  15  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  16  *              Hirokazu Takahashi, <taka@valinux.co.jp>
  17  *
  18  *      See ip_input.c for original log
  19  *
  20  *      Fixes:
  21  *              Alan Cox        :       Missing nonblock feature in ip_build_xmit.
  22  *              Mike Kilburn    :       htons() missing in ip_build_xmit.
  23  *              Bradford Johnson:       Fix faulty handling of some frames when
  24  *                                      no route is found.
  25  *              Alexander Demenshin:    Missing sk/skb free in ip_queue_xmit
  26  *                                      (in case if packet not accepted by
  27  *                                      output firewall rules)
  28  *              Mike McLagan    :       Routing by source
  29  *              Alexey Kuznetsov:       use new route cache
  30  *              Andi Kleen:             Fix broken PMTU recovery and remove
  31  *                                      some redundant tests.
  32  *      Vitaly E. Lavrov        :       Transparent proxy revived after year coma.
  33  *              Andi Kleen      :       Replace ip_reply with ip_send_reply.
  34  *              Andi Kleen      :       Split fast and slow ip_build_xmit path
  35  *                                      for decreased register pressure on x86
  36  *                                      and more readibility.
  37  *              Marc Boucher    :       When call_out_firewall returns FW_QUEUE,
  38  *                                      silently drop skb instead of failing with -EPERM.
  39  *              Detlev Wengorz  :       Copy protocol for fragments.
  40  *              Hirokazu Takahashi:     HW checksumming for outgoing UDP
  41  *                                      datagrams.
  42  *              Hirokazu Takahashi:     sendfile() on UDP works now.
  43  */
  44
  45 #include <asm/uaccess.h>
  46 #include <asm/system.h>
  47 #include <linux/module.h>
  48 #include <linux/types.h>
  49 #include <linux/kernel.h>
  50 #include <linux/mm.h>
  51 #include <linux/string.h>
  52 #include <linux/errno.h>
  53 #include <linux/highmem.h>
  54 #include <linux/slab.h>
  55
  56 #include <linux/socket.h>
  57 #include <linux/sockios.h>
  58 #include <linux/in.h>
  59 #include <linux/inet.h>
  60 #include <linux/netdevice.h>
  61 #include <linux/etherdevice.h>
  62 #include <linux/proc_fs.h>
  63 #include <linux/stat.h>
  64 #include <linux/init.h>
  65
  66 #include <net/snmp.h>
  67 #include <net/ip.h>
  68 #include <net/protocol.h>
  69 #include <net/route.h>
  70 #include <net/xfrm.h>
  71 #include <linux/skbuff.h>
  72 #include <net/sock.h>
  73 #include <net/arp.h>
  74 #include <net/icmp.h>
  75 #include <net/checksum.h>
  76 #include <net/inetpeer.h>
  77 #include <linux/igmp.h>
  78 #include <linux/netfilter_ipv4.h>
  79 #include <linux/netfilter_bridge.h>
  80 #include <linux/mroute.h>
  81 #include <linux/netlink.h>
  82 #include <linux/tcp.h>
  83
  84 int sysctl_ip_default_ttl __read_mostly = IPDEFTTL;
  85
  86 /* Generate a checksum for an outgoing IP datagram. */
  87 __inline__ void ip_send_check(struct iphdr *iph)
  88 {
  89         iph->check = 0;
  90         iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl);
  91 }
  92
  93 int __ip_local_out(struct sk_buff *skb)
  94 {
  95         struct iphdr *iph = ip_hdr(skb);
  96
  97         iph->tot_len = htons(skb->len);
  98         ip_send_check(iph);
  99         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 100                        skb_dst(skb)->dev, dst_output);
 101 }
 102
 103 int ip_local_out(struct sk_buff *skb)
 104 {
 105         int err;
 106
 107         err = __ip_local_out(skb);
 108         if (likely(err == 1))
 109                 err = dst_output(skb);
 110
 111         return err;
 112 }
 113 EXPORT_SYMBOL_GPL(ip_local_out);
 114
 115 /* dev_loopback_xmit for use with netfilter. */
 116 static int ip_dev_loopback_xmit(struct sk_buff *newskb)
 117 {
 118         skb_reset_mac_header(newskb);
 119         __skb_pull(newskb, skb_network_offset(newskb));
 120         newskb->pkt_type = PACKET_LOOPBACK;
 121         newskb->ip_summed = CHECKSUM_UNNECESSARY;
 122         WARN_ON(!skb_dst(newskb));
 123         netif_rx_ni(newskb);
 124         return 0;
 125 }
 126
 127 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 128 {
 129         int ttl = inet->uc_ttl;
 130
 131         if (ttl < 0)
 132                 ttl = dst_metric(dst, RTAX_HOPLIMIT);
 133         return ttl;
 134 }
 135
 136 /*
 137  *              Add an ip header to a skbuff and send it out.
 138  *
 139  */
 140 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 141                           __be32 saddr, __be32 daddr, struct ip_options *opt)
 142 {
 143         struct inet_sock *inet = inet_sk(sk);
 144         struct rtable *rt = skb_rtable(skb);
 145         struct iphdr *iph;
 146
 147         /* Build the IP header. */
 148         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 149         skb_reset_network_header(skb);
 150         iph = ip_hdr(skb);
 151         iph->version  = 4;
 152         iph->ihl      = 5;
 153         iph->tos      = inet->tos;
 154         if (ip_dont_fragment(sk, &rt->u.dst))
 155                 iph->frag_off = htons(IP_DF);
 156         else
 157                 iph->frag_off = 0;
 158         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 159         iph->daddr    = rt->rt_dst;
 160         iph->saddr    = rt->rt_src;
 161         iph->protocol = sk->sk_protocol;
 162         ip_select_ident(iph, &rt->u.dst, sk);
 163
 164         if (opt && opt->optlen) {
 165                 iph->ihl += opt->optlen>>2;
 166                 ip_options_build(skb, opt, daddr, rt, 0);
 167         }
 168
 169         skb->priority = sk->sk_priority;
 170         skb->mark = sk->sk_mark;
 171
 172         /* Send it out. */
 173         return ip_local_out(skb);
 174 }
 175
 176 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 177
 178 static inline int ip_finish_output2(struct sk_buff *skb)
 179 {
 180         struct dst_entry *dst = skb_dst(skb);
 181         struct rtable *rt = (struct rtable *)dst;
 182         struct net_device *dev = dst->dev;
 183         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 184
 185         if (rt->rt_type == RTN_MULTICAST) {
 186                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 187         } else if (rt->rt_type == RTN_BROADCAST)
 188                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 189
 190         /* Be paranoid, rather than too clever. */
 191         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 192                 struct sk_buff *skb2;
 193
 194                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 195                 if (skb2 == NULL) {
 196                         kfree_skb(skb);
 197                         return -ENOMEM;
 198                 }
 199                 if (skb->sk)
 200                         skb_set_owner_w(skb2, skb->sk);
 201                 kfree_skb(skb);
 202                 skb = skb2;
 203         }
 204
 205         if (dst->hh)
 206                 return neigh_hh_output(dst->hh, skb);
 207         else if (dst->neighbour)
 208                 return dst->neighbour->output(skb);
 209
 210         if (net_ratelimit())
 211                 printk(KERN_DEBUG "ip_finish_output2: No header cache and no neighbour!\n");
 212         kfree_skb(skb);
 213         return -EINVAL;
 214 }
 215
 216 static inline int ip_skb_dst_mtu(struct sk_buff *skb)
 217 {
 218         struct inet_sock *inet = skb->sk ? inet_sk(skb->sk) : NULL;
 219
 220         return (inet && inet->pmtudisc == IP_PMTUDISC_PROBE) ?
 221                skb_dst(skb)->dev->mtu : dst_mtu(skb_dst(skb));
 222 }
 223
 224 static int ip_finish_output(struct sk_buff *skb)
 225 {
 226 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 227         /* Policy lookup after SNAT yielded a new policy */
 228         if (skb_dst(skb)->xfrm != NULL) {
 229                 IPCB(skb)->flags |= IPSKB_REROUTED;
 230                 return dst_output(skb);
 231         }
 232 #endif
 233         if (skb->len > ip_skb_dst_mtu(skb) && !skb_is_gso(skb))
 234                 return ip_fragment(skb, ip_finish_output2);
 235         else
 236                 return ip_finish_output2(skb);
 237 }
 238
 239 int ip_mc_output(struct sk_buff *skb)
 240 {
 241         struct sock *sk = skb->sk;
 242         struct rtable *rt = skb_rtable(skb);
 243         struct net_device *dev = rt->u.dst.dev;
 244
 245         /*
 246          *      If the indicated interface is up and running, send the packet.
 247          */
 248         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 249
 250         skb->dev = dev;
 251         skb->protocol = htons(ETH_P_IP);
 252
 253         /*
 254          *      Multicasts are looped back for other local users
 255          */
 256
 257         if (rt->rt_flags&RTCF_MULTICAST) {
 258                 if (sk_mc_loop(sk)
 259 #ifdef CONFIG_IP_MROUTE
 260                 /* Small optimization: do not loopback not local frames,
 261                    which returned after forwarding; they will be  dropped
 262                    by ip_mr_input in any case.
 263                    Note, that local frames are looped back to be delivered
 264                    to local recipients.
 265
 266                    This check is duplicated in ip_mr_input at the moment.
 267                  */
 268                     &&
 269                     ((rt->rt_flags & RTCF_LOCAL) ||
 270                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 271 #endif
 272                    ) {
 273                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 274                         if (newskb)
 275                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 276                                         newskb, NULL, newskb->dev,
 277                                         ip_dev_loopback_xmit);
 278                 }
 279
 280                 /* Multicasts with ttl 0 must not go beyond the host */
 281
 282                 if (ip_hdr(skb)->ttl == 0) {
 283                         kfree_skb(skb);
 284                         return 0;
 285                 }
 286         }
 287
 288         if (rt->rt_flags&RTCF_BROADCAST) {
 289                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 290                 if (newskb)
 291                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 292                                 NULL, newskb->dev, ip_dev_loopback_xmit);
 293         }
 294
 295         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 296                             skb->dev, ip_finish_output,
 297                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 298 }
 299
 300 int ip_output(struct sk_buff *skb)
 301 {
 302         struct net_device *dev = skb_dst(skb)->dev;
 303
 304         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 305
 306         skb->dev = dev;
 307         skb->protocol = htons(ETH_P_IP);
 308
 309         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 310                             ip_finish_output,
 311                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 312 }
 313
 314 int ip_queue_xmit(struct sk_buff *skb)
 315 {
 316         struct sock *sk = skb->sk;
 317         struct inet_sock *inet = inet_sk(sk);
 318         struct ip_options *opt = inet->opt;
 319         struct rtable *rt;
 320         struct iphdr *iph;
 321         int res;
 322
 323         /* Skip all of this if the packet is already routed,
 324          * f.e. by something like SCTP.
 325          */
 326         rcu_read_lock();
 327         rt = skb_rtable(skb);
 328         if (rt != NULL)
 329                 goto packet_routed;
 330
 331         /* Make sure we can route this packet. */
 332         rt = (struct rtable *)__sk_dst_check(sk, 0);
 333         if (rt == NULL) {
 334                 __be32 daddr;
 335
 336                 /* Use correct destination address if we have options. */
 337                 daddr = inet->inet_daddr;
 338                 if(opt && opt->srr)
 339                         daddr = opt->faddr;
 340
 341                 {
 342                         struct flowi fl = { .oif = sk->sk_bound_dev_if,
 343                                             .mark = sk->sk_mark,
 344                                             .nl_u = { .ip4_u =
 345                                                       { .daddr = daddr,
 346                                                         .saddr = inet->inet_saddr,
 347                                                         .tos = RT_CONN_FLAGS(sk) } },
 348                                             .proto = sk->sk_protocol,
 349                                             .flags = inet_sk_flowi_flags(sk),
 350                                             .uli_u = { .ports =
 351                                                        { .sport = inet->inet_sport,
 352                                                          .dport = inet->inet_dport } } };
 353
 354                         /* If this fails, retransmit mechanism of transport layer will
 355                          * keep trying until route appears or the connection times
 356                          * itself out.
 357                          */
 358                         security_sk_classify_flow(sk, &fl);
 359                         if (ip_route_output_flow(sock_net(sk), &rt, &fl, sk, 0))
 360                                 goto no_route;
 361                 }
 362                 sk_setup_caps(sk, &rt->u.dst);
 363         }
 364         skb_dst_set_noref(skb, &rt->u.dst);
 365
 366 packet_routed:
 367         if (opt && opt->is_strictroute && rt->rt_dst != rt->rt_gateway)
 368                 goto no_route;
 369
 370         /* OK, we know where to send it, allocate and build IP header. */
 371         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->optlen : 0));
 372         skb_reset_network_header(skb);
 373         iph = ip_hdr(skb);
 374         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 375         if (ip_dont_fragment(sk, &rt->u.dst) && !skb->local_df)
 376                 iph->frag_off = htons(IP_DF);
 377         else
 378                 iph->frag_off = 0;
 379         iph->ttl      = ip_select_ttl(inet, &rt->u.dst);
 380         iph->protocol = sk->sk_protocol;
 381         iph->saddr    = rt->rt_src;
 382         iph->daddr    = rt->rt_dst;
 383         /* Transport layer set skb->h.foo itself. */
 384
 385         if (opt && opt->optlen) {
 386                 iph->ihl += opt->optlen >> 2;
 387                 ip_options_build(skb, opt, inet->inet_daddr, rt, 0);
 388         }
 389
 390         ip_select_ident_more(iph, &rt->u.dst, sk,
 391                              (skb_shinfo(skb)->gso_segs ?: 1) - 1);
 392
 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
 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_set(to, dst_clone(skb_dst(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 defined(CONFIG_NETFILTER_XT_TARGET_TRACE) || \
 426     defined(CONFIG_NETFILTER_XT_TARGET_TRACE_MODULE)
 427         to->nf_trace = from->nf_trace;
 428 #endif
 429 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 430         to->ipvs_property = from->ipvs_property;
 431 #endif
 432         skb_copy_secmark(to, from);
 433 }
 434
 435 /*
 436  *      This IP datagram is too large to be sent in one piece.  Break it up into
 437  *      smaller pieces (each of size equal to IP header plus
 438  *      a block of the data of the original IP data part) that will yet fit in a
 439  *      single device frame, and queue such a frame for sending.
 440  */
 441
 442 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 443 {
 444         struct iphdr *iph;
 445         int raw = 0;
 446         int ptr;
 447         struct net_device *dev;
 448         struct sk_buff *skb2;
 449         unsigned int mtu, hlen, left, len, ll_rs, pad;
 450         int offset;
 451         __be16 not_last_frag;
 452         struct rtable *rt = skb_rtable(skb);
 453         int err = 0;
 454
 455         dev = rt->u.dst.dev;
 456
 457         /*
 458          *      Point into the IP datagram header.
 459          */
 460
 461         iph = ip_hdr(skb);
 462
 463         if (unlikely((iph->frag_off & htons(IP_DF)) && !skb->local_df)) {
 464                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 465                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 466                           htonl(ip_skb_dst_mtu(skb)));
 467                 kfree_skb(skb);
 468                 return -EMSGSIZE;
 469         }
 470
 471         /*
 472          *      Setup starting values.
 473          */
 474
 475         hlen = iph->ihl * 4;
 476         mtu = dst_mtu(&rt->u.dst) - hlen;       /* Size of data space */
 477 #ifdef CONFIG_BRIDGE_NETFILTER
 478         if (skb->nf_bridge)
 479                 mtu -= nf_bridge_mtu_reduction(skb);
 480 #endif
 481         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 482
 483         /* When frag_list is given, use it. First, check its validity:
 484          * some transformers could create wrong frag_list or break existing
 485          * one, it is not prohibited. In this case fall back to copying.
 486          *
 487          * LATER: this step can be merged to real generation of fragments,
 488          * we can switch to copy when see the first bad fragment.
 489          */
 490         if (skb_has_frags(skb)) {
 491                 struct sk_buff *frag;
 492                 int first_len = skb_pagelen(skb);
 493                 int truesizes = 0;
 494
 495                 if (first_len - hlen > mtu ||
 496                     ((first_len - hlen) & 7) ||
 497                     (iph->frag_off & htons(IP_MF|IP_OFFSET)) ||
 498                     skb_cloned(skb))
 499                         goto slow_path;
 500
 501                 skb_walk_frags(skb, frag) {
 502                         /* Correct geometry. */
 503                         if (frag->len > mtu ||
 504                             ((frag->len & 7) && frag->next) ||
 505                             skb_headroom(frag) < hlen)
 506                             goto slow_path;
 507
 508                         /* Partially cloned skb? */
 509                         if (skb_shared(frag))
 510                                 goto slow_path;
 511
 512                         BUG_ON(frag->sk);
 513                         if (skb->sk) {
 514                                 frag->sk = skb->sk;
 515                                 frag->destructor = sock_wfree;
 516                         }
 517                         truesizes += frag->truesize;
 518                 }
 519
 520                 /* Everything is OK. Generate! */
 521
 522                 err = 0;
 523                 offset = 0;
 524                 frag = skb_shinfo(skb)->frag_list;
 525                 skb_frag_list_init(skb);
 526                 skb->data_len = first_len - skb_headlen(skb);
 527                 skb->truesize -= truesizes;
 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
 581 slow_path:
 582         left = skb->len - hlen;         /* Space per frame */
 583         ptr = raw + hlen;               /* Where to start from */
 584
 585         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 586          * we need to make room for the encapsulating header
 587          */
 588         pad = nf_bridge_pad(skb);
 589         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->u.dst.dev, pad);
 590         mtu -= pad;
 591
 592         /*
 593          *      Fragment the datagram.
 594          */
 595
 596         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 597         not_last_frag = iph->frag_off & htons(IP_MF);
 598
 599         /*
 600          *      Keep copying data until we run out.
 601          */
 602
 603         while (left > 0) {
 604                 len = left;
 605                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 606                 if (len > mtu)
 607                         len = mtu;
 608                 /* IF: we are not sending upto and including the packet end
 609                    then align the next start on an eight byte boundary */
 610                 if (len < left) {
 611                         len &= ~7;
 612                 }
 613                 /*
 614                  *      Allocate buffer.
 615                  */
 616
 617                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 618                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 619                         err = -ENOMEM;
 620                         goto fail;
 621                 }
 622
 623                 /*
 624                  *      Set up data on packet
 625                  */
 626
 627                 ip_copy_metadata(skb2, skb);
 628                 skb_reserve(skb2, ll_rs);
 629                 skb_put(skb2, len + hlen);
 630                 skb_reset_network_header(skb2);
 631                 skb2->transport_header = skb2->network_header + hlen;
 632
 633                 /*
 634                  *      Charge the memory for the fragment to any owner
 635                  *      it might possess
 636                  */
 637
 638                 if (skb->sk)
 639                         skb_set_owner_w(skb2, skb->sk);
 640
 641                 /*
 642                  *      Copy the packet header into the new buffer.
 643                  */
 644
 645                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 646
 647                 /*
 648                  *      Copy a block of the IP datagram.
 649                  */
 650                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 651                         BUG();
 652                 left -= len;
 653
 654                 /*
 655                  *      Fill in the new header fields.
 656                  */
 657                 iph = ip_hdr(skb2);
 658                 iph->frag_off = htons((offset >> 3));
 659
 660                 /* ANK: dirty, but effective trick. Upgrade options only if
 661                  * the segment to be fragmented was THE FIRST (otherwise,
 662                  * options are already fixed) and make it ONCE
 663                  * on the initial skb, so that all the following fragments
 664                  * will inherit fixed options.
 665                  */
 666                 if (offset == 0)
 667                         ip_options_fragment(skb);
 668
 669                 /*
 670                  *      Added AC : If we are fragmenting a fragment that's not the
 671                  *                 last fragment then keep MF on each bit
 672                  */
 673                 if (left > 0 || not_last_frag)
 674                         iph->frag_off |= htons(IP_MF);
 675                 ptr += len;
 676                 offset += len;
 677
 678                 /*
 679                  *      Put this fragment into the sending queue.
 680                  */
 681                 iph->tot_len = htons(len + hlen);
 682
 683                 ip_send_check(iph);
 684
 685                 err = output(skb2);
 686                 if (err)
 687                         goto fail;
 688
 689                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 690         }
 691         kfree_skb(skb);
 692         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 693         return err;
 694
 695 fail:
 696         kfree_skb(skb);
 697         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 698         return err;
 699 }
 700
 701 EXPORT_SYMBOL(ip_fragment);
 702
 703 int
 704 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 705 {
 706         struct iovec *iov = from;
 707
 708         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 709                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 710                         return -EFAULT;
 711         } else {
 712                 __wsum csum = 0;
 713                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 714                         return -EFAULT;
 715                 skb->csum = csum_block_add(skb->csum, csum, odd);
 716         }
 717         return 0;
 718 }
 719
 720 static inline __wsum
 721 csum_page(struct page *page, int offset, int copy)
 722 {
 723         char *kaddr;
 724         __wsum csum;
 725         kaddr = kmap(page);
 726         csum = csum_partial(kaddr + offset, copy, 0);
 727         kunmap(page);
 728         return csum;
 729 }
 730
 731 static inline int ip_ufo_append_data(struct sock *sk,
 732                         int getfrag(void *from, char *to, int offset, int len,
 733                                int odd, struct sk_buff *skb),
 734                         void *from, int length, int hh_len, int fragheaderlen,
 735                         int transhdrlen, int mtu, unsigned int flags)
 736 {
 737         struct sk_buff *skb;
 738         int err;
 739
 740         /* There is support for UDP fragmentation offload by network
 741          * device, so create one single skb packet containing complete
 742          * udp datagram
 743          */
 744         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL) {
 745                 skb = sock_alloc_send_skb(sk,
 746                         hh_len + fragheaderlen + transhdrlen + 20,
 747                         (flags & MSG_DONTWAIT), &err);
 748
 749                 if (skb == NULL)
 750                         return err;
 751
 752                 /* reserve space for Hardware header */
 753                 skb_reserve(skb, hh_len);
 754
 755                 /* create space for UDP/IP header */
 756                 skb_put(skb, fragheaderlen + transhdrlen);
 757
 758                 /* initialize network header pointer */
 759                 skb_reset_network_header(skb);
 760
 761                 /* initialize protocol header pointer */
 762                 skb->transport_header = skb->network_header + fragheaderlen;
 763
 764                 skb->ip_summed = CHECKSUM_PARTIAL;
 765                 skb->csum = 0;
 766                 sk->sk_sndmsg_off = 0;
 767
 768                 /* specify the length of each IP datagram fragment */
 769                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
 770                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 771                 __skb_queue_tail(&sk->sk_write_queue, skb);
 772         }
 773
 774         return skb_append_datato_frags(sk, skb, getfrag, from,
 775                                        (length - transhdrlen));
 776 }
 777
 778 /*
 779  *      ip_append_data() and ip_append_page() can make one large IP datagram
 780  *      from many pieces of data. Each pieces will be holded on the socket
 781  *      until ip_push_pending_frames() is called. Each piece can be a page
 782  *      or non-page data.
 783  *
 784  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
 785  *      this interface potentially.
 786  *
 787  *      LATER: length must be adjusted by pad at tail, when it is required.
 788  */
 789 int ip_append_data(struct sock *sk,
 790                    int getfrag(void *from, char *to, int offset, int len,
 791                                int odd, struct sk_buff *skb),
 792                    void *from, int length, int transhdrlen,
 793                    struct ipcm_cookie *ipc, struct rtable **rtp,
 794                    unsigned int flags)
 795 {
 796         struct inet_sock *inet = inet_sk(sk);
 797         struct sk_buff *skb;
 798
 799         struct ip_options *opt = NULL;
 800         int hh_len;
 801         int exthdrlen;
 802         int mtu;
 803         int copy;
 804         int err;
 805         int offset = 0;
 806         unsigned int maxfraglen, fragheaderlen;
 807         int csummode = CHECKSUM_NONE;
 808         struct rtable *rt;
 809
 810         if (flags&MSG_PROBE)
 811                 return 0;
 812
 813         if (skb_queue_empty(&sk->sk_write_queue)) {
 814                 /*
 815                  * setup for corking.
 816                  */
 817                 opt = ipc->opt;
 818                 if (opt) {
 819                         if (inet->cork.opt == NULL) {
 820                                 inet->cork.opt = kmalloc(sizeof(struct ip_options) + 40, sk->sk_allocation);
 821                                 if (unlikely(inet->cork.opt == NULL))
 822                                         return -ENOBUFS;
 823                         }
 824                         memcpy(inet->cork.opt, opt, sizeof(struct ip_options)+opt->optlen);
 825                         inet->cork.flags |= IPCORK_OPT;
 826                         inet->cork.addr = ipc->addr;
 827                 }
 828                 rt = *rtp;
 829                 if (unlikely(!rt))
 830                         return -EFAULT;
 831                 /*
 832                  * We steal reference to this route, caller should not release it
 833                  */
 834                 *rtp = NULL;
 835                 inet->cork.fragsize = mtu = inet->pmtudisc == IP_PMTUDISC_PROBE ?
 836                                             rt->u.dst.dev->mtu :
 837                                             dst_mtu(rt->u.dst.path);
 838                 inet->cork.dst = &rt->u.dst;
 839                 inet->cork.length = 0;
 840                 sk->sk_sndmsg_page = NULL;
 841                 sk->sk_sndmsg_off = 0;
 842                 if ((exthdrlen = rt->u.dst.header_len) != 0) {
 843                         length += exthdrlen;
 844                         transhdrlen += exthdrlen;
 845                 }
 846         } else {
 847                 rt = (struct rtable *)inet->cork.dst;
 848                 if (inet->cork.flags & IPCORK_OPT)
 849                         opt = inet->cork.opt;
 850
 851                 transhdrlen = 0;
 852                 exthdrlen = 0;
 853                 mtu = inet->cork.fragsize;
 854         }
 855         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
 856
 857         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 858         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 859
 860         if (inet->cork.length + length > 0xFFFF - fragheaderlen) {
 861                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport,
 862                                mtu-exthdrlen);
 863                 return -EMSGSIZE;
 864         }
 865
 866         /*
 867          * transhdrlen > 0 means that this is the first fragment and we wish
 868          * it won't be fragmented in the future.
 869          */
 870         if (transhdrlen &&
 871             length + fragheaderlen <= mtu &&
 872             rt->u.dst.dev->features & NETIF_F_V4_CSUM &&
 873             !exthdrlen)
 874                 csummode = CHECKSUM_PARTIAL;
 875
 876         skb = skb_peek_tail(&sk->sk_write_queue);
 877
 878         inet->cork.length += length;
 879         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 880             (sk->sk_protocol == IPPROTO_UDP) &&
 881             (rt->u.dst.dev->features & NETIF_F_UFO)) {
 882                 err = ip_ufo_append_data(sk, getfrag, from, length, hh_len,
 883                                          fragheaderlen, transhdrlen, mtu,
 884                                          flags);
 885                 if (err)
 886                         goto error;
 887                 return 0;
 888         }
 889
 890         /* So, what's going on in the loop below?
 891          *
 892          * We use calculated fragment length to generate chained skb,
 893          * each of segments is IP fragment ready for sending to network after
 894          * adding appropriate IP header.
 895          */
 896
 897         if (!skb)
 898                 goto alloc_new_skb;
 899
 900         while (length > 0) {
 901                 /* Check if the remaining data fits into current packet. */
 902                 copy = mtu - skb->len;
 903                 if (copy < length)
 904                         copy = maxfraglen - skb->len;
 905                 if (copy <= 0) {
 906                         char *data;
 907                         unsigned int datalen;
 908                         unsigned int fraglen;
 909                         unsigned int fraggap;
 910                         unsigned int alloclen;
 911                         struct sk_buff *skb_prev;
 912 alloc_new_skb:
 913                         skb_prev = skb;
 914                         if (skb_prev)
 915                                 fraggap = skb_prev->len - maxfraglen;
 916                         else
 917                                 fraggap = 0;
 918
 919                         /*
 920                          * If remaining data exceeds the mtu,
 921                          * we know we need more fragment(s).
 922                          */
 923                         datalen = length + fraggap;
 924                         if (datalen > mtu - fragheaderlen)
 925                                 datalen = maxfraglen - fragheaderlen;
 926                         fraglen = datalen + fragheaderlen;
 927
 928                         if ((flags & MSG_MORE) &&
 929                             !(rt->u.dst.dev->features&NETIF_F_SG))
 930                                 alloclen = mtu;
 931                         else
 932                                 alloclen = datalen + fragheaderlen;
 933
 934                         /* The last fragment gets additional space at tail.
 935                          * Note, with MSG_MORE we overallocate on fragments,
 936                          * because we have no idea what fragment will be
 937                          * the last.
 938                          */
 939                         if (datalen == length + fraggap)
 940                                 alloclen += rt->u.dst.trailer_len;
 941
 942                         if (transhdrlen) {
 943                                 skb = sock_alloc_send_skb(sk,
 944                                                 alloclen + hh_len + 15,
 945                                                 (flags & MSG_DONTWAIT), &err);
 946                         } else {
 947                                 skb = NULL;
 948                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 949                                     2 * sk->sk_sndbuf)
 950                                         skb = sock_wmalloc(sk,
 951                                                            alloclen + hh_len + 15, 1,
 952                                                            sk->sk_allocation);
 953                                 if (unlikely(skb == NULL))
 954                                         err = -ENOBUFS;
 955                                 else
 956                                         /* only the initial fragment is
 957                                            time stamped */
 958                                         ipc->shtx.flags = 0;
 959                         }
 960                         if (skb == NULL)
 961                                 goto error;
 962
 963                         /*
 964                          *      Fill in the control structures
 965                          */
 966                         skb->ip_summed = csummode;
 967                         skb->csum = 0;
 968                         skb_reserve(skb, hh_len);
 969                         *skb_tx(skb) = ipc->shtx;
 970
 971                         /*
 972                          *      Find where to start putting bytes.
 973                          */
 974                         data = skb_put(skb, fraglen);
 975                         skb_set_network_header(skb, exthdrlen);
 976                         skb->transport_header = (skb->network_header +
 977                                                  fragheaderlen);
 978                         data += fragheaderlen;
 979
 980                         if (fraggap) {
 981                                 skb->csum = skb_copy_and_csum_bits(
 982                                         skb_prev, maxfraglen,
 983                                         data + transhdrlen, fraggap, 0);
 984                                 skb_prev->csum = csum_sub(skb_prev->csum,
 985                                                           skb->csum);
 986                                 data += fraggap;
 987                                 pskb_trim_unique(skb_prev, maxfraglen);
 988                         }
 989
 990                         copy = datalen - transhdrlen - fraggap;
 991                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
 992                                 err = -EFAULT;
 993                                 kfree_skb(skb);
 994                                 goto error;
 995                         }
 996
 997                         offset += copy;
 998                         length -= datalen - fraggap;
 999                         transhdrlen = 0;
1000                         exthdrlen = 0;
1001                         csummode = CHECKSUM_NONE;
1002
1003                         /*
1004                          * Put the packet on the pending queue.
1005                          */
1006                         __skb_queue_tail(&sk->sk_write_queue, skb);
1007                         continue;
1008                 }
1009
1010                 if (copy > length)
1011                         copy = length;
1012
1013                 if (!(rt->u.dst.dev->features&NETIF_F_SG)) {
1014                         unsigned int off;
1015
1016                         off = skb->len;
1017                         if (getfrag(from, skb_put(skb, copy),
1018                                         offset, copy, off, skb) < 0) {
1019                                 __skb_trim(skb, off);
1020                                 err = -EFAULT;
1021                                 goto error;
1022                         }
1023                 } else {
1024                         int i = skb_shinfo(skb)->nr_frags;
1025                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i-1];
1026                         struct page *page = sk->sk_sndmsg_page;
1027                         int off = sk->sk_sndmsg_off;
1028                         unsigned int left;
1029
1030                         if (page && (left = PAGE_SIZE - off) > 0) {
1031                                 if (copy >= left)
1032                                         copy = left;
1033                                 if (page != frag->page) {
1034                                         if (i == MAX_SKB_FRAGS) {
1035                                                 err = -EMSGSIZE;
1036                                                 goto error;
1037                                         }
1038                                         get_page(page);
1039                                         skb_fill_page_desc(skb, i, page, sk->sk_sndmsg_off, 0);
1040                                         frag = &skb_shinfo(skb)->frags[i];
1041                                 }
1042                         } else if (i < MAX_SKB_FRAGS) {
1043                                 if (copy > PAGE_SIZE)
1044                                         copy = PAGE_SIZE;
1045                                 page = alloc_pages(sk->sk_allocation, 0);
1046                                 if (page == NULL)  {
1047                                         err = -ENOMEM;
1048                                         goto error;
1049                                 }
1050                                 sk->sk_sndmsg_page = page;
1051                                 sk->sk_sndmsg_off = 0;
1052
1053                                 skb_fill_page_desc(skb, i, page, 0, 0);
1054                                 frag = &skb_shinfo(skb)->frags[i];
1055                         } else {
1056                                 err = -EMSGSIZE;
1057                                 goto error;
1058                         }
1059                         if (getfrag(from, page_address(frag->page)+frag->page_offset+frag->size, offset, copy, skb->len, skb) < 0) {
1060                                 err = -EFAULT;
1061                                 goto error;
1062                         }
1063                         sk->sk_sndmsg_off += copy;
1064                         frag->size += copy;
1065                         skb->len += copy;
1066                         skb->data_len += copy;
1067                         skb->truesize += copy;
1068                         atomic_add(copy, &sk->sk_wmem_alloc);
1069                 }
1070                 offset += copy;
1071                 length -= copy;
1072         }
1073
1074         return 0;
1075
1076 error:
1077         inet->cork.length -= length;
1078         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1079         return err;
1080 }
1081
1082 ssize_t ip_append_page(struct sock *sk, struct page *page,
1083                        int offset, size_t size, int flags)
1084 {
1085         struct inet_sock *inet = inet_sk(sk);
1086         struct sk_buff *skb;
1087         struct rtable *rt;
1088         struct ip_options *opt = NULL;
1089         int hh_len;
1090         int mtu;
1091         int len;
1092         int err;
1093         unsigned int maxfraglen, fragheaderlen, fraggap;
1094
1095         if (inet->hdrincl)
1096                 return -EPERM;
1097
1098         if (flags&MSG_PROBE)
1099                 return 0;
1100
1101         if (skb_queue_empty(&sk->sk_write_queue))
1102                 return -EINVAL;
1103
1104         rt = (struct rtable *)inet->cork.dst;
1105         if (inet->cork.flags & IPCORK_OPT)
1106                 opt = inet->cork.opt;
1107
1108         if (!(rt->u.dst.dev->features&NETIF_F_SG))
1109                 return -EOPNOTSUPP;
1110
1111         hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
1112         mtu = inet->cork.fragsize;
1113
1114         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1115         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1116
1117         if (inet->cork.length + size > 0xFFFF - fragheaderlen) {
1118                 ip_local_error(sk, EMSGSIZE, rt->rt_dst, inet->inet_dport, mtu);
1119                 return -EMSGSIZE;
1120         }
1121
1122         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1123                 return -EINVAL;
1124
1125         inet->cork.length += size;
1126         if ((size + skb->len > mtu) &&
1127             (sk->sk_protocol == IPPROTO_UDP) &&
1128             (rt->u.dst.dev->features & NETIF_F_UFO)) {
1129                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1130                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1131         }
1132
1133
1134         while (size > 0) {
1135                 int i;
1136
1137                 if (skb_is_gso(skb))
1138                         len = size;
1139                 else {
1140
1141                         /* Check if the remaining data fits into current packet. */
1142                         len = mtu - skb->len;
1143                         if (len < size)
1144                                 len = maxfraglen - skb->len;
1145                 }
1146                 if (len <= 0) {
1147                         struct sk_buff *skb_prev;
1148                         int alloclen;
1149
1150                         skb_prev = skb;
1151                         fraggap = skb_prev->len - maxfraglen;
1152
1153                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1154                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1155                         if (unlikely(!skb)) {
1156                                 err = -ENOBUFS;
1157                                 goto error;
1158                         }
1159
1160                         /*
1161                          *      Fill in the control structures
1162                          */
1163                         skb->ip_summed = CHECKSUM_NONE;
1164                         skb->csum = 0;
1165                         skb_reserve(skb, hh_len);
1166
1167                         /*
1168                          *      Find where to start putting bytes.
1169                          */
1170                         skb_put(skb, fragheaderlen + fraggap);
1171                         skb_reset_network_header(skb);
1172                         skb->transport_header = (skb->network_header +
1173                                                  fragheaderlen);
1174                         if (fraggap) {
1175                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1176                                                                    maxfraglen,
1177                                                     skb_transport_header(skb),
1178                                                                    fraggap, 0);
1179                                 skb_prev->csum = csum_sub(skb_prev->csum,
1180                                                           skb->csum);
1181                                 pskb_trim_unique(skb_prev, maxfraglen);
1182                         }
1183
1184                         /*
1185                          * Put the packet on the pending queue.
1186                          */
1187                         __skb_queue_tail(&sk->sk_write_queue, skb);
1188                         continue;
1189                 }
1190
1191                 i = skb_shinfo(skb)->nr_frags;
1192                 if (len > size)
1193                         len = size;
1194                 if (skb_can_coalesce(skb, i, page, offset)) {
1195                         skb_shinfo(skb)->frags[i-1].size += len;
1196                 } else if (i < MAX_SKB_FRAGS) {
1197                         get_page(page);
1198                         skb_fill_page_desc(skb, i, page, offset, len);
1199                 } else {
1200                         err = -EMSGSIZE;
1201                         goto error;
1202                 }
1203
1204                 if (skb->ip_summed == CHECKSUM_NONE) {
1205                         __wsum csum;
1206                         csum = csum_page(page, offset, len);
1207                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1208                 }
1209
1210                 skb->len += len;
1211                 skb->data_len += len;
1212                 skb->truesize += len;
1213                 atomic_add(len, &sk->sk_wmem_alloc);
1214                 offset += len;
1215                 size -= len;
1216         }
1217         return 0;
1218
1219 error:
1220         inet->cork.length -= size;
1221         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1222         return err;
1223 }
1224
1225 static void ip_cork_release(struct inet_sock *inet)
1226 {
1227         inet->cork.flags &= ~IPCORK_OPT;
1228         kfree(inet->cork.opt);
1229         inet->cork.opt = NULL;
1230         dst_release(inet->cork.dst);
1231         inet->cork.dst = NULL;
1232 }
1233
1234 /*
1235  *      Combined all pending IP fragments on the socket as one IP datagram
1236  *      and push them out.
1237  */
1238 int ip_push_pending_frames(struct sock *sk)
1239 {
1240         struct sk_buff *skb, *tmp_skb;
1241         struct sk_buff **tail_skb;
1242         struct inet_sock *inet = inet_sk(sk);
1243         struct net *net = sock_net(sk);
1244         struct ip_options *opt = NULL;
1245         struct rtable *rt = (struct rtable *)inet->cork.dst;
1246         struct iphdr *iph;
1247         __be16 df = 0;
1248         __u8 ttl;
1249         int err = 0;
1250
1251         if ((skb = __skb_dequeue(&sk->sk_write_queue)) == NULL)
1252                 goto out;
1253         tail_skb = &(skb_shinfo(skb)->frag_list);
1254
1255         /* move skb->data to ip header from ext header */
1256         if (skb->data < skb_network_header(skb))
1257                 __skb_pull(skb, skb_network_offset(skb));
1258         while ((tmp_skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
1259                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1260                 *tail_skb = tmp_skb;
1261                 tail_skb = &(tmp_skb->next);
1262                 skb->len += tmp_skb->len;
1263                 skb->data_len += tmp_skb->len;
1264                 skb->truesize += tmp_skb->truesize;
1265                 tmp_skb->destructor = NULL;
1266                 tmp_skb->sk = NULL;
1267         }
1268
1269         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1270          * to fragment the frame generated here. No matter, what transforms
1271          * how transforms change size of the packet, it will come out.
1272          */
1273         if (inet->pmtudisc < IP_PMTUDISC_DO)
1274                 skb->local_df = 1;
1275
1276         /* DF bit is set when we want to see DF on outgoing frames.
1277          * If local_df is set too, we still allow to fragment this frame
1278          * locally. */
1279         if (inet->pmtudisc >= IP_PMTUDISC_DO ||
1280             (skb->len <= dst_mtu(&rt->u.dst) &&
1281              ip_dont_fragment(sk, &rt->u.dst)))
1282                 df = htons(IP_DF);
1283
1284         if (inet->cork.flags & IPCORK_OPT)
1285                 opt = inet->cork.opt;
1286
1287         if (rt->rt_type == RTN_MULTICAST)
1288                 ttl = inet->mc_ttl;
1289         else
1290                 ttl = ip_select_ttl(inet, &rt->u.dst);
1291
1292         iph = (struct iphdr *)skb->data;
1293         iph->version = 4;
1294         iph->ihl = 5;
1295         if (opt) {
1296                 iph->ihl += opt->optlen>>2;
1297                 ip_options_build(skb, opt, inet->cork.addr, rt, 0);
1298         }
1299         iph->tos = inet->tos;
1300         iph->frag_off = df;
1301         ip_select_ident(iph, &rt->u.dst, sk);
1302         iph->ttl = ttl;
1303         iph->protocol = sk->sk_protocol;
1304         iph->saddr = rt->rt_src;
1305         iph->daddr = rt->rt_dst;
1306
1307         skb->priority = sk->sk_priority;
1308         skb->mark = sk->sk_mark;
1309         /*
1310          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1311          * on dst refcount
1312          */
1313         inet->cork.dst = NULL;
1314         skb_dst_set(skb, &rt->u.dst);
1315
1316         if (iph->protocol == IPPROTO_ICMP)
1317                 icmp_out_count(net, ((struct icmphdr *)
1318                         skb_transport_header(skb))->type);
1319
1320         /* Netfilter gets whole the not fragmented skb. */
1321         err = ip_local_out(skb);
1322         if (err) {
1323                 if (err > 0)
1324                         err = net_xmit_errno(err);
1325                 if (err)
1326                         goto error;
1327         }
1328
1329 out:
1330         ip_cork_release(inet);
1331         return err;
1332
1333 error:
1334         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1335         goto out;
1336 }
1337
1338 /*
1339  *      Throw away all pending data on the socket.
1340  */
1341 void ip_flush_pending_frames(struct sock *sk)
1342 {
1343         struct sk_buff *skb;
1344
1345         while ((skb = __skb_dequeue_tail(&sk->sk_write_queue)) != NULL)
1346                 kfree_skb(skb);
1347
1348         ip_cork_release(inet_sk(sk));
1349 }
1350
1351
1352 /*
1353  *      Fetch data from kernel space and fill in checksum if needed.
1354  */
1355 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1356                               int len, int odd, struct sk_buff *skb)
1357 {
1358         __wsum csum;
1359
1360         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1361         skb->csum = csum_block_add(skb->csum, csum, odd);
1362         return 0;
1363 }
1364
1365 /*
1366  *      Generic function to send a packet as reply to another packet.
1367  *      Used to send TCP resets so far. ICMP should use this function too.
1368  *
1369  *      Should run single threaded per socket because it uses the sock
1370  *      structure to pass arguments.
1371  */
1372 void ip_send_reply(struct sock *sk, struct sk_buff *skb, struct ip_reply_arg *arg,
1373                    unsigned int len)
1374 {
1375         struct inet_sock *inet = inet_sk(sk);
1376         struct {
1377                 struct ip_options       opt;
1378                 char                    data[40];
1379         } replyopts;
1380         struct ipcm_cookie ipc;
1381         __be32 daddr;
1382         struct rtable *rt = skb_rtable(skb);
1383
1384         if (ip_options_echo(&replyopts.opt, skb))
1385                 return;
1386
1387         daddr = ipc.addr = rt->rt_src;
1388         ipc.opt = NULL;
1389         ipc.shtx.flags = 0;
1390
1391         if (replyopts.opt.optlen) {
1392                 ipc.opt = &replyopts.opt;
1393
1394                 if (ipc.opt->srr)
1395                         daddr = replyopts.opt.faddr;
1396         }
1397
1398         {
1399                 struct flowi fl = { .oif = arg->bound_dev_if,
1400                                     .nl_u = { .ip4_u =
1401                                               { .daddr = daddr,
1402                                                 .saddr = rt->rt_spec_dst,
1403                                                 .tos = RT_TOS(ip_hdr(skb)->tos) } },
1404                                     /* Not quite clean, but right. */
1405                                     .uli_u = { .ports =
1406                                                { .sport = tcp_hdr(skb)->dest,
1407                                                  .dport = tcp_hdr(skb)->source } },
1408                                     .proto = sk->sk_protocol,
1409                                     .flags = ip_reply_arg_flowi_flags(arg) };
1410                 security_skb_classify_flow(skb, &fl);
1411                 if (ip_route_output_key(sock_net(sk), &rt, &fl))
1412                         return;
1413         }
1414
1415         /* And let IP do all the hard work.
1416
1417            This chunk is not reenterable, hence spinlock.
1418            Note that it uses the fact, that this function is called
1419            with locally disabled BH and that sk cannot be already spinlocked.
1420          */
1421         bh_lock_sock(sk);
1422         inet->tos = ip_hdr(skb)->tos;
1423         sk->sk_priority = skb->priority;
1424         sk->sk_protocol = ip_hdr(skb)->protocol;
1425         sk->sk_bound_dev_if = arg->bound_dev_if;
1426         ip_append_data(sk, ip_reply_glue_bits, arg->iov->iov_base, len, 0,
1427                        &ipc, &rt, MSG_DONTWAIT);
1428         if ((skb = skb_peek(&sk->sk_write_queue)) != NULL) {
1429                 if (arg->csumoffset >= 0)
1430                         *((__sum16 *)skb_transport_header(skb) +
1431                           arg->csumoffset) = csum_fold(csum_add(skb->csum,
1432                                                                 arg->csum));
1433                 skb->ip_summed = CHECKSUM_NONE;
1434                 ip_push_pending_frames(sk);
1435         }
1436
1437         bh_unlock_sock(sk);
1438
1439         ip_rt_put(rt);
1440 }
1441
1442 void __init ip_init(void)
1443 {
1444         ip_rt_init();
1445         inet_initpeers();
1446
1447 #if defined(CONFIG_IP_MULTICAST) && defined(CONFIG_PROC_FS)
1448         igmp_mc_proc_init();
1449 #endif
1450 }
1451
1452 EXPORT_SYMBOL(ip_generic_getfrag);
1453 EXPORT_SYMBOL(ip_queue_xmit);
1454 EXPORT_SYMBOL(ip_send_check);