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         skb->protocol = htons(ETH_P_IP);
 101
 102         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, skb, NULL,
 103                        skb_dst(skb)->dev, dst_output);
 104 }
 105
 106 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
 107 {
 108         int err;
 109
 110         err = __ip_local_out(skb);
 111         if (likely(err == 1))
 112                 err = dst_output_sk(sk, skb);
 113
 114         return err;
 115 }
 116 EXPORT_SYMBOL_GPL(ip_local_out_sk);
 117
 118 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 119 {
 120         int ttl = inet->uc_ttl;
 121
 122         if (ttl < 0)
 123                 ttl = ip4_dst_hoplimit(dst);
 124         return ttl;
 125 }
 126
 127 /*
 128  *              Add an ip header to a skbuff and send it out.
 129  *
 130  */
 131 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 132                           __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
 133 {
 134         struct inet_sock *inet = inet_sk(sk);
 135         struct rtable *rt = skb_rtable(skb);
 136         struct iphdr *iph;
 137
 138         /* Build the IP header. */
 139         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
 140         skb_reset_network_header(skb);
 141         iph = ip_hdr(skb);
 142         iph->version  = 4;
 143         iph->ihl      = 5;
 144         iph->tos      = inet->tos;
 145         if (ip_dont_fragment(sk, &rt->dst))
 146                 iph->frag_off = htons(IP_DF);
 147         else
 148                 iph->frag_off = 0;
 149         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 150         iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
 151         iph->saddr    = saddr;
 152         iph->protocol = sk->sk_protocol;
 153         ip_select_ident(sock_net(sk), skb, sk);
 154
 155         if (opt && opt->opt.optlen) {
 156                 iph->ihl += opt->opt.optlen>>2;
 157                 ip_options_build(skb, &opt->opt, daddr, rt, 0);
 158         }
 159
 160         skb->priority = sk->sk_priority;
 161         if (!skb->mark)
 162                 skb->mark = sk->sk_mark;
 163
 164         /* Send it out. */
 165         return ip_local_out(skb);
 166 }
 167 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 168
 169 static inline int ip_finish_output2(struct sk_buff *skb)
 170 {
 171         struct dst_entry *dst = skb_dst(skb);
 172         struct rtable *rt = (struct rtable *)dst;
 173         struct net_device *dev = dst->dev;
 174         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 175         struct neighbour *neigh;
 176         u32 nexthop;
 177
 178         if (rt->rt_type == RTN_MULTICAST) {
 179                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 180         } else if (rt->rt_type == RTN_BROADCAST)
 181                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 182
 183         /* Be paranoid, rather than too clever. */
 184         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 185                 struct sk_buff *skb2;
 186
 187                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 188                 if (skb2 == NULL) {
 189                         kfree_skb(skb);
 190                         return -ENOMEM;
 191                 }
 192                 if (skb->sk)
 193                         skb_set_owner_w(skb2, skb->sk);
 194                 consume_skb(skb);
 195                 skb = skb2;
 196         }
 197
 198         rcu_read_lock_bh();
 199         nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
 200         neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
 201         if (unlikely(!neigh))
 202                 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
 203         if (!IS_ERR(neigh)) {
 204                 int res = dst_neigh_output(dst, neigh, skb);
 205
 206                 rcu_read_unlock_bh();
 207                 return res;
 208         }
 209         rcu_read_unlock_bh();
 210
 211         net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
 212                             __func__);
 213         kfree_skb(skb);
 214         return -EINVAL;
 215 }
 216
 217 static int ip_finish_output_gso(struct sk_buff *skb)
 218 {
 219         netdev_features_t features;
 220         struct sk_buff *segs;
 221         int ret = 0;
 222
 223         /* common case: seglen is <= mtu */
 224         if (skb_gso_network_seglen(skb) <= ip_skb_dst_mtu(skb))
 225                 return ip_finish_output2(skb);
 226
 227         /* Slowpath -  GSO segment length is exceeding the dst MTU.
 228          *
 229          * This can happen in two cases:
 230          * 1) TCP GRO packet, DF bit not set
 231          * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
 232          * from host network stack.
 233          */
 234         features = netif_skb_features(skb);
 235         segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 236         if (IS_ERR_OR_NULL(segs)) {
 237                 kfree_skb(skb);
 238                 return -ENOMEM;
 239         }
 240
 241         consume_skb(skb);
 242
 243         do {
 244                 struct sk_buff *nskb = segs->next;
 245                 int err;
 246
 247                 segs->next = NULL;
 248                 err = ip_fragment(segs, ip_finish_output2);
 249
 250                 if (err && ret == 0)
 251                         ret = err;
 252                 segs = nskb;
 253         } while (segs);
 254
 255         return ret;
 256 }
 257
 258 static int ip_finish_output(struct sk_buff *skb)
 259 {
 260 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 261         /* Policy lookup after SNAT yielded a new policy */
 262         if (skb_dst(skb)->xfrm != NULL) {
 263                 IPCB(skb)->flags |= IPSKB_REROUTED;
 264                 return dst_output(skb);
 265         }
 266 #endif
 267         if (skb_is_gso(skb))
 268                 return ip_finish_output_gso(skb);
 269
 270         if (skb->len > ip_skb_dst_mtu(skb))
 271                 return ip_fragment(skb, ip_finish_output2);
 272
 273         return ip_finish_output2(skb);
 274 }
 275
 276 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
 277 {
 278         struct rtable *rt = skb_rtable(skb);
 279         struct net_device *dev = rt->dst.dev;
 280
 281         /*
 282          *      If the indicated interface is up and running, send the packet.
 283          */
 284         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 285
 286         skb->dev = dev;
 287         skb->protocol = htons(ETH_P_IP);
 288
 289         /*
 290          *      Multicasts are looped back for other local users
 291          */
 292
 293         if (rt->rt_flags&RTCF_MULTICAST) {
 294                 if (sk_mc_loop(sk)
 295 #ifdef CONFIG_IP_MROUTE
 296                 /* Small optimization: do not loopback not local frames,
 297                    which returned after forwarding; they will be  dropped
 298                    by ip_mr_input in any case.
 299                    Note, that local frames are looped back to be delivered
 300                    to local recipients.
 301
 302                    This check is duplicated in ip_mr_input at the moment.
 303                  */
 304                     &&
 305                     ((rt->rt_flags & RTCF_LOCAL) ||
 306                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 307 #endif
 308                    ) {
 309                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 310                         if (newskb)
 311                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 312                                         newskb, NULL, newskb->dev,
 313                                         dev_loopback_xmit);
 314                 }
 315
 316                 /* Multicasts with ttl 0 must not go beyond the host */
 317
 318                 if (ip_hdr(skb)->ttl == 0) {
 319                         kfree_skb(skb);
 320                         return 0;
 321                 }
 322         }
 323
 324         if (rt->rt_flags&RTCF_BROADCAST) {
 325                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 326                 if (newskb)
 327                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, newskb,
 328                                 NULL, newskb->dev, dev_loopback_xmit);
 329         }
 330
 331         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL,
 332                             skb->dev, ip_finish_output,
 333                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 334 }
 335
 336 int ip_output(struct sock *sk, struct sk_buff *skb)
 337 {
 338         struct net_device *dev = skb_dst(skb)->dev;
 339
 340         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 341
 342         skb->dev = dev;
 343         skb->protocol = htons(ETH_P_IP);
 344
 345         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, skb, NULL, dev,
 346                             ip_finish_output,
 347                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 348 }
 349
 350 /*
 351  * copy saddr and daddr, possibly using 64bit load/stores
 352  * Equivalent to :
 353  *   iph->saddr = fl4->saddr;
 354  *   iph->daddr = fl4->daddr;
 355  */
 356 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 357 {
 358         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 359                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 360         memcpy(&iph->saddr, &fl4->saddr,
 361                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 362 }
 363
 364 /* Note: skb->sk can be different from sk, in case of tunnels */
 365 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
 366 {
 367         struct inet_sock *inet = inet_sk(sk);
 368         struct ip_options_rcu *inet_opt;
 369         struct flowi4 *fl4;
 370         struct rtable *rt;
 371         struct iphdr *iph;
 372         int res;
 373
 374         /* Skip all of this if the packet is already routed,
 375          * f.e. by something like SCTP.
 376          */
 377         rcu_read_lock();
 378         inet_opt = rcu_dereference(inet->inet_opt);
 379         fl4 = &fl->u.ip4;
 380         rt = skb_rtable(skb);
 381         if (rt != NULL)
 382                 goto packet_routed;
 383
 384         /* Make sure we can route this packet. */
 385         rt = (struct rtable *)__sk_dst_check(sk, 0);
 386         if (rt == NULL) {
 387                 __be32 daddr;
 388
 389                 /* Use correct destination address if we have options. */
 390                 daddr = inet->inet_daddr;
 391                 if (inet_opt && inet_opt->opt.srr)
 392                         daddr = inet_opt->opt.faddr;
 393
 394                 /* If this fails, retransmit mechanism of transport layer will
 395                  * keep trying until route appears or the connection times
 396                  * itself out.
 397                  */
 398                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 399                                            daddr, inet->inet_saddr,
 400                                            inet->inet_dport,
 401                                            inet->inet_sport,
 402                                            sk->sk_protocol,
 403                                            RT_CONN_FLAGS(sk),
 404                                            sk->sk_bound_dev_if);
 405                 if (IS_ERR(rt))
 406                         goto no_route;
 407                 sk_setup_caps(sk, &rt->dst);
 408         }
 409         skb_dst_set_noref(skb, &rt->dst);
 410
 411 packet_routed:
 412         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
 413                 goto no_route;
 414
 415         /* OK, we know where to send it, allocate and build IP header. */
 416         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 417         skb_reset_network_header(skb);
 418         iph = ip_hdr(skb);
 419         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 420         if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
 421                 iph->frag_off = htons(IP_DF);
 422         else
 423                 iph->frag_off = 0;
 424         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 425         iph->protocol = sk->sk_protocol;
 426         ip_copy_addrs(iph, fl4);
 427
 428         /* Transport layer set skb->h.foo itself. */
 429
 430         if (inet_opt && inet_opt->opt.optlen) {
 431                 iph->ihl += inet_opt->opt.optlen >> 2;
 432                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 433         }
 434
 435         ip_select_ident_segs(sock_net(sk), skb, sk,
 436                              skb_shinfo(skb)->gso_segs ?: 1);
 437
 438         /* TODO : should we use skb->sk here instead of sk ? */
 439         skb->priority = sk->sk_priority;
 440         skb->mark = sk->sk_mark;
 441
 442         res = ip_local_out_sk(sk, skb);
 443         rcu_read_unlock();
 444         return res;
 445
 446 no_route:
 447         rcu_read_unlock();
 448         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 449         kfree_skb(skb);
 450         return -EHOSTUNREACH;
 451 }
 452 EXPORT_SYMBOL(ip_queue_xmit);
 453
 454
 455 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 456 {
 457         to->pkt_type = from->pkt_type;
 458         to->priority = from->priority;
 459         to->protocol = from->protocol;
 460         skb_dst_drop(to);
 461         skb_dst_copy(to, from);
 462         to->dev = from->dev;
 463         to->mark = from->mark;
 464
 465         /* Copy the flags to each fragment. */
 466         IPCB(to)->flags = IPCB(from)->flags;
 467
 468 #ifdef CONFIG_NET_SCHED
 469         to->tc_index = from->tc_index;
 470 #endif
 471         nf_copy(to, from);
 472 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 473         to->ipvs_property = from->ipvs_property;
 474 #endif
 475         skb_copy_secmark(to, from);
 476 }
 477
 478 /*
 479  *      This IP datagram is too large to be sent in one piece.  Break it up into
 480  *      smaller pieces (each of size equal to IP header plus
 481  *      a block of the data of the original IP data part) that will yet fit in a
 482  *      single device frame, and queue such a frame for sending.
 483  */
 484
 485 int ip_fragment(struct sk_buff *skb, int (*output)(struct sk_buff *))
 486 {
 487         struct iphdr *iph;
 488         int ptr;
 489         struct net_device *dev;
 490         struct sk_buff *skb2;
 491         unsigned int mtu, hlen, left, len, ll_rs;
 492         int offset;
 493         __be16 not_last_frag;
 494         struct rtable *rt = skb_rtable(skb);
 495         int err = 0;
 496
 497         dev = rt->dst.dev;
 498
 499         /*
 500          *      Point into the IP datagram header.
 501          */
 502
 503         iph = ip_hdr(skb);
 504
 505         mtu = ip_skb_dst_mtu(skb);
 506         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->ignore_df) ||
 507                      (IPCB(skb)->frag_max_size &&
 508                       IPCB(skb)->frag_max_size > mtu))) {
 509                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 510                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 511                           htonl(mtu));
 512                 kfree_skb(skb);
 513                 return -EMSGSIZE;
 514         }
 515
 516         /*
 517          *      Setup starting values.
 518          */
 519
 520         hlen = iph->ihl * 4;
 521         mtu = mtu - hlen;       /* Size of data space */
 522 #ifdef CONFIG_BRIDGE_NETFILTER
 523         if (skb->nf_bridge)
 524                 mtu -= nf_bridge_mtu_reduction(skb);
 525 #endif
 526         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 527
 528         /* When frag_list is given, use it. First, check its validity:
 529          * some transformers could create wrong frag_list or break existing
 530          * one, it is not prohibited. In this case fall back to copying.
 531          *
 532          * LATER: this step can be merged to real generation of fragments,
 533          * we can switch to copy when see the first bad fragment.
 534          */
 535         if (skb_has_frag_list(skb)) {
 536                 struct sk_buff *frag, *frag2;
 537                 int first_len = skb_pagelen(skb);
 538
 539                 if (first_len - hlen > mtu ||
 540                     ((first_len - hlen) & 7) ||
 541                     ip_is_fragment(iph) ||
 542                     skb_cloned(skb))
 543                         goto slow_path;
 544
 545                 skb_walk_frags(skb, frag) {
 546                         /* Correct geometry. */
 547                         if (frag->len > mtu ||
 548                             ((frag->len & 7) && frag->next) ||
 549                             skb_headroom(frag) < hlen)
 550                                 goto slow_path_clean;
 551
 552                         /* Partially cloned skb? */
 553                         if (skb_shared(frag))
 554                                 goto slow_path_clean;
 555
 556                         BUG_ON(frag->sk);
 557                         if (skb->sk) {
 558                                 frag->sk = skb->sk;
 559                                 frag->destructor = sock_wfree;
 560                         }
 561                         skb->truesize -= frag->truesize;
 562                 }
 563
 564                 /* Everything is OK. Generate! */
 565
 566                 err = 0;
 567                 offset = 0;
 568                 frag = skb_shinfo(skb)->frag_list;
 569                 skb_frag_list_init(skb);
 570                 skb->data_len = first_len - skb_headlen(skb);
 571                 skb->len = first_len;
 572                 iph->tot_len = htons(first_len);
 573                 iph->frag_off = htons(IP_MF);
 574                 ip_send_check(iph);
 575
 576                 for (;;) {
 577                         /* Prepare header of the next frame,
 578                          * before previous one went down. */
 579                         if (frag) {
 580                                 frag->ip_summed = CHECKSUM_NONE;
 581                                 skb_reset_transport_header(frag);
 582                                 __skb_push(frag, hlen);
 583                                 skb_reset_network_header(frag);
 584                                 memcpy(skb_network_header(frag), iph, hlen);
 585                                 iph = ip_hdr(frag);
 586                                 iph->tot_len = htons(frag->len);
 587                                 ip_copy_metadata(frag, skb);
 588                                 if (offset == 0)
 589                                         ip_options_fragment(frag);
 590                                 offset += skb->len - hlen;
 591                                 iph->frag_off = htons(offset>>3);
 592                                 if (frag->next != NULL)
 593                                         iph->frag_off |= htons(IP_MF);
 594                                 /* Ready, complete checksum */
 595                                 ip_send_check(iph);
 596                         }
 597
 598                         err = output(skb);
 599
 600                         if (!err)
 601                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 602                         if (err || !frag)
 603                                 break;
 604
 605                         skb = frag;
 606                         frag = skb->next;
 607                         skb->next = NULL;
 608                 }
 609
 610                 if (err == 0) {
 611                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 612                         return 0;
 613                 }
 614
 615                 while (frag) {
 616                         skb = frag->next;
 617                         kfree_skb(frag);
 618                         frag = skb;
 619                 }
 620                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 621                 return err;
 622
 623 slow_path_clean:
 624                 skb_walk_frags(skb, frag2) {
 625                         if (frag2 == frag)
 626                                 break;
 627                         frag2->sk = NULL;
 628                         frag2->destructor = NULL;
 629                         skb->truesize += frag2->truesize;
 630                 }
 631         }
 632
 633 slow_path:
 634         /* for offloaded checksums cleanup checksum before fragmentation */
 635         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 636                 goto fail;
 637         iph = ip_hdr(skb);
 638
 639         left = skb->len - hlen;         /* Space per frame */
 640         ptr = hlen;             /* Where to start from */
 641
 642         /* for bridged IP traffic encapsulated inside f.e. a vlan header,
 643          * we need to make room for the encapsulating header
 644          */
 645         ll_rs = LL_RESERVED_SPACE_EXTRA(rt->dst.dev, nf_bridge_pad(skb));
 646
 647         /*
 648          *      Fragment the datagram.
 649          */
 650
 651         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 652         not_last_frag = iph->frag_off & htons(IP_MF);
 653
 654         /*
 655          *      Keep copying data until we run out.
 656          */
 657
 658         while (left > 0) {
 659                 len = left;
 660                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 661                 if (len > mtu)
 662                         len = mtu;
 663                 /* IF: we are not sending up to and including the packet end
 664                    then align the next start on an eight byte boundary */
 665                 if (len < left) {
 666                         len &= ~7;
 667                 }
 668                 /*
 669                  *      Allocate buffer.
 670                  */
 671
 672                 if ((skb2 = alloc_skb(len+hlen+ll_rs, GFP_ATOMIC)) == NULL) {
 673                         NETDEBUG(KERN_INFO "IP: frag: no memory for new fragment!\n");
 674                         err = -ENOMEM;
 675                         goto fail;
 676                 }
 677
 678                 /*
 679                  *      Set up data on packet
 680                  */
 681
 682                 ip_copy_metadata(skb2, skb);
 683                 skb_reserve(skb2, ll_rs);
 684                 skb_put(skb2, len + hlen);
 685                 skb_reset_network_header(skb2);
 686                 skb2->transport_header = skb2->network_header + hlen;
 687
 688                 /*
 689                  *      Charge the memory for the fragment to any owner
 690                  *      it might possess
 691                  */
 692
 693                 if (skb->sk)
 694                         skb_set_owner_w(skb2, skb->sk);
 695
 696                 /*
 697                  *      Copy the packet header into the new buffer.
 698                  */
 699
 700                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 701
 702                 /*
 703                  *      Copy a block of the IP datagram.
 704                  */
 705                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 706                         BUG();
 707                 left -= len;
 708
 709                 /*
 710                  *      Fill in the new header fields.
 711                  */
 712                 iph = ip_hdr(skb2);
 713                 iph->frag_off = htons((offset >> 3));
 714
 715                 /* ANK: dirty, but effective trick. Upgrade options only if
 716                  * the segment to be fragmented was THE FIRST (otherwise,
 717                  * options are already fixed) and make it ONCE
 718                  * on the initial skb, so that all the following fragments
 719                  * will inherit fixed options.
 720                  */
 721                 if (offset == 0)
 722                         ip_options_fragment(skb);
 723
 724                 /*
 725                  *      Added AC : If we are fragmenting a fragment that's not the
 726                  *                 last fragment then keep MF on each bit
 727                  */
 728                 if (left > 0 || not_last_frag)
 729                         iph->frag_off |= htons(IP_MF);
 730                 ptr += len;
 731                 offset += len;
 732
 733                 /*
 734                  *      Put this fragment into the sending queue.
 735                  */
 736                 iph->tot_len = htons(len + hlen);
 737
 738                 ip_send_check(iph);
 739
 740                 err = output(skb2);
 741                 if (err)
 742                         goto fail;
 743
 744                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 745         }
 746         consume_skb(skb);
 747         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 748         return err;
 749
 750 fail:
 751         kfree_skb(skb);
 752         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 753         return err;
 754 }
 755 EXPORT_SYMBOL(ip_fragment);
 756
 757 int
 758 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 759 {
 760         struct iovec *iov = from;
 761
 762         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 763                 if (memcpy_fromiovecend(to, iov, offset, len) < 0)
 764                         return -EFAULT;
 765         } else {
 766                 __wsum csum = 0;
 767                 if (csum_partial_copy_fromiovecend(to, iov, offset, len, &csum) < 0)
 768                         return -EFAULT;
 769                 skb->csum = csum_block_add(skb->csum, csum, odd);
 770         }
 771         return 0;
 772 }
 773 EXPORT_SYMBOL(ip_generic_getfrag);
 774
 775 static inline __wsum
 776 csum_page(struct page *page, int offset, int copy)
 777 {
 778         char *kaddr;
 779         __wsum csum;
 780         kaddr = kmap(page);
 781         csum = csum_partial(kaddr + offset, copy, 0);
 782         kunmap(page);
 783         return csum;
 784 }
 785
 786 static inline int ip_ufo_append_data(struct sock *sk,
 787                         struct sk_buff_head *queue,
 788                         int getfrag(void *from, char *to, int offset, int len,
 789                                int odd, struct sk_buff *skb),
 790                         void *from, int length, int hh_len, int fragheaderlen,
 791                         int transhdrlen, int maxfraglen, unsigned int flags)
 792 {
 793         struct sk_buff *skb;
 794         int err;
 795
 796         /* There is support for UDP fragmentation offload by network
 797          * device, so create one single skb packet containing complete
 798          * udp datagram
 799          */
 800         if ((skb = skb_peek_tail(queue)) == NULL) {
 801                 skb = sock_alloc_send_skb(sk,
 802                         hh_len + fragheaderlen + transhdrlen + 20,
 803                         (flags & MSG_DONTWAIT), &err);
 804
 805                 if (skb == NULL)
 806                         return err;
 807
 808                 /* reserve space for Hardware header */
 809                 skb_reserve(skb, hh_len);
 810
 811                 /* create space for UDP/IP header */
 812                 skb_put(skb, fragheaderlen + transhdrlen);
 813
 814                 /* initialize network header pointer */
 815                 skb_reset_network_header(skb);
 816
 817                 /* initialize protocol header pointer */
 818                 skb->transport_header = skb->network_header + fragheaderlen;
 819
 820                 skb->csum = 0;
 821
 822
 823                 __skb_queue_tail(queue, skb);
 824         } else if (skb_is_gso(skb)) {
 825                 goto append;
 826         }
 827
 828         skb->ip_summed = CHECKSUM_PARTIAL;
 829         /* specify the length of each IP datagram fragment */
 830         skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 831         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 832
 833 append:
 834         return skb_append_datato_frags(sk, skb, getfrag, from,
 835                                        (length - transhdrlen));
 836 }
 837
 838 static int __ip_append_data(struct sock *sk,
 839                             struct flowi4 *fl4,
 840                             struct sk_buff_head *queue,
 841                             struct inet_cork *cork,
 842                             struct page_frag *pfrag,
 843                             int getfrag(void *from, char *to, int offset,
 844                                         int len, int odd, struct sk_buff *skb),
 845                             void *from, int length, int transhdrlen,
 846                             unsigned int flags)
 847 {
 848         struct inet_sock *inet = inet_sk(sk);
 849         struct sk_buff *skb;
 850
 851         struct ip_options *opt = cork->opt;
 852         int hh_len;
 853         int exthdrlen;
 854         int mtu;
 855         int copy;
 856         int err;
 857         int offset = 0;
 858         unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
 859         int csummode = CHECKSUM_NONE;
 860         struct rtable *rt = (struct rtable *)cork->dst;
 861
 862         skb = skb_peek_tail(queue);
 863
 864         exthdrlen = !skb ? rt->dst.header_len : 0;
 865         mtu = cork->fragsize;
 866
 867         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 868
 869         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 870         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 871         maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
 872
 873         if (cork->length + length > maxnonfragsize - fragheaderlen) {
 874                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 875                                mtu - (opt ? opt->optlen : 0));
 876                 return -EMSGSIZE;
 877         }
 878
 879         /*
 880          * transhdrlen > 0 means that this is the first fragment and we wish
 881          * it won't be fragmented in the future.
 882          */
 883         if (transhdrlen &&
 884             length + fragheaderlen <= mtu &&
 885             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 886             !exthdrlen)
 887                 csummode = CHECKSUM_PARTIAL;
 888
 889         cork->length += length;
 890         if ((skb && skb_is_gso(skb)) ||
 891             (((length + (skb ? skb->len : fragheaderlen)) > mtu) &&
 892             (skb_queue_len(queue) <= 1) &&
 893             (sk->sk_protocol == IPPROTO_UDP) &&
 894             (rt->dst.dev->features & NETIF_F_UFO) && !dst_xfrm(&rt->dst) &&
 895             (sk->sk_type == SOCK_DGRAM))) {
 896                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 897                                          hh_len, fragheaderlen, transhdrlen,
 898                                          maxfraglen, flags);
 899                 if (err)
 900                         goto error;
 901                 return 0;
 902         }
 903
 904         /* So, what's going on in the loop below?
 905          *
 906          * We use calculated fragment length to generate chained skb,
 907          * each of segments is IP fragment ready for sending to network after
 908          * adding appropriate IP header.
 909          */
 910
 911         if (!skb)
 912                 goto alloc_new_skb;
 913
 914         while (length > 0) {
 915                 /* Check if the remaining data fits into current packet. */
 916                 copy = mtu - skb->len;
 917                 if (copy < length)
 918                         copy = maxfraglen - skb->len;
 919                 if (copy <= 0) {
 920                         char *data;
 921                         unsigned int datalen;
 922                         unsigned int fraglen;
 923                         unsigned int fraggap;
 924                         unsigned int alloclen;
 925                         struct sk_buff *skb_prev;
 926 alloc_new_skb:
 927                         skb_prev = skb;
 928                         if (skb_prev)
 929                                 fraggap = skb_prev->len - maxfraglen;
 930                         else
 931                                 fraggap = 0;
 932
 933                         /*
 934                          * If remaining data exceeds the mtu,
 935                          * we know we need more fragment(s).
 936                          */
 937                         datalen = length + fraggap;
 938                         if (datalen > mtu - fragheaderlen)
 939                                 datalen = maxfraglen - fragheaderlen;
 940                         fraglen = datalen + fragheaderlen;
 941
 942                         if ((flags & MSG_MORE) &&
 943                             !(rt->dst.dev->features&NETIF_F_SG))
 944                                 alloclen = mtu;
 945                         else
 946                                 alloclen = fraglen;
 947
 948                         alloclen += exthdrlen;
 949
 950                         /* The last fragment gets additional space at tail.
 951                          * Note, with MSG_MORE we overallocate on fragments,
 952                          * because we have no idea what fragment will be
 953                          * the last.
 954                          */
 955                         if (datalen == length + fraggap)
 956                                 alloclen += rt->dst.trailer_len;
 957
 958                         if (transhdrlen) {
 959                                 skb = sock_alloc_send_skb(sk,
 960                                                 alloclen + hh_len + 15,
 961                                                 (flags & MSG_DONTWAIT), &err);
 962                         } else {
 963                                 skb = NULL;
 964                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 965                                     2 * sk->sk_sndbuf)
 966                                         skb = sock_wmalloc(sk,
 967                                                            alloclen + hh_len + 15, 1,
 968                                                            sk->sk_allocation);
 969                                 if (unlikely(skb == NULL))
 970                                         err = -ENOBUFS;
 971                                 else
 972                                         /* only the initial fragment is
 973                                            time stamped */
 974                                         cork->tx_flags = 0;
 975                         }
 976                         if (skb == NULL)
 977                                 goto error;
 978
 979                         /*
 980                          *      Fill in the control structures
 981                          */
 982                         skb->ip_summed = csummode;
 983                         skb->csum = 0;
 984                         skb_reserve(skb, hh_len);
 985                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 986
 987                         /*
 988                          *      Find where to start putting bytes.
 989                          */
 990                         data = skb_put(skb, fraglen + exthdrlen);
 991                         skb_set_network_header(skb, exthdrlen);
 992                         skb->transport_header = (skb->network_header +
 993                                                  fragheaderlen);
 994                         data += fragheaderlen + exthdrlen;
 995
 996                         if (fraggap) {
 997                                 skb->csum = skb_copy_and_csum_bits(
 998                                         skb_prev, maxfraglen,
 999                                         data + transhdrlen, fraggap, 0);
1000                                 skb_prev->csum = csum_sub(skb_prev->csum,
1001                                                           skb->csum);
1002                                 data += fraggap;
1003                                 pskb_trim_unique(skb_prev, maxfraglen);
1004                         }
1005
1006                         copy = datalen - transhdrlen - fraggap;
1007                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1008                                 err = -EFAULT;
1009                                 kfree_skb(skb);
1010                                 goto error;
1011                         }
1012
1013                         offset += copy;
1014                         length -= datalen - fraggap;
1015                         transhdrlen = 0;
1016                         exthdrlen = 0;
1017                         csummode = CHECKSUM_NONE;
1018
1019                         /*
1020                          * Put the packet on the pending queue.
1021                          */
1022                         __skb_queue_tail(queue, skb);
1023                         continue;
1024                 }
1025
1026                 if (copy > length)
1027                         copy = length;
1028
1029                 if (!(rt->dst.dev->features&NETIF_F_SG) &&
1030                     skb_tailroom(skb) >= copy) {
1031                         unsigned int off;
1032
1033                         off = skb->len;
1034                         if (getfrag(from, skb_put(skb, copy),
1035                                         offset, copy, off, skb) < 0) {
1036                                 __skb_trim(skb, off);
1037                                 err = -EFAULT;
1038                                 goto error;
1039                         }
1040                 } else {
1041                         int i = skb_shinfo(skb)->nr_frags;
1042
1043                         err = -ENOMEM;
1044                         if (!sk_page_frag_refill(sk, pfrag))
1045                                 goto error;
1046
1047                         if (!skb_can_coalesce(skb, i, pfrag->page,
1048                                               pfrag->offset)) {
1049                                 err = -EMSGSIZE;
1050                                 if (i == MAX_SKB_FRAGS)
1051                                         goto error;
1052
1053                                 __skb_fill_page_desc(skb, i, pfrag->page,
1054                                                      pfrag->offset, 0);
1055                                 skb_shinfo(skb)->nr_frags = ++i;
1056                                 get_page(pfrag->page);
1057                         }
1058                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1059                         if (getfrag(from,
1060                                     page_address(pfrag->page) + pfrag->offset,
1061                                     offset, copy, skb->len, skb) < 0)
1062                                 goto error_efault;
1063
1064                         pfrag->offset += copy;
1065                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1066                         skb->len += copy;
1067                         skb->data_len += copy;
1068                         skb->truesize += copy;
1069                         atomic_add(copy, &sk->sk_wmem_alloc);
1070                 }
1071                 offset += copy;
1072                 length -= copy;
1073         }
1074
1075         return 0;
1076
1077 error_efault:
1078         err = -EFAULT;
1079 error:
1080         cork->length -= length;
1081         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1082         return err;
1083 }
1084
1085 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1086                          struct ipcm_cookie *ipc, struct rtable **rtp)
1087 {
1088         struct ip_options_rcu *opt;
1089         struct rtable *rt;
1090
1091         /*
1092          * setup for corking.
1093          */
1094         opt = ipc->opt;
1095         if (opt) {
1096                 if (cork->opt == NULL) {
1097                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1098                                             sk->sk_allocation);
1099                         if (unlikely(cork->opt == NULL))
1100                                 return -ENOBUFS;
1101                 }
1102                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1103                 cork->flags |= IPCORK_OPT;
1104                 cork->addr = ipc->addr;
1105         }
1106         rt = *rtp;
1107         if (unlikely(!rt))
1108                 return -EFAULT;
1109
1110         cork->fragsize = ip_sk_use_pmtu(sk) ?
1111                          dst_mtu(&rt->dst) : ACCESS_ONCE(rt->dst.dev->mtu);
1112
1113         if (!inetdev_valid_mtu(cork->fragsize))
1114                 return -ENETUNREACH;
1115
1116         cork->dst = &rt->dst;
1117         /* We stole this route, caller should not release it. */
1118         *rtp = NULL;
1119
1120         cork->length = 0;
1121         cork->ttl = ipc->ttl;
1122         cork->tos = ipc->tos;
1123         cork->priority = ipc->priority;
1124         cork->tx_flags = ipc->tx_flags;
1125
1126         return 0;
1127 }
1128
1129 /*
1130  *      ip_append_data() and ip_append_page() can make one large IP datagram
1131  *      from many pieces of data. Each pieces will be holded on the socket
1132  *      until ip_push_pending_frames() is called. Each piece can be a page
1133  *      or non-page data.
1134  *
1135  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1136  *      this interface potentially.
1137  *
1138  *      LATER: length must be adjusted by pad at tail, when it is required.
1139  */
1140 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1141                    int getfrag(void *from, char *to, int offset, int len,
1142                                int odd, struct sk_buff *skb),
1143                    void *from, int length, int transhdrlen,
1144                    struct ipcm_cookie *ipc, struct rtable **rtp,
1145                    unsigned int flags)
1146 {
1147         struct inet_sock *inet = inet_sk(sk);
1148         int err;
1149
1150         if (flags&MSG_PROBE)
1151                 return 0;
1152
1153         if (skb_queue_empty(&sk->sk_write_queue)) {
1154                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1155                 if (err)
1156                         return err;
1157         } else {
1158                 transhdrlen = 0;
1159         }
1160
1161         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1162                                 sk_page_frag(sk), getfrag,
1163                                 from, length, transhdrlen, flags);
1164 }
1165
1166 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1167                        int offset, size_t size, int flags)
1168 {
1169         struct inet_sock *inet = inet_sk(sk);
1170         struct sk_buff *skb;
1171         struct rtable *rt;
1172         struct ip_options *opt = NULL;
1173         struct inet_cork *cork;
1174         int hh_len;
1175         int mtu;
1176         int len;
1177         int err;
1178         unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1179
1180         if (inet->hdrincl)
1181                 return -EPERM;
1182
1183         if (flags&MSG_PROBE)
1184                 return 0;
1185
1186         if (skb_queue_empty(&sk->sk_write_queue))
1187                 return -EINVAL;
1188
1189         cork = &inet->cork.base;
1190         rt = (struct rtable *)cork->dst;
1191         if (cork->flags & IPCORK_OPT)
1192                 opt = cork->opt;
1193
1194         if (!(rt->dst.dev->features&NETIF_F_SG))
1195                 return -EOPNOTSUPP;
1196
1197         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1198         mtu = cork->fragsize;
1199
1200         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1201         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1202         maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1203
1204         if (cork->length + size > maxnonfragsize - fragheaderlen) {
1205                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1206                                mtu - (opt ? opt->optlen : 0));
1207                 return -EMSGSIZE;
1208         }
1209
1210         if ((skb = skb_peek_tail(&sk->sk_write_queue)) == NULL)
1211                 return -EINVAL;
1212
1213         cork->length += size;
1214         if ((size + skb->len > mtu) &&
1215             (skb_queue_len(&sk->sk_write_queue) == 1) &&
1216             (sk->sk_protocol == IPPROTO_UDP) &&
1217             (rt->dst.dev->features & NETIF_F_UFO)) {
1218                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1219                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1220         }
1221
1222
1223         while (size > 0) {
1224                 int i;
1225
1226                 if (skb_is_gso(skb))
1227                         len = size;
1228                 else {
1229
1230                         /* Check if the remaining data fits into current packet. */
1231                         len = mtu - skb->len;
1232                         if (len < size)
1233                                 len = maxfraglen - skb->len;
1234                 }
1235                 if (len <= 0) {
1236                         struct sk_buff *skb_prev;
1237                         int alloclen;
1238
1239                         skb_prev = skb;
1240                         fraggap = skb_prev->len - maxfraglen;
1241
1242                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1243                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1244                         if (unlikely(!skb)) {
1245                                 err = -ENOBUFS;
1246                                 goto error;
1247                         }
1248
1249                         /*
1250                          *      Fill in the control structures
1251                          */
1252                         skb->ip_summed = CHECKSUM_NONE;
1253                         skb->csum = 0;
1254                         skb_reserve(skb, hh_len);
1255
1256                         /*
1257                          *      Find where to start putting bytes.
1258                          */
1259                         skb_put(skb, fragheaderlen + fraggap);
1260                         skb_reset_network_header(skb);
1261                         skb->transport_header = (skb->network_header +
1262                                                  fragheaderlen);
1263                         if (fraggap) {
1264                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1265                                                                    maxfraglen,
1266                                                     skb_transport_header(skb),
1267                                                                    fraggap, 0);
1268                                 skb_prev->csum = csum_sub(skb_prev->csum,
1269                                                           skb->csum);
1270                                 pskb_trim_unique(skb_prev, maxfraglen);
1271                         }
1272
1273                         /*
1274                          * Put the packet on the pending queue.
1275                          */
1276                         __skb_queue_tail(&sk->sk_write_queue, skb);
1277                         continue;
1278                 }
1279
1280                 i = skb_shinfo(skb)->nr_frags;
1281                 if (len > size)
1282                         len = size;
1283                 if (skb_can_coalesce(skb, i, page, offset)) {
1284                         skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1285                 } else if (i < MAX_SKB_FRAGS) {
1286                         get_page(page);
1287                         skb_fill_page_desc(skb, i, page, offset, len);
1288                 } else {
1289                         err = -EMSGSIZE;
1290                         goto error;
1291                 }
1292
1293                 if (skb->ip_summed == CHECKSUM_NONE) {
1294                         __wsum csum;
1295                         csum = csum_page(page, offset, len);
1296                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1297                 }
1298
1299                 skb->len += len;
1300                 skb->data_len += len;
1301                 skb->truesize += len;
1302                 atomic_add(len, &sk->sk_wmem_alloc);
1303                 offset += len;
1304                 size -= len;
1305         }
1306         return 0;
1307
1308 error:
1309         cork->length -= size;
1310         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1311         return err;
1312 }
1313
1314 static void ip_cork_release(struct inet_cork *cork)
1315 {
1316         cork->flags &= ~IPCORK_OPT;
1317         kfree(cork->opt);
1318         cork->opt = NULL;
1319         dst_release(cork->dst);
1320         cork->dst = NULL;
1321 }
1322
1323 /*
1324  *      Combined all pending IP fragments on the socket as one IP datagram
1325  *      and push them out.
1326  */
1327 struct sk_buff *__ip_make_skb(struct sock *sk,
1328                               struct flowi4 *fl4,
1329                               struct sk_buff_head *queue,
1330                               struct inet_cork *cork)
1331 {
1332         struct sk_buff *skb, *tmp_skb;
1333         struct sk_buff **tail_skb;
1334         struct inet_sock *inet = inet_sk(sk);
1335         struct net *net = sock_net(sk);
1336         struct ip_options *opt = NULL;
1337         struct rtable *rt = (struct rtable *)cork->dst;
1338         struct iphdr *iph;
1339         __be16 df = 0;
1340         __u8 ttl;
1341
1342         if ((skb = __skb_dequeue(queue)) == NULL)
1343                 goto out;
1344         tail_skb = &(skb_shinfo(skb)->frag_list);
1345
1346         /* move skb->data to ip header from ext header */
1347         if (skb->data < skb_network_header(skb))
1348                 __skb_pull(skb, skb_network_offset(skb));
1349         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1350                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1351                 *tail_skb = tmp_skb;
1352                 tail_skb = &(tmp_skb->next);
1353                 skb->len += tmp_skb->len;
1354                 skb->data_len += tmp_skb->len;
1355                 skb->truesize += tmp_skb->truesize;
1356                 tmp_skb->destructor = NULL;
1357                 tmp_skb->sk = NULL;
1358         }
1359
1360         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1361          * to fragment the frame generated here. No matter, what transforms
1362          * how transforms change size of the packet, it will come out.
1363          */
1364         skb->ignore_df = ip_sk_ignore_df(sk);
1365
1366         /* DF bit is set when we want to see DF on outgoing frames.
1367          * If ignore_df is set too, we still allow to fragment this frame
1368          * locally. */
1369         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1370             inet->pmtudisc == IP_PMTUDISC_PROBE ||
1371             (skb->len <= dst_mtu(&rt->dst) &&
1372              ip_dont_fragment(sk, &rt->dst)))
1373                 df = htons(IP_DF);
1374
1375         if (cork->flags & IPCORK_OPT)
1376                 opt = cork->opt;
1377
1378         if (cork->ttl != 0)
1379                 ttl = cork->ttl;
1380         else if (rt->rt_type == RTN_MULTICAST)
1381                 ttl = inet->mc_ttl;
1382         else
1383                 ttl = ip_select_ttl(inet, &rt->dst);
1384
1385         iph = ip_hdr(skb);
1386         iph->version = 4;
1387         iph->ihl = 5;
1388         iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1389         iph->frag_off = df;
1390         iph->ttl = ttl;
1391         iph->protocol = sk->sk_protocol;
1392         ip_copy_addrs(iph, fl4);
1393         ip_select_ident(net, skb, sk);
1394
1395         if (opt) {
1396                 iph->ihl += opt->optlen>>2;
1397                 ip_options_build(skb, opt, cork->addr, rt, 0);
1398         }
1399
1400         skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1401         skb->mark = sk->sk_mark;
1402         /*
1403          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1404          * on dst refcount
1405          */
1406         cork->dst = NULL;
1407         skb_dst_set(skb, &rt->dst);
1408
1409         if (iph->protocol == IPPROTO_ICMP)
1410                 icmp_out_count(net, ((struct icmphdr *)
1411                         skb_transport_header(skb))->type);
1412
1413         ip_cork_release(cork);
1414 out:
1415         return skb;
1416 }
1417
1418 int ip_send_skb(struct net *net, struct sk_buff *skb)
1419 {
1420         int err;
1421
1422         err = ip_local_out(skb);
1423         if (err) {
1424                 if (err > 0)
1425                         err = net_xmit_errno(err);
1426                 if (err)
1427                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1428         }
1429
1430         return err;
1431 }
1432
1433 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1434 {
1435         struct sk_buff *skb;
1436
1437         skb = ip_finish_skb(sk, fl4);
1438         if (!skb)
1439                 return 0;
1440
1441         /* Netfilter gets whole the not fragmented skb. */
1442         return ip_send_skb(sock_net(sk), skb);
1443 }
1444
1445 /*
1446  *      Throw away all pending data on the socket.
1447  */
1448 static void __ip_flush_pending_frames(struct sock *sk,
1449                                       struct sk_buff_head *queue,
1450                                       struct inet_cork *cork)
1451 {
1452         struct sk_buff *skb;
1453
1454         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1455                 kfree_skb(skb);
1456
1457         ip_cork_release(cork);
1458 }
1459
1460 void ip_flush_pending_frames(struct sock *sk)
1461 {
1462         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1463 }
1464
1465 struct sk_buff *ip_make_skb(struct sock *sk,
1466                             struct flowi4 *fl4,
1467                             int getfrag(void *from, char *to, int offset,
1468                                         int len, int odd, struct sk_buff *skb),
1469                             void *from, int length, int transhdrlen,
1470                             struct ipcm_cookie *ipc, struct rtable **rtp,
1471                             unsigned int flags)
1472 {
1473         struct inet_cork cork;
1474         struct sk_buff_head queue;
1475         int err;
1476
1477         if (flags & MSG_PROBE)
1478                 return NULL;
1479
1480         __skb_queue_head_init(&queue);
1481
1482         cork.flags = 0;
1483         cork.addr = 0;
1484         cork.opt = NULL;
1485         err = ip_setup_cork(sk, &cork, ipc, rtp);
1486         if (err)
1487                 return ERR_PTR(err);
1488
1489         err = __ip_append_data(sk, fl4, &queue, &cork,
1490                                &current->task_frag, getfrag,
1491                                from, length, transhdrlen, flags);
1492         if (err) {
1493                 __ip_flush_pending_frames(sk, &queue, &cork);
1494                 return ERR_PTR(err);
1495         }
1496
1497         return __ip_make_skb(sk, fl4, &queue, &cork);
1498 }
1499
1500 /*
1501  *      Fetch data from kernel space and fill in checksum if needed.
1502  */
1503 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1504                               int len, int odd, struct sk_buff *skb)
1505 {
1506         __wsum csum;
1507
1508         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1509         skb->csum = csum_block_add(skb->csum, csum, odd);
1510         return 0;
1511 }
1512
1513 /*
1514  *      Generic function to send a packet as reply to another packet.
1515  *      Used to send some TCP resets/acks so far.
1516  */
1517 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb, __be32 daddr,
1518                            __be32 saddr, const struct ip_reply_arg *arg,
1519                            unsigned int len)
1520 {
1521         struct ip_options_data replyopts;
1522         struct ipcm_cookie ipc;
1523         struct flowi4 fl4;
1524         struct rtable *rt = skb_rtable(skb);
1525         struct net *net = sock_net(sk);
1526         struct sk_buff *nskb;
1527         int err;
1528
1529         if (ip_options_echo(&replyopts.opt.opt, skb))
1530                 return;
1531
1532         ipc.addr = daddr;
1533         ipc.opt = NULL;
1534         ipc.tx_flags = 0;
1535         ipc.ttl = 0;
1536         ipc.tos = -1;
1537
1538         if (replyopts.opt.opt.optlen) {
1539                 ipc.opt = &replyopts.opt;
1540
1541                 if (replyopts.opt.opt.srr)
1542                         daddr = replyopts.opt.opt.faddr;
1543         }
1544
1545         flowi4_init_output(&fl4, arg->bound_dev_if,
1546                            IP4_REPLY_MARK(net, skb->mark),
1547                            RT_TOS(arg->tos),
1548                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1549                            ip_reply_arg_flowi_flags(arg),
1550                            daddr, saddr,
1551                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1552         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1553         rt = ip_route_output_key(net, &fl4);
1554         if (IS_ERR(rt))
1555                 return;
1556
1557         inet_sk(sk)->tos = arg->tos;
1558
1559         sk->sk_priority = skb->priority;
1560         sk->sk_protocol = ip_hdr(skb)->protocol;
1561         sk->sk_bound_dev_if = arg->bound_dev_if;
1562         sk->sk_sndbuf = sysctl_wmem_default;
1563         err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1564                              len, 0, &ipc, &rt, MSG_DONTWAIT);
1565         if (unlikely(err)) {
1566                 ip_flush_pending_frames(sk);
1567                 goto out;
1568         }
1569
1570         nskb = skb_peek(&sk->sk_write_queue);
1571         if (nskb) {
1572                 if (arg->csumoffset >= 0)
1573                         *((__sum16 *)skb_transport_header(nskb) +
1574                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1575                                                                 arg->csum));
1576                 nskb->ip_summed = CHECKSUM_NONE;
1577                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1578                 ip_push_pending_frames(sk, &fl4);
1579         }
1580 out:
1581         ip_rt_put(rt);
1582 }
1583
1584 void __init ip_init(void)
1585 {
1586         ip_rt_init();
1587         inet_initpeers();
1588
1589 #if defined(CONFIG_IP_MULTICAST)
1590         igmp_mc_init();
1591 #endif
1592 }