kernel/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_sk(struct sock *sk, struct sk_buff *skb)
  95 {
  96         struct iphdr *iph = ip_hdr(skb);
  97
  98         iph->tot_len = htons(skb->len);
  99         ip_send_check(iph);
 100         return nf_hook(NFPROTO_IPV4, NF_INET_LOCAL_OUT, sk, skb, NULL,
 101                        skb_dst(skb)->dev, dst_output_sk);
 102 }
 103
 104 int __ip_local_out(struct sk_buff *skb)
 105 {
 106         return __ip_local_out_sk(skb->sk, skb);
 107 }
 108
 109 int ip_local_out_sk(struct sock *sk, struct sk_buff *skb)
 110 {
 111         int err;
 112
 113         err = __ip_local_out(skb);
 114         if (likely(err == 1))
 115                 err = dst_output_sk(sk, skb);
 116
 117         return err;
 118 }
 119 EXPORT_SYMBOL_GPL(ip_local_out_sk);
 120
 121 static inline int ip_select_ttl(struct inet_sock *inet, struct dst_entry *dst)
 122 {
 123         int ttl = inet->uc_ttl;
 124
 125         if (ttl < 0)
 126                 ttl = ip4_dst_hoplimit(dst);
 127         return ttl;
 128 }
 129
 130 /*
 131  *              Add an ip header to a skbuff and send it out.
 132  *
 133  */
 134 int ip_build_and_send_pkt(struct sk_buff *skb, struct sock *sk,
 135                           __be32 saddr, __be32 daddr, struct ip_options_rcu *opt)
 136 {
 137         struct inet_sock *inet = inet_sk(sk);
 138         struct rtable *rt = skb_rtable(skb);
 139         struct iphdr *iph;
 140
 141         /* Build the IP header. */
 142         skb_push(skb, sizeof(struct iphdr) + (opt ? opt->opt.optlen : 0));
 143         skb_reset_network_header(skb);
 144         iph = ip_hdr(skb);
 145         iph->version  = 4;
 146         iph->ihl      = 5;
 147         iph->tos      = inet->tos;
 148         if (ip_dont_fragment(sk, &rt->dst))
 149                 iph->frag_off = htons(IP_DF);
 150         else
 151                 iph->frag_off = 0;
 152         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 153         iph->daddr    = (opt && opt->opt.srr ? opt->opt.faddr : daddr);
 154         iph->saddr    = saddr;
 155         iph->protocol = sk->sk_protocol;
 156         ip_select_ident(sock_net(sk), skb, sk);
 157
 158         if (opt && opt->opt.optlen) {
 159                 iph->ihl += opt->opt.optlen>>2;
 160                 ip_options_build(skb, &opt->opt, daddr, rt, 0);
 161         }
 162
 163         skb->priority = sk->sk_priority;
 164         skb->mark = sk->sk_mark;
 165
 166         /* Send it out. */
 167         return ip_local_out(skb);
 168 }
 169 EXPORT_SYMBOL_GPL(ip_build_and_send_pkt);
 170
 171 static inline int ip_finish_output2(struct sock *sk, struct sk_buff *skb)
 172 {
 173         struct dst_entry *dst = skb_dst(skb);
 174         struct rtable *rt = (struct rtable *)dst;
 175         struct net_device *dev = dst->dev;
 176         unsigned int hh_len = LL_RESERVED_SPACE(dev);
 177         struct neighbour *neigh;
 178         u32 nexthop;
 179
 180         if (rt->rt_type == RTN_MULTICAST) {
 181                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTMCAST, skb->len);
 182         } else if (rt->rt_type == RTN_BROADCAST)
 183                 IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUTBCAST, skb->len);
 184
 185         /* Be paranoid, rather than too clever. */
 186         if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
 187                 struct sk_buff *skb2;
 188
 189                 skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
 190                 if (!skb2) {
 191                         kfree_skb(skb);
 192                         return -ENOMEM;
 193                 }
 194                 if (skb->sk)
 195                         skb_set_owner_w(skb2, skb->sk);
 196                 consume_skb(skb);
 197                 skb = skb2;
 198         }
 199
 200         rcu_read_lock_bh();
 201         nexthop = (__force u32) rt_nexthop(rt, ip_hdr(skb)->daddr);
 202         neigh = __ipv4_neigh_lookup_noref(dev, nexthop);
 203         if (unlikely(!neigh))
 204                 neigh = __neigh_create(&arp_tbl, &nexthop, dev, false);
 205         if (!IS_ERR(neigh)) {
 206                 int res = dst_neigh_output(dst, neigh, skb);
 207
 208                 rcu_read_unlock_bh();
 209                 return res;
 210         }
 211         rcu_read_unlock_bh();
 212
 213         net_dbg_ratelimited("%s: No header cache and no neighbour!\n",
 214                             __func__);
 215         kfree_skb(skb);
 216         return -EINVAL;
 217 }
 218
 219 static int ip_finish_output_gso(struct sock *sk, struct sk_buff *skb)
 220 {
 221         netdev_features_t features;
 222         struct sk_buff *segs;
 223         int ret = 0;
 224
 225         /* common case: locally created skb or seglen is <= mtu */
 226         if (((IPCB(skb)->flags & IPSKB_FORWARDED) == 0) ||
 227               skb_gso_network_seglen(skb) <= ip_skb_dst_mtu(skb))
 228                 return ip_finish_output2(sk, skb);
 229
 230         /* Slowpath -  GSO segment length is exceeding the dst MTU.
 231          *
 232          * This can happen in two cases:
 233          * 1) TCP GRO packet, DF bit not set
 234          * 2) skb arrived via virtio-net, we thus get TSO/GSO skbs directly
 235          * from host network stack.
 236          */
 237         features = netif_skb_features(skb);
 238         segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 239         if (IS_ERR_OR_NULL(segs)) {
 240                 kfree_skb(skb);
 241                 return -ENOMEM;
 242         }
 243
 244         consume_skb(skb);
 245
 246         do {
 247                 struct sk_buff *nskb = segs->next;
 248                 int err;
 249
 250                 segs->next = NULL;
 251                 err = ip_fragment(sk, segs, ip_finish_output2);
 252
 253                 if (err && ret == 0)
 254                         ret = err;
 255                 segs = nskb;
 256         } while (segs);
 257
 258         return ret;
 259 }
 260
 261 static int ip_finish_output(struct sock *sk, struct sk_buff *skb)
 262 {
 263 #if defined(CONFIG_NETFILTER) && defined(CONFIG_XFRM)
 264         /* Policy lookup after SNAT yielded a new policy */
 265         if (skb_dst(skb)->xfrm) {
 266                 IPCB(skb)->flags |= IPSKB_REROUTED;
 267                 return dst_output_sk(sk, skb);
 268         }
 269 #endif
 270         if (skb_is_gso(skb))
 271                 return ip_finish_output_gso(sk, skb);
 272
 273         if (skb->len > ip_skb_dst_mtu(skb))
 274                 return ip_fragment(sk, skb, ip_finish_output2);
 275
 276         return ip_finish_output2(sk, skb);
 277 }
 278
 279 int ip_mc_output(struct sock *sk, struct sk_buff *skb)
 280 {
 281         struct rtable *rt = skb_rtable(skb);
 282         struct net_device *dev = rt->dst.dev;
 283
 284         /*
 285          *      If the indicated interface is up and running, send the packet.
 286          */
 287         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 288
 289         skb->dev = dev;
 290         skb->protocol = htons(ETH_P_IP);
 291
 292         /*
 293          *      Multicasts are looped back for other local users
 294          */
 295
 296         if (rt->rt_flags&RTCF_MULTICAST) {
 297                 if (sk_mc_loop(sk)
 298 #ifdef CONFIG_IP_MROUTE
 299                 /* Small optimization: do not loopback not local frames,
 300                    which returned after forwarding; they will be  dropped
 301                    by ip_mr_input in any case.
 302                    Note, that local frames are looped back to be delivered
 303                    to local recipients.
 304
 305                    This check is duplicated in ip_mr_input at the moment.
 306                  */
 307                     &&
 308                     ((rt->rt_flags & RTCF_LOCAL) ||
 309                      !(IPCB(skb)->flags & IPSKB_FORWARDED))
 310 #endif
 311                    ) {
 312                         struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 313                         if (newskb)
 314                                 NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING,
 315                                         sk, newskb, NULL, newskb->dev,
 316                                         dev_loopback_xmit);
 317                 }
 318
 319                 /* Multicasts with ttl 0 must not go beyond the host */
 320
 321                 if (ip_hdr(skb)->ttl == 0) {
 322                         kfree_skb(skb);
 323                         return 0;
 324                 }
 325         }
 326
 327         if (rt->rt_flags&RTCF_BROADCAST) {
 328                 struct sk_buff *newskb = skb_clone(skb, GFP_ATOMIC);
 329                 if (newskb)
 330                         NF_HOOK(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, newskb,
 331                                 NULL, newskb->dev, dev_loopback_xmit);
 332         }
 333
 334         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, skb, NULL,
 335                             skb->dev, ip_finish_output,
 336                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 337 }
 338
 339 int ip_output(struct sock *sk, struct sk_buff *skb)
 340 {
 341         struct net_device *dev = skb_dst(skb)->dev;
 342
 343         IP_UPD_PO_STATS(dev_net(dev), IPSTATS_MIB_OUT, skb->len);
 344
 345         skb->dev = dev;
 346         skb->protocol = htons(ETH_P_IP);
 347
 348         return NF_HOOK_COND(NFPROTO_IPV4, NF_INET_POST_ROUTING, sk, skb,
 349                             NULL, dev,
 350                             ip_finish_output,
 351                             !(IPCB(skb)->flags & IPSKB_REROUTED));
 352 }
 353
 354 /*
 355  * copy saddr and daddr, possibly using 64bit load/stores
 356  * Equivalent to :
 357  *   iph->saddr = fl4->saddr;
 358  *   iph->daddr = fl4->daddr;
 359  */
 360 static void ip_copy_addrs(struct iphdr *iph, const struct flowi4 *fl4)
 361 {
 362         BUILD_BUG_ON(offsetof(typeof(*fl4), daddr) !=
 363                      offsetof(typeof(*fl4), saddr) + sizeof(fl4->saddr));
 364         memcpy(&iph->saddr, &fl4->saddr,
 365                sizeof(fl4->saddr) + sizeof(fl4->daddr));
 366 }
 367
 368 /* Note: skb->sk can be different from sk, in case of tunnels */
 369 int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)
 370 {
 371         struct inet_sock *inet = inet_sk(sk);
 372         struct ip_options_rcu *inet_opt;
 373         struct flowi4 *fl4;
 374         struct rtable *rt;
 375         struct iphdr *iph;
 376         int res;
 377
 378         /* Skip all of this if the packet is already routed,
 379          * f.e. by something like SCTP.
 380          */
 381         rcu_read_lock();
 382         inet_opt = rcu_dereference(inet->inet_opt);
 383         fl4 = &fl->u.ip4;
 384         rt = skb_rtable(skb);
 385         if (rt)
 386                 goto packet_routed;
 387
 388         /* Make sure we can route this packet. */
 389         rt = (struct rtable *)__sk_dst_check(sk, 0);
 390         if (!rt) {
 391                 __be32 daddr;
 392
 393                 /* Use correct destination address if we have options. */
 394                 daddr = inet->inet_daddr;
 395                 if (inet_opt && inet_opt->opt.srr)
 396                         daddr = inet_opt->opt.faddr;
 397
 398                 /* If this fails, retransmit mechanism of transport layer will
 399                  * keep trying until route appears or the connection times
 400                  * itself out.
 401                  */
 402                 rt = ip_route_output_ports(sock_net(sk), fl4, sk,
 403                                            daddr, inet->inet_saddr,
 404                                            inet->inet_dport,
 405                                            inet->inet_sport,
 406                                            sk->sk_protocol,
 407                                            RT_CONN_FLAGS(sk),
 408                                            sk->sk_bound_dev_if);
 409                 if (IS_ERR(rt))
 410                         goto no_route;
 411                 sk_setup_caps(sk, &rt->dst);
 412         }
 413         skb_dst_set_noref(skb, &rt->dst);
 414
 415 packet_routed:
 416         if (inet_opt && inet_opt->opt.is_strictroute && rt->rt_uses_gateway)
 417                 goto no_route;
 418
 419         /* OK, we know where to send it, allocate and build IP header. */
 420         skb_push(skb, sizeof(struct iphdr) + (inet_opt ? inet_opt->opt.optlen : 0));
 421         skb_reset_network_header(skb);
 422         iph = ip_hdr(skb);
 423         *((__be16 *)iph) = htons((4 << 12) | (5 << 8) | (inet->tos & 0xff));
 424         if (ip_dont_fragment(sk, &rt->dst) && !skb->ignore_df)
 425                 iph->frag_off = htons(IP_DF);
 426         else
 427                 iph->frag_off = 0;
 428         iph->ttl      = ip_select_ttl(inet, &rt->dst);
 429         iph->protocol = sk->sk_protocol;
 430         ip_copy_addrs(iph, fl4);
 431
 432         /* Transport layer set skb->h.foo itself. */
 433
 434         if (inet_opt && inet_opt->opt.optlen) {
 435                 iph->ihl += inet_opt->opt.optlen >> 2;
 436                 ip_options_build(skb, &inet_opt->opt, inet->inet_daddr, rt, 0);
 437         }
 438
 439         ip_select_ident_segs(sock_net(sk), skb, sk,
 440                              skb_shinfo(skb)->gso_segs ?: 1);
 441
 442         /* TODO : should we use skb->sk here instead of sk ? */
 443         skb->priority = sk->sk_priority;
 444         skb->mark = sk->sk_mark;
 445
 446         res = ip_local_out(skb);
 447         rcu_read_unlock();
 448         return res;
 449
 450 no_route:
 451         rcu_read_unlock();
 452         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
 453         kfree_skb(skb);
 454         return -EHOSTUNREACH;
 455 }
 456 EXPORT_SYMBOL(ip_queue_xmit);
 457
 458 static void ip_copy_metadata(struct sk_buff *to, struct sk_buff *from)
 459 {
 460         to->pkt_type = from->pkt_type;
 461         to->priority = from->priority;
 462         to->protocol = from->protocol;
 463         skb_dst_drop(to);
 464         skb_dst_copy(to, from);
 465         to->dev = from->dev;
 466         to->mark = from->mark;
 467
 468         /* Copy the flags to each fragment. */
 469         IPCB(to)->flags = IPCB(from)->flags;
 470
 471 #ifdef CONFIG_NET_SCHED
 472         to->tc_index = from->tc_index;
 473 #endif
 474         nf_copy(to, from);
 475 #if defined(CONFIG_IP_VS) || defined(CONFIG_IP_VS_MODULE)
 476         to->ipvs_property = from->ipvs_property;
 477 #endif
 478         skb_copy_secmark(to, from);
 479 }
 480
 481 /*
 482  *      This IP datagram is too large to be sent in one piece.  Break it up into
 483  *      smaller pieces (each of size equal to IP header plus
 484  *      a block of the data of the original IP data part) that will yet fit in a
 485  *      single device frame, and queue such a frame for sending.
 486  */
 487
 488 int ip_fragment(struct sock *sk, struct sk_buff *skb,
 489                 int (*output)(struct sock *, struct sk_buff *))
 490 {
 491         struct iphdr *iph;
 492         int ptr;
 493         struct net_device *dev;
 494         struct sk_buff *skb2;
 495         unsigned int mtu, hlen, left, len, ll_rs;
 496         int offset;
 497         __be16 not_last_frag;
 498         struct rtable *rt = skb_rtable(skb);
 499         int err = 0;
 500
 501         dev = rt->dst.dev;
 502
 503         /*
 504          *      Point into the IP datagram header.
 505          */
 506
 507         iph = ip_hdr(skb);
 508
 509         mtu = ip_skb_dst_mtu(skb);
 510         if (unlikely(((iph->frag_off & htons(IP_DF)) && !skb->ignore_df) ||
 511                      (IPCB(skb)->frag_max_size &&
 512                       IPCB(skb)->frag_max_size > mtu))) {
 513                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 514                 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED,
 515                           htonl(mtu));
 516                 kfree_skb(skb);
 517                 return -EMSGSIZE;
 518         }
 519
 520         /*
 521          *      Setup starting values.
 522          */
 523
 524         hlen = iph->ihl * 4;
 525         mtu = mtu - hlen;       /* Size of data space */
 526 #if IS_ENABLED(CONFIG_BRIDGE_NETFILTER)
 527         if (skb->nf_bridge)
 528                 mtu -= nf_bridge_mtu_reduction(skb);
 529 #endif
 530         IPCB(skb)->flags |= IPSKB_FRAG_COMPLETE;
 531
 532         /* When frag_list is given, use it. First, check its validity:
 533          * some transformers could create wrong frag_list or break existing
 534          * one, it is not prohibited. In this case fall back to copying.
 535          *
 536          * LATER: this step can be merged to real generation of fragments,
 537          * we can switch to copy when see the first bad fragment.
 538          */
 539         if (skb_has_frag_list(skb)) {
 540                 struct sk_buff *frag, *frag2;
 541                 int first_len = skb_pagelen(skb);
 542
 543                 if (first_len - hlen > mtu ||
 544                     ((first_len - hlen) & 7) ||
 545                     ip_is_fragment(iph) ||
 546                     skb_cloned(skb))
 547                         goto slow_path;
 548
 549                 skb_walk_frags(skb, frag) {
 550                         /* Correct geometry. */
 551                         if (frag->len > mtu ||
 552                             ((frag->len & 7) && frag->next) ||
 553                             skb_headroom(frag) < hlen)
 554                                 goto slow_path_clean;
 555
 556                         /* Partially cloned skb? */
 557                         if (skb_shared(frag))
 558                                 goto slow_path_clean;
 559
 560                         BUG_ON(frag->sk);
 561                         if (skb->sk) {
 562                                 frag->sk = skb->sk;
 563                                 frag->destructor = sock_wfree;
 564                         }
 565                         skb->truesize -= frag->truesize;
 566                 }
 567
 568                 /* Everything is OK. Generate! */
 569
 570                 err = 0;
 571                 offset = 0;
 572                 frag = skb_shinfo(skb)->frag_list;
 573                 skb_frag_list_init(skb);
 574                 skb->data_len = first_len - skb_headlen(skb);
 575                 skb->len = first_len;
 576                 iph->tot_len = htons(first_len);
 577                 iph->frag_off = htons(IP_MF);
 578                 ip_send_check(iph);
 579
 580                 for (;;) {
 581                         /* Prepare header of the next frame,
 582                          * before previous one went down. */
 583                         if (frag) {
 584                                 frag->ip_summed = CHECKSUM_NONE;
 585                                 skb_reset_transport_header(frag);
 586                                 __skb_push(frag, hlen);
 587                                 skb_reset_network_header(frag);
 588                                 memcpy(skb_network_header(frag), iph, hlen);
 589                                 iph = ip_hdr(frag);
 590                                 iph->tot_len = htons(frag->len);
 591                                 ip_copy_metadata(frag, skb);
 592                                 if (offset == 0)
 593                                         ip_options_fragment(frag);
 594                                 offset += skb->len - hlen;
 595                                 iph->frag_off = htons(offset>>3);
 596                                 if (frag->next)
 597                                         iph->frag_off |= htons(IP_MF);
 598                                 /* Ready, complete checksum */
 599                                 ip_send_check(iph);
 600                         }
 601
 602                         err = output(sk, skb);
 603
 604                         if (!err)
 605                                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 606                         if (err || !frag)
 607                                 break;
 608
 609                         skb = frag;
 610                         frag = skb->next;
 611                         skb->next = NULL;
 612                 }
 613
 614                 if (err == 0) {
 615                         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 616                         return 0;
 617                 }
 618
 619                 while (frag) {
 620                         skb = frag->next;
 621                         kfree_skb(frag);
 622                         frag = skb;
 623                 }
 624                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 625                 return err;
 626
 627 slow_path_clean:
 628                 skb_walk_frags(skb, frag2) {
 629                         if (frag2 == frag)
 630                                 break;
 631                         frag2->sk = NULL;
 632                         frag2->destructor = NULL;
 633                         skb->truesize += frag2->truesize;
 634                 }
 635         }
 636
 637 slow_path:
 638         /* for offloaded checksums cleanup checksum before fragmentation */
 639         if ((skb->ip_summed == CHECKSUM_PARTIAL) && skb_checksum_help(skb))
 640                 goto fail;
 641         iph = ip_hdr(skb);
 642
 643         left = skb->len - hlen;         /* Space per frame */
 644         ptr = hlen;             /* Where to start from */
 645
 646         ll_rs = LL_RESERVED_SPACE(rt->dst.dev);
 647
 648         /*
 649          *      Fragment the datagram.
 650          */
 651
 652         offset = (ntohs(iph->frag_off) & IP_OFFSET) << 3;
 653         not_last_frag = iph->frag_off & htons(IP_MF);
 654
 655         /*
 656          *      Keep copying data until we run out.
 657          */
 658
 659         while (left > 0) {
 660                 len = left;
 661                 /* IF: it doesn't fit, use 'mtu' - the data space left */
 662                 if (len > mtu)
 663                         len = mtu;
 664                 /* IF: we are not sending up to and including the packet end
 665                    then align the next start on an eight byte boundary */
 666                 if (len < left) {
 667                         len &= ~7;
 668                 }
 669
 670                 /* Allocate buffer */
 671                 skb2 = alloc_skb(len + hlen + ll_rs, GFP_ATOMIC);
 672                 if (!skb2) {
 673                         err = -ENOMEM;
 674                         goto fail;
 675                 }
 676
 677                 /*
 678                  *      Set up data on packet
 679                  */
 680
 681                 ip_copy_metadata(skb2, skb);
 682                 skb_reserve(skb2, ll_rs);
 683                 skb_put(skb2, len + hlen);
 684                 skb_reset_network_header(skb2);
 685                 skb2->transport_header = skb2->network_header + hlen;
 686
 687                 /*
 688                  *      Charge the memory for the fragment to any owner
 689                  *      it might possess
 690                  */
 691
 692                 if (skb->sk)
 693                         skb_set_owner_w(skb2, skb->sk);
 694
 695                 /*
 696                  *      Copy the packet header into the new buffer.
 697                  */
 698
 699                 skb_copy_from_linear_data(skb, skb_network_header(skb2), hlen);
 700
 701                 /*
 702                  *      Copy a block of the IP datagram.
 703                  */
 704                 if (skb_copy_bits(skb, ptr, skb_transport_header(skb2), len))
 705                         BUG();
 706                 left -= len;
 707
 708                 /*
 709                  *      Fill in the new header fields.
 710                  */
 711                 iph = ip_hdr(skb2);
 712                 iph->frag_off = htons((offset >> 3));
 713
 714                 /* ANK: dirty, but effective trick. Upgrade options only if
 715                  * the segment to be fragmented was THE FIRST (otherwise,
 716                  * options are already fixed) and make it ONCE
 717                  * on the initial skb, so that all the following fragments
 718                  * will inherit fixed options.
 719                  */
 720                 if (offset == 0)
 721                         ip_options_fragment(skb);
 722
 723                 /*
 724                  *      Added AC : If we are fragmenting a fragment that's not the
 725                  *                 last fragment then keep MF on each bit
 726                  */
 727                 if (left > 0 || not_last_frag)
 728                         iph->frag_off |= htons(IP_MF);
 729                 ptr += len;
 730                 offset += len;
 731
 732                 /*
 733                  *      Put this fragment into the sending queue.
 734                  */
 735                 iph->tot_len = htons(len + hlen);
 736
 737                 ip_send_check(iph);
 738
 739                 err = output(sk, skb2);
 740                 if (err)
 741                         goto fail;
 742
 743                 IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGCREATES);
 744         }
 745         consume_skb(skb);
 746         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGOKS);
 747         return err;
 748
 749 fail:
 750         kfree_skb(skb);
 751         IP_INC_STATS(dev_net(dev), IPSTATS_MIB_FRAGFAILS);
 752         return err;
 753 }
 754 EXPORT_SYMBOL(ip_fragment);
 755
 756 int
 757 ip_generic_getfrag(void *from, char *to, int offset, int len, int odd, struct sk_buff *skb)
 758 {
 759         struct msghdr *msg = from;
 760
 761         if (skb->ip_summed == CHECKSUM_PARTIAL) {
 762                 if (copy_from_iter(to, len, &msg->msg_iter) != len)
 763                         return -EFAULT;
 764         } else {
 765                 __wsum csum = 0;
 766                 if (csum_and_copy_from_iter(to, len, &csum, &msg->msg_iter) != len)
 767                         return -EFAULT;
 768                 skb->csum = csum_block_add(skb->csum, csum, odd);
 769         }
 770         return 0;
 771 }
 772 EXPORT_SYMBOL(ip_generic_getfrag);
 773
 774 static inline __wsum
 775 csum_page(struct page *page, int offset, int copy)
 776 {
 777         char *kaddr;
 778         __wsum csum;
 779         kaddr = kmap(page);
 780         csum = csum_partial(kaddr + offset, copy, 0);
 781         kunmap(page);
 782         return csum;
 783 }
 784
 785 static inline int ip_ufo_append_data(struct sock *sk,
 786                         struct sk_buff_head *queue,
 787                         int getfrag(void *from, char *to, int offset, int len,
 788                                int odd, struct sk_buff *skb),
 789                         void *from, int length, int hh_len, int fragheaderlen,
 790                         int transhdrlen, int maxfraglen, unsigned int flags)
 791 {
 792         struct sk_buff *skb;
 793         int err;
 794
 795         /* There is support for UDP fragmentation offload by network
 796          * device, so create one single skb packet containing complete
 797          * udp datagram
 798          */
 799         skb = skb_peek_tail(queue);
 800         if (!skb) {
 801                 skb = sock_alloc_send_skb(sk,
 802                         hh_len + fragheaderlen + transhdrlen + 20,
 803                         (flags & MSG_DONTWAIT), &err);
 804
 805                 if (!skb)
 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                 __skb_queue_tail(queue, skb);
 823         } else if (skb_is_gso(skb)) {
 824                 goto append;
 825         }
 826
 827         skb->ip_summed = CHECKSUM_PARTIAL;
 828         /* specify the length of each IP datagram fragment */
 829         skb_shinfo(skb)->gso_size = maxfraglen - fragheaderlen;
 830         skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
 831
 832 append:
 833         return skb_append_datato_frags(sk, skb, getfrag, from,
 834                                        (length - transhdrlen));
 835 }
 836
 837 static int __ip_append_data(struct sock *sk,
 838                             struct flowi4 *fl4,
 839                             struct sk_buff_head *queue,
 840                             struct inet_cork *cork,
 841                             struct page_frag *pfrag,
 842                             int getfrag(void *from, char *to, int offset,
 843                                         int len, int odd, struct sk_buff *skb),
 844                             void *from, int length, int transhdrlen,
 845                             unsigned int flags)
 846 {
 847         struct inet_sock *inet = inet_sk(sk);
 848         struct sk_buff *skb;
 849
 850         struct ip_options *opt = cork->opt;
 851         int hh_len;
 852         int exthdrlen;
 853         int mtu;
 854         int copy;
 855         int err;
 856         int offset = 0;
 857         unsigned int maxfraglen, fragheaderlen, maxnonfragsize;
 858         int csummode = CHECKSUM_NONE;
 859         struct rtable *rt = (struct rtable *)cork->dst;
 860         u32 tskey = 0;
 861
 862         skb = skb_peek_tail(queue);
 863
 864         exthdrlen = !skb ? rt->dst.header_len : 0;
 865         mtu = cork->fragsize;
 866         if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
 867             sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
 868                 tskey = sk->sk_tskey++;
 869
 870         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
 871
 872         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 873         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
 874         maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
 875
 876         if (cork->length + length > maxnonfragsize - fragheaderlen) {
 877                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
 878                                mtu - (opt ? opt->optlen : 0));
 879                 return -EMSGSIZE;
 880         }
 881
 882         /*
 883          * transhdrlen > 0 means that this is the first fragment and we wish
 884          * it won't be fragmented in the future.
 885          */
 886         if (transhdrlen &&
 887             length + fragheaderlen <= mtu &&
 888             rt->dst.dev->features & NETIF_F_V4_CSUM &&
 889             !exthdrlen)
 890                 csummode = CHECKSUM_PARTIAL;
 891
 892         cork->length += length;
 893         if (((length > mtu) || (skb && skb_is_gso(skb))) &&
 894             (sk->sk_protocol == IPPROTO_UDP) &&
 895             (rt->dst.dev->features & NETIF_F_UFO) && !rt->dst.header_len &&
 896             (sk->sk_type == SOCK_DGRAM)) {
 897                 err = ip_ufo_append_data(sk, queue, getfrag, from, length,
 898                                          hh_len, fragheaderlen, transhdrlen,
 899                                          maxfraglen, flags);
 900                 if (err)
 901                         goto error;
 902                 return 0;
 903         }
 904
 905         /* So, what's going on in the loop below?
 906          *
 907          * We use calculated fragment length to generate chained skb,
 908          * each of segments is IP fragment ready for sending to network after
 909          * adding appropriate IP header.
 910          */
 911
 912         if (!skb)
 913                 goto alloc_new_skb;
 914
 915         while (length > 0) {
 916                 /* Check if the remaining data fits into current packet. */
 917                 copy = mtu - skb->len;
 918                 if (copy < length)
 919                         copy = maxfraglen - skb->len;
 920                 if (copy <= 0) {
 921                         char *data;
 922                         unsigned int datalen;
 923                         unsigned int fraglen;
 924                         unsigned int fraggap;
 925                         unsigned int alloclen;
 926                         struct sk_buff *skb_prev;
 927 alloc_new_skb:
 928                         skb_prev = skb;
 929                         if (skb_prev)
 930                                 fraggap = skb_prev->len - maxfraglen;
 931                         else
 932                                 fraggap = 0;
 933
 934                         /*
 935                          * If remaining data exceeds the mtu,
 936                          * we know we need more fragment(s).
 937                          */
 938                         datalen = length + fraggap;
 939                         if (datalen > mtu - fragheaderlen)
 940                                 datalen = maxfraglen - fragheaderlen;
 941                         fraglen = datalen + fragheaderlen;
 942
 943                         if ((flags & MSG_MORE) &&
 944                             !(rt->dst.dev->features&NETIF_F_SG))
 945                                 alloclen = mtu;
 946                         else
 947                                 alloclen = fraglen;
 948
 949                         alloclen += exthdrlen;
 950
 951                         /* The last fragment gets additional space at tail.
 952                          * Note, with MSG_MORE we overallocate on fragments,
 953                          * because we have no idea what fragment will be
 954                          * the last.
 955                          */
 956                         if (datalen == length + fraggap)
 957                                 alloclen += rt->dst.trailer_len;
 958
 959                         if (transhdrlen) {
 960                                 skb = sock_alloc_send_skb(sk,
 961                                                 alloclen + hh_len + 15,
 962                                                 (flags & MSG_DONTWAIT), &err);
 963                         } else {
 964                                 skb = NULL;
 965                                 if (atomic_read(&sk->sk_wmem_alloc) <=
 966                                     2 * sk->sk_sndbuf)
 967                                         skb = sock_wmalloc(sk,
 968                                                            alloclen + hh_len + 15, 1,
 969                                                            sk->sk_allocation);
 970                                 if (unlikely(!skb))
 971                                         err = -ENOBUFS;
 972                         }
 973                         if (!skb)
 974                                 goto error;
 975
 976                         /*
 977                          *      Fill in the control structures
 978                          */
 979                         skb->ip_summed = csummode;
 980                         skb->csum = 0;
 981                         skb_reserve(skb, hh_len);
 982
 983                         /* only the initial fragment is time stamped */
 984                         skb_shinfo(skb)->tx_flags = cork->tx_flags;
 985                         cork->tx_flags = 0;
 986                         skb_shinfo(skb)->tskey = tskey;
 987                         tskey = 0;
 988
 989                         /*
 990                          *      Find where to start putting bytes.
 991                          */
 992                         data = skb_put(skb, fraglen + exthdrlen);
 993                         skb_set_network_header(skb, exthdrlen);
 994                         skb->transport_header = (skb->network_header +
 995                                                  fragheaderlen);
 996                         data += fragheaderlen + exthdrlen;
 997
 998                         if (fraggap) {
 999                                 skb->csum = skb_copy_and_csum_bits(
1000                                         skb_prev, maxfraglen,
1001                                         data + transhdrlen, fraggap, 0);
1002                                 skb_prev->csum = csum_sub(skb_prev->csum,
1003                                                           skb->csum);
1004                                 data += fraggap;
1005                                 pskb_trim_unique(skb_prev, maxfraglen);
1006                         }
1007
1008                         copy = datalen - transhdrlen - fraggap;
1009                         if (copy > 0 && getfrag(from, data + transhdrlen, offset, copy, fraggap, skb) < 0) {
1010                                 err = -EFAULT;
1011                                 kfree_skb(skb);
1012                                 goto error;
1013                         }
1014
1015                         offset += copy;
1016                         length -= datalen - fraggap;
1017                         transhdrlen = 0;
1018                         exthdrlen = 0;
1019                         csummode = CHECKSUM_NONE;
1020
1021                         /*
1022                          * Put the packet on the pending queue.
1023                          */
1024                         __skb_queue_tail(queue, skb);
1025                         continue;
1026                 }
1027
1028                 if (copy > length)
1029                         copy = length;
1030
1031                 if (!(rt->dst.dev->features&NETIF_F_SG)) {
1032                         unsigned int off;
1033
1034                         off = skb->len;
1035                         if (getfrag(from, skb_put(skb, copy),
1036                                         offset, copy, off, skb) < 0) {
1037                                 __skb_trim(skb, off);
1038                                 err = -EFAULT;
1039                                 goto error;
1040                         }
1041                 } else {
1042                         int i = skb_shinfo(skb)->nr_frags;
1043
1044                         err = -ENOMEM;
1045                         if (!sk_page_frag_refill(sk, pfrag))
1046                                 goto error;
1047
1048                         if (!skb_can_coalesce(skb, i, pfrag->page,
1049                                               pfrag->offset)) {
1050                                 err = -EMSGSIZE;
1051                                 if (i == MAX_SKB_FRAGS)
1052                                         goto error;
1053
1054                                 __skb_fill_page_desc(skb, i, pfrag->page,
1055                                                      pfrag->offset, 0);
1056                                 skb_shinfo(skb)->nr_frags = ++i;
1057                                 get_page(pfrag->page);
1058                         }
1059                         copy = min_t(int, copy, pfrag->size - pfrag->offset);
1060                         if (getfrag(from,
1061                                     page_address(pfrag->page) + pfrag->offset,
1062                                     offset, copy, skb->len, skb) < 0)
1063                                 goto error_efault;
1064
1065                         pfrag->offset += copy;
1066                         skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1067                         skb->len += copy;
1068                         skb->data_len += copy;
1069                         skb->truesize += copy;
1070                         atomic_add(copy, &sk->sk_wmem_alloc);
1071                 }
1072                 offset += copy;
1073                 length -= copy;
1074         }
1075
1076         return 0;
1077
1078 error_efault:
1079         err = -EFAULT;
1080 error:
1081         cork->length -= length;
1082         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1083         return err;
1084 }
1085
1086 static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
1087                          struct ipcm_cookie *ipc, struct rtable **rtp)
1088 {
1089         struct ip_options_rcu *opt;
1090         struct rtable *rt;
1091
1092         /*
1093          * setup for corking.
1094          */
1095         opt = ipc->opt;
1096         if (opt) {
1097                 if (!cork->opt) {
1098                         cork->opt = kmalloc(sizeof(struct ip_options) + 40,
1099                                             sk->sk_allocation);
1100                         if (unlikely(!cork->opt))
1101                                 return -ENOBUFS;
1102                 }
1103                 memcpy(cork->opt, &opt->opt, sizeof(struct ip_options) + opt->opt.optlen);
1104                 cork->flags |= IPCORK_OPT;
1105                 cork->addr = ipc->addr;
1106         }
1107         rt = *rtp;
1108         if (unlikely(!rt))
1109                 return -EFAULT;
1110         /*
1111          * We steal reference to this route, caller should not release it
1112          */
1113         *rtp = NULL;
1114         cork->fragsize = ip_sk_use_pmtu(sk) ?
1115                          dst_mtu(&rt->dst) : rt->dst.dev->mtu;
1116         cork->dst = &rt->dst;
1117         cork->length = 0;
1118         cork->ttl = ipc->ttl;
1119         cork->tos = ipc->tos;
1120         cork->priority = ipc->priority;
1121         cork->tx_flags = ipc->tx_flags;
1122
1123         return 0;
1124 }
1125
1126 /*
1127  *      ip_append_data() and ip_append_page() can make one large IP datagram
1128  *      from many pieces of data. Each pieces will be holded on the socket
1129  *      until ip_push_pending_frames() is called. Each piece can be a page
1130  *      or non-page data.
1131  *
1132  *      Not only UDP, other transport protocols - e.g. raw sockets - can use
1133  *      this interface potentially.
1134  *
1135  *      LATER: length must be adjusted by pad at tail, when it is required.
1136  */
1137 int ip_append_data(struct sock *sk, struct flowi4 *fl4,
1138                    int getfrag(void *from, char *to, int offset, int len,
1139                                int odd, struct sk_buff *skb),
1140                    void *from, int length, int transhdrlen,
1141                    struct ipcm_cookie *ipc, struct rtable **rtp,
1142                    unsigned int flags)
1143 {
1144         struct inet_sock *inet = inet_sk(sk);
1145         int err;
1146
1147         if (flags&MSG_PROBE)
1148                 return 0;
1149
1150         if (skb_queue_empty(&sk->sk_write_queue)) {
1151                 err = ip_setup_cork(sk, &inet->cork.base, ipc, rtp);
1152                 if (err)
1153                         return err;
1154         } else {
1155                 transhdrlen = 0;
1156         }
1157
1158         return __ip_append_data(sk, fl4, &sk->sk_write_queue, &inet->cork.base,
1159                                 sk_page_frag(sk), getfrag,
1160                                 from, length, transhdrlen, flags);
1161 }
1162
1163 ssize_t ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
1164                        int offset, size_t size, int flags)
1165 {
1166         struct inet_sock *inet = inet_sk(sk);
1167         struct sk_buff *skb;
1168         struct rtable *rt;
1169         struct ip_options *opt = NULL;
1170         struct inet_cork *cork;
1171         int hh_len;
1172         int mtu;
1173         int len;
1174         int err;
1175         unsigned int maxfraglen, fragheaderlen, fraggap, maxnonfragsize;
1176
1177         if (inet->hdrincl)
1178                 return -EPERM;
1179
1180         if (flags&MSG_PROBE)
1181                 return 0;
1182
1183         if (skb_queue_empty(&sk->sk_write_queue))
1184                 return -EINVAL;
1185
1186         cork = &inet->cork.base;
1187         rt = (struct rtable *)cork->dst;
1188         if (cork->flags & IPCORK_OPT)
1189                 opt = cork->opt;
1190
1191         if (!(rt->dst.dev->features&NETIF_F_SG))
1192                 return -EOPNOTSUPP;
1193
1194         hh_len = LL_RESERVED_SPACE(rt->dst.dev);
1195         mtu = cork->fragsize;
1196
1197         fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
1198         maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;
1199         maxnonfragsize = ip_sk_ignore_df(sk) ? 0xFFFF : mtu;
1200
1201         if (cork->length + size > maxnonfragsize - fragheaderlen) {
1202                 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport,
1203                                mtu - (opt ? opt->optlen : 0));
1204                 return -EMSGSIZE;
1205         }
1206
1207         skb = skb_peek_tail(&sk->sk_write_queue);
1208         if (!skb)
1209                 return -EINVAL;
1210
1211         cork->length += size;
1212         if ((size + skb->len > mtu) &&
1213             (sk->sk_protocol == IPPROTO_UDP) &&
1214             (rt->dst.dev->features & NETIF_F_UFO)) {
1215                 skb_shinfo(skb)->gso_size = mtu - fragheaderlen;
1216                 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
1217         }
1218
1219         while (size > 0) {
1220                 int i;
1221
1222                 if (skb_is_gso(skb))
1223                         len = size;
1224                 else {
1225
1226                         /* Check if the remaining data fits into current packet. */
1227                         len = mtu - skb->len;
1228                         if (len < size)
1229                                 len = maxfraglen - skb->len;
1230                 }
1231                 if (len <= 0) {
1232                         struct sk_buff *skb_prev;
1233                         int alloclen;
1234
1235                         skb_prev = skb;
1236                         fraggap = skb_prev->len - maxfraglen;
1237
1238                         alloclen = fragheaderlen + hh_len + fraggap + 15;
1239                         skb = sock_wmalloc(sk, alloclen, 1, sk->sk_allocation);
1240                         if (unlikely(!skb)) {
1241                                 err = -ENOBUFS;
1242                                 goto error;
1243                         }
1244
1245                         /*
1246                          *      Fill in the control structures
1247                          */
1248                         skb->ip_summed = CHECKSUM_NONE;
1249                         skb->csum = 0;
1250                         skb_reserve(skb, hh_len);
1251
1252                         /*
1253                          *      Find where to start putting bytes.
1254                          */
1255                         skb_put(skb, fragheaderlen + fraggap);
1256                         skb_reset_network_header(skb);
1257                         skb->transport_header = (skb->network_header +
1258                                                  fragheaderlen);
1259                         if (fraggap) {
1260                                 skb->csum = skb_copy_and_csum_bits(skb_prev,
1261                                                                    maxfraglen,
1262                                                     skb_transport_header(skb),
1263                                                                    fraggap, 0);
1264                                 skb_prev->csum = csum_sub(skb_prev->csum,
1265                                                           skb->csum);
1266                                 pskb_trim_unique(skb_prev, maxfraglen);
1267                         }
1268
1269                         /*
1270                          * Put the packet on the pending queue.
1271                          */
1272                         __skb_queue_tail(&sk->sk_write_queue, skb);
1273                         continue;
1274                 }
1275
1276                 i = skb_shinfo(skb)->nr_frags;
1277                 if (len > size)
1278                         len = size;
1279                 if (skb_can_coalesce(skb, i, page, offset)) {
1280                         skb_frag_size_add(&skb_shinfo(skb)->frags[i-1], len);
1281                 } else if (i < MAX_SKB_FRAGS) {
1282                         get_page(page);
1283                         skb_fill_page_desc(skb, i, page, offset, len);
1284                 } else {
1285                         err = -EMSGSIZE;
1286                         goto error;
1287                 }
1288
1289                 if (skb->ip_summed == CHECKSUM_NONE) {
1290                         __wsum csum;
1291                         csum = csum_page(page, offset, len);
1292                         skb->csum = csum_block_add(skb->csum, csum, skb->len);
1293                 }
1294
1295                 skb->len += len;
1296                 skb->data_len += len;
1297                 skb->truesize += len;
1298                 atomic_add(len, &sk->sk_wmem_alloc);
1299                 offset += len;
1300                 size -= len;
1301         }
1302         return 0;
1303
1304 error:
1305         cork->length -= size;
1306         IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
1307         return err;
1308 }
1309
1310 static void ip_cork_release(struct inet_cork *cork)
1311 {
1312         cork->flags &= ~IPCORK_OPT;
1313         kfree(cork->opt);
1314         cork->opt = NULL;
1315         dst_release(cork->dst);
1316         cork->dst = NULL;
1317 }
1318
1319 /*
1320  *      Combined all pending IP fragments on the socket as one IP datagram
1321  *      and push them out.
1322  */
1323 struct sk_buff *__ip_make_skb(struct sock *sk,
1324                               struct flowi4 *fl4,
1325                               struct sk_buff_head *queue,
1326                               struct inet_cork *cork)
1327 {
1328         struct sk_buff *skb, *tmp_skb;
1329         struct sk_buff **tail_skb;
1330         struct inet_sock *inet = inet_sk(sk);
1331         struct net *net = sock_net(sk);
1332         struct ip_options *opt = NULL;
1333         struct rtable *rt = (struct rtable *)cork->dst;
1334         struct iphdr *iph;
1335         __be16 df = 0;
1336         __u8 ttl;
1337
1338         skb = __skb_dequeue(queue);
1339         if (!skb)
1340                 goto out;
1341         tail_skb = &(skb_shinfo(skb)->frag_list);
1342
1343         /* move skb->data to ip header from ext header */
1344         if (skb->data < skb_network_header(skb))
1345                 __skb_pull(skb, skb_network_offset(skb));
1346         while ((tmp_skb = __skb_dequeue(queue)) != NULL) {
1347                 __skb_pull(tmp_skb, skb_network_header_len(skb));
1348                 *tail_skb = tmp_skb;
1349                 tail_skb = &(tmp_skb->next);
1350                 skb->len += tmp_skb->len;
1351                 skb->data_len += tmp_skb->len;
1352                 skb->truesize += tmp_skb->truesize;
1353                 tmp_skb->destructor = NULL;
1354                 tmp_skb->sk = NULL;
1355         }
1356
1357         /* Unless user demanded real pmtu discovery (IP_PMTUDISC_DO), we allow
1358          * to fragment the frame generated here. No matter, what transforms
1359          * how transforms change size of the packet, it will come out.
1360          */
1361         skb->ignore_df = ip_sk_ignore_df(sk);
1362
1363         /* DF bit is set when we want to see DF on outgoing frames.
1364          * If ignore_df is set too, we still allow to fragment this frame
1365          * locally. */
1366         if (inet->pmtudisc == IP_PMTUDISC_DO ||
1367             inet->pmtudisc == IP_PMTUDISC_PROBE ||
1368             (skb->len <= dst_mtu(&rt->dst) &&
1369              ip_dont_fragment(sk, &rt->dst)))
1370                 df = htons(IP_DF);
1371
1372         if (cork->flags & IPCORK_OPT)
1373                 opt = cork->opt;
1374
1375         if (cork->ttl != 0)
1376                 ttl = cork->ttl;
1377         else if (rt->rt_type == RTN_MULTICAST)
1378                 ttl = inet->mc_ttl;
1379         else
1380                 ttl = ip_select_ttl(inet, &rt->dst);
1381
1382         iph = ip_hdr(skb);
1383         iph->version = 4;
1384         iph->ihl = 5;
1385         iph->tos = (cork->tos != -1) ? cork->tos : inet->tos;
1386         iph->frag_off = df;
1387         iph->ttl = ttl;
1388         iph->protocol = sk->sk_protocol;
1389         ip_copy_addrs(iph, fl4);
1390         ip_select_ident(net, skb, sk);
1391
1392         if (opt) {
1393                 iph->ihl += opt->optlen>>2;
1394                 ip_options_build(skb, opt, cork->addr, rt, 0);
1395         }
1396
1397         skb->priority = (cork->tos != -1) ? cork->priority: sk->sk_priority;
1398         skb->mark = sk->sk_mark;
1399         /*
1400          * Steal rt from cork.dst to avoid a pair of atomic_inc/atomic_dec
1401          * on dst refcount
1402          */
1403         cork->dst = NULL;
1404         skb_dst_set(skb, &rt->dst);
1405
1406         if (iph->protocol == IPPROTO_ICMP)
1407                 icmp_out_count(net, ((struct icmphdr *)
1408                         skb_transport_header(skb))->type);
1409
1410         ip_cork_release(cork);
1411 out:
1412         return skb;
1413 }
1414
1415 int ip_send_skb(struct net *net, struct sk_buff *skb)
1416 {
1417         int err;
1418
1419         err = ip_local_out(skb);
1420         if (err) {
1421                 if (err > 0)
1422                         err = net_xmit_errno(err);
1423                 if (err)
1424                         IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS);
1425         }
1426
1427         return err;
1428 }
1429
1430 int ip_push_pending_frames(struct sock *sk, struct flowi4 *fl4)
1431 {
1432         struct sk_buff *skb;
1433
1434         skb = ip_finish_skb(sk, fl4);
1435         if (!skb)
1436                 return 0;
1437
1438         /* Netfilter gets whole the not fragmented skb. */
1439         return ip_send_skb(sock_net(sk), skb);
1440 }
1441
1442 /*
1443  *      Throw away all pending data on the socket.
1444  */
1445 static void __ip_flush_pending_frames(struct sock *sk,
1446                                       struct sk_buff_head *queue,
1447                                       struct inet_cork *cork)
1448 {
1449         struct sk_buff *skb;
1450
1451         while ((skb = __skb_dequeue_tail(queue)) != NULL)
1452                 kfree_skb(skb);
1453
1454         ip_cork_release(cork);
1455 }
1456
1457 void ip_flush_pending_frames(struct sock *sk)
1458 {
1459         __ip_flush_pending_frames(sk, &sk->sk_write_queue, &inet_sk(sk)->cork.base);
1460 }
1461
1462 struct sk_buff *ip_make_skb(struct sock *sk,
1463                             struct flowi4 *fl4,
1464                             int getfrag(void *from, char *to, int offset,
1465                                         int len, int odd, struct sk_buff *skb),
1466                             void *from, int length, int transhdrlen,
1467                             struct ipcm_cookie *ipc, struct rtable **rtp,
1468                             unsigned int flags)
1469 {
1470         struct inet_cork cork;
1471         struct sk_buff_head queue;
1472         int err;
1473
1474         if (flags & MSG_PROBE)
1475                 return NULL;
1476
1477         __skb_queue_head_init(&queue);
1478
1479         cork.flags = 0;
1480         cork.addr = 0;
1481         cork.opt = NULL;
1482         err = ip_setup_cork(sk, &cork, ipc, rtp);
1483         if (err)
1484                 return ERR_PTR(err);
1485
1486         err = __ip_append_data(sk, fl4, &queue, &cork,
1487                                &current->task_frag, getfrag,
1488                                from, length, transhdrlen, flags);
1489         if (err) {
1490                 __ip_flush_pending_frames(sk, &queue, &cork);
1491                 return ERR_PTR(err);
1492         }
1493
1494         return __ip_make_skb(sk, fl4, &queue, &cork);
1495 }
1496
1497 /*
1498  *      Fetch data from kernel space and fill in checksum if needed.
1499  */
1500 static int ip_reply_glue_bits(void *dptr, char *to, int offset,
1501                               int len, int odd, struct sk_buff *skb)
1502 {
1503         __wsum csum;
1504
1505         csum = csum_partial_copy_nocheck(dptr+offset, to, len, 0);
1506         skb->csum = csum_block_add(skb->csum, csum, odd);
1507         return 0;
1508 }
1509
1510 /*
1511  *      Generic function to send a packet as reply to another packet.
1512  *      Used to send some TCP resets/acks so far.
1513  */
1514 void ip_send_unicast_reply(struct sock *sk, struct sk_buff *skb,
1515                            const struct ip_options *sopt,
1516                            __be32 daddr, __be32 saddr,
1517                            const struct ip_reply_arg *arg,
1518                            unsigned int len)
1519 {
1520         struct ip_options_data replyopts;
1521         struct ipcm_cookie ipc;
1522         struct flowi4 fl4;
1523         struct rtable *rt = skb_rtable(skb);
1524         struct net *net = sock_net(sk);
1525         struct sk_buff *nskb;
1526         int err;
1527
1528         if (__ip_options_echo(&replyopts.opt.opt, skb, sopt))
1529                 return;
1530
1531         ipc.addr = daddr;
1532         ipc.opt = NULL;
1533         ipc.tx_flags = 0;
1534         ipc.ttl = 0;
1535         ipc.tos = -1;
1536
1537         if (replyopts.opt.opt.optlen) {
1538                 ipc.opt = &replyopts.opt;
1539
1540                 if (replyopts.opt.opt.srr)
1541                         daddr = replyopts.opt.opt.faddr;
1542         }
1543
1544         flowi4_init_output(&fl4, arg->bound_dev_if,
1545                            IP4_REPLY_MARK(net, skb->mark),
1546                            RT_TOS(arg->tos),
1547                            RT_SCOPE_UNIVERSE, ip_hdr(skb)->protocol,
1548                            ip_reply_arg_flowi_flags(arg),
1549                            daddr, saddr,
1550                            tcp_hdr(skb)->source, tcp_hdr(skb)->dest);
1551         security_skb_classify_flow(skb, flowi4_to_flowi(&fl4));
1552         rt = ip_route_output_key(net, &fl4);
1553         if (IS_ERR(rt))
1554                 return;
1555
1556         inet_sk(sk)->tos = arg->tos;
1557
1558         sk->sk_priority = skb->priority;
1559         sk->sk_protocol = ip_hdr(skb)->protocol;
1560         sk->sk_bound_dev_if = arg->bound_dev_if;
1561         sk->sk_sndbuf = sysctl_wmem_default;
1562         err = ip_append_data(sk, &fl4, ip_reply_glue_bits, arg->iov->iov_base,
1563                              len, 0, &ipc, &rt, MSG_DONTWAIT);
1564         if (unlikely(err)) {
1565                 ip_flush_pending_frames(sk);
1566                 goto out;
1567         }
1568
1569         nskb = skb_peek(&sk->sk_write_queue);
1570         if (nskb) {
1571                 if (arg->csumoffset >= 0)
1572                         *((__sum16 *)skb_transport_header(nskb) +
1573                           arg->csumoffset) = csum_fold(csum_add(nskb->csum,
1574                                                                 arg->csum));
1575                 nskb->ip_summed = CHECKSUM_NONE;
1576                 skb_set_queue_mapping(nskb, skb_get_queue_mapping(skb));
1577                 ip_push_pending_frames(sk, &fl4);
1578         }
1579 out:
1580         ip_rt_put(rt);
1581 }
1582
1583 void __init ip_init(void)
1584 {
1585         ip_rt_init();
1586         inet_initpeers();
1587
1588 #if defined(CONFIG_IP_MULTICAST)
1589         igmp_mc_init();
1590 #endif
1591 }