SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/ipv4/ip_sockglue.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 IP to API glue.
   7  *
   8  * Version:     $Id: ip_sockglue.c,v 1.62 2002/02/01 22:01:04 davem Exp $
   9  *
  10  * Authors:     see ip.c
  11  *
  12  * Fixes:
  13  *              Many            :       Split from ip.c , see ip.c for history.
  14  *              Martin Mares    :       TOS setting fixed.
  15  *              Alan Cox        :       Fixed a couple of oopses in Martin's
  16  *                                      TOS tweaks.
  17  *              Mike McLagan    :       Routing by source
  18  */
  19
  20 #include <linux/config.h>
  21 #include <linux/module.h>
  22 #include <linux/types.h>
  23 #include <linux/mm.h>
  24 #include <linux/sched.h>
  25 #include <linux/skbuff.h>
  26 #include <linux/ip.h>
  27 #include <linux/icmp.h>
  28 #include <linux/netdevice.h>
  29 #include <net/sock.h>
  30 #include <net/ip.h>
  31 #include <net/icmp.h>
  32 #include <net/tcp_states.h>
  33 #include <linux/udp.h>
  34 #include <linux/igmp.h>
  35 #include <linux/netfilter.h>
  36 #include <linux/route.h>
  37 #include <linux/mroute.h>
  38 #include <net/route.h>
  39 #include <net/xfrm.h>
  40 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
  41 #include <net/transp_v6.h>
  42 #endif
  43
  44 #include <linux/errqueue.h>
  45 #include <asm/uaccess.h>
  46
  47 #define IP_CMSG_PKTINFO         1
  48 #define IP_CMSG_TTL             2
  49 #define IP_CMSG_TOS             4
  50 #define IP_CMSG_RECVOPTS        8
  51 #define IP_CMSG_RETOPTS         16
  52
  53 /*
  54  *      SOL_IP control messages.
  55  */
  56
  57 static void ip_cmsg_recv_pktinfo(struct msghdr *msg, struct sk_buff *skb)
  58 {
  59         struct in_pktinfo info;
  60         struct rtable *rt = (struct rtable *)skb->dst;
  61
  62         info.ipi_addr.s_addr = skb->nh.iph->daddr;
  63         if (rt) {
  64                 info.ipi_ifindex = rt->rt_iif;
  65                 info.ipi_spec_dst.s_addr = rt->rt_spec_dst;
  66         } else {
  67                 info.ipi_ifindex = 0;
  68                 info.ipi_spec_dst.s_addr = 0;
  69         }
  70
  71         put_cmsg(msg, SOL_IP, IP_PKTINFO, sizeof(info), &info);
  72 }
  73
  74 static void ip_cmsg_recv_ttl(struct msghdr *msg, struct sk_buff *skb)
  75 {
  76         int ttl = skb->nh.iph->ttl;
  77         put_cmsg(msg, SOL_IP, IP_TTL, sizeof(int), &ttl);
  78 }
  79
  80 static void ip_cmsg_recv_tos(struct msghdr *msg, struct sk_buff *skb)
  81 {
  82         put_cmsg(msg, SOL_IP, IP_TOS, 1, &skb->nh.iph->tos);
  83 }
  84
  85 static void ip_cmsg_recv_opts(struct msghdr *msg, struct sk_buff *skb)
  86 {
  87         if (IPCB(skb)->opt.optlen == 0)
  88                 return;
  89
  90         put_cmsg(msg, SOL_IP, IP_RECVOPTS, IPCB(skb)->opt.optlen, skb->nh.iph+1);
  91 }
  92
  93
  94 static void ip_cmsg_recv_retopts(struct msghdr *msg, struct sk_buff *skb)
  95 {
  96         unsigned char optbuf[sizeof(struct ip_options) + 40];
  97         struct ip_options * opt = (struct ip_options*)optbuf;
  98
  99         if (IPCB(skb)->opt.optlen == 0)
 100                 return;
 101
 102         if (ip_options_echo(opt, skb)) {
 103                 msg->msg_flags |= MSG_CTRUNC;
 104                 return;
 105         }
 106         ip_options_undo(opt);
 107
 108         put_cmsg(msg, SOL_IP, IP_RETOPTS, opt->optlen, opt->__data);
 109 }
 110
 111
 112 void ip_cmsg_recv(struct msghdr *msg, struct sk_buff *skb)
 113 {
 114         struct inet_sock *inet = inet_sk(skb->sk);
 115         unsigned flags = inet->cmsg_flags;
 116
 117         /* Ordered by supposed usage frequency */
 118         if (flags & 1)
 119                 ip_cmsg_recv_pktinfo(msg, skb);
 120         if ((flags>>=1) == 0)
 121                 return;
 122
 123         if (flags & 1)
 124                 ip_cmsg_recv_ttl(msg, skb);
 125         if ((flags>>=1) == 0)
 126                 return;
 127
 128         if (flags & 1)
 129                 ip_cmsg_recv_tos(msg, skb);
 130         if ((flags>>=1) == 0)
 131                 return;
 132
 133         if (flags & 1)
 134                 ip_cmsg_recv_opts(msg, skb);
 135         if ((flags>>=1) == 0)
 136                 return;
 137
 138         if (flags & 1)
 139                 ip_cmsg_recv_retopts(msg, skb);
 140 }
 141
 142 int ip_cmsg_send(struct msghdr *msg, struct ipcm_cookie *ipc)
 143 {
 144         int err;
 145         struct cmsghdr *cmsg;
 146
 147         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 148                 if (!CMSG_OK(msg, cmsg))
 149                         return -EINVAL;
 150                 if (cmsg->cmsg_level != SOL_IP)
 151                         continue;
 152                 switch (cmsg->cmsg_type) {
 153                 case IP_RETOPTS:
 154                         err = cmsg->cmsg_len - CMSG_ALIGN(sizeof(struct cmsghdr));
 155                         err = ip_options_get(&ipc->opt, CMSG_DATA(cmsg), err < 40 ? err : 40);
 156                         if (err)
 157                                 return err;
 158                         break;
 159                 case IP_PKTINFO:
 160                 {
 161                         struct in_pktinfo *info;
 162                         if (cmsg->cmsg_len != CMSG_LEN(sizeof(struct in_pktinfo)))
 163                                 return -EINVAL;
 164                         info = (struct in_pktinfo *)CMSG_DATA(cmsg);
 165                         ipc->oif = info->ipi_ifindex;
 166                         ipc->addr = info->ipi_spec_dst.s_addr;
 167                         break;
 168                 }
 169                 default:
 170                         return -EINVAL;
 171                 }
 172         }
 173         return 0;
 174 }
 175
 176
 177 /* Special input handler for packets caught by router alert option.
 178    They are selected only by protocol field, and then processed likely
 179    local ones; but only if someone wants them! Otherwise, router
 180    not running rsvpd will kill RSVP.
 181
 182    It is user level problem, what it will make with them.
 183    I have no idea, how it will masquearde or NAT them (it is joke, joke :-)),
 184    but receiver should be enough clever f.e. to forward mtrace requests,
 185    sent to multicast group to reach destination designated router.
 186  */
 187 struct ip_ra_chain *ip_ra_chain;
 188 DEFINE_RWLOCK(ip_ra_lock);
 189
 190 int ip_ra_control(struct sock *sk, unsigned char on, void (*destructor)(struct sock *))
 191 {
 192         struct ip_ra_chain *ra, *new_ra, **rap;
 193
 194         if (sk->sk_type != SOCK_RAW || inet_sk(sk)->num == IPPROTO_RAW)
 195                 return -EINVAL;
 196
 197         new_ra = on ? kmalloc(sizeof(*new_ra), GFP_KERNEL) : NULL;
 198
 199         write_lock_bh(&ip_ra_lock);
 200         for (rap = &ip_ra_chain; (ra=*rap) != NULL; rap = &ra->next) {
 201                 if (ra->sk == sk) {
 202                         if (on) {
 203                                 write_unlock_bh(&ip_ra_lock);
 204                                 kfree(new_ra);
 205                                 return -EADDRINUSE;
 206                         }
 207                         *rap = ra->next;
 208                         write_unlock_bh(&ip_ra_lock);
 209
 210                         if (ra->destructor)
 211                                 ra->destructor(sk);
 212                         sock_put(sk);
 213                         kfree(ra);
 214                         return 0;
 215                 }
 216         }
 217         if (new_ra == NULL) {
 218                 write_unlock_bh(&ip_ra_lock);
 219                 return -ENOBUFS;
 220         }
 221         new_ra->sk = sk;
 222         new_ra->destructor = destructor;
 223
 224         new_ra->next = ra;
 225         *rap = new_ra;
 226         sock_hold(sk);
 227         write_unlock_bh(&ip_ra_lock);
 228
 229         return 0;
 230 }
 231
 232 void ip_icmp_error(struct sock *sk, struct sk_buff *skb, int err,
 233                    u16 port, u32 info, u8 *payload)
 234 {
 235         struct inet_sock *inet = inet_sk(sk);
 236         struct sock_exterr_skb *serr;
 237
 238         if (!inet->recverr)
 239                 return;
 240
 241         skb = skb_clone(skb, GFP_ATOMIC);
 242         if (!skb)
 243                 return;
 244
 245         serr = SKB_EXT_ERR(skb);
 246         serr->ee.ee_errno = err;
 247         serr->ee.ee_origin = SO_EE_ORIGIN_ICMP;
 248         serr->ee.ee_type = skb->h.icmph->type;
 249         serr->ee.ee_code = skb->h.icmph->code;
 250         serr->ee.ee_pad = 0;
 251         serr->ee.ee_info = info;
 252         serr->ee.ee_data = 0;
 253         serr->addr_offset = (u8*)&(((struct iphdr*)(skb->h.icmph+1))->daddr) - skb->nh.raw;
 254         serr->port = port;
 255
 256         skb->h.raw = payload;
 257         if (!skb_pull(skb, payload - skb->data) ||
 258             sock_queue_err_skb(sk, skb))
 259                 kfree_skb(skb);
 260 }
 261
 262 void ip_local_error(struct sock *sk, int err, u32 daddr, u16 port, u32 info)
 263 {
 264         struct inet_sock *inet = inet_sk(sk);
 265         struct sock_exterr_skb *serr;
 266         struct iphdr *iph;
 267         struct sk_buff *skb;
 268
 269         if (!inet->recverr)
 270                 return;
 271
 272         skb = alloc_skb(sizeof(struct iphdr), GFP_ATOMIC);
 273         if (!skb)
 274                 return;
 275
 276         iph = (struct iphdr*)skb_put(skb, sizeof(struct iphdr));
 277         skb->nh.iph = iph;
 278         iph->daddr = daddr;
 279
 280         serr = SKB_EXT_ERR(skb);
 281         serr->ee.ee_errno = err;
 282         serr->ee.ee_origin = SO_EE_ORIGIN_LOCAL;
 283         serr->ee.ee_type = 0;
 284         serr->ee.ee_code = 0;
 285         serr->ee.ee_pad = 0;
 286         serr->ee.ee_info = info;
 287         serr->ee.ee_data = 0;
 288         serr->addr_offset = (u8*)&iph->daddr - skb->nh.raw;
 289         serr->port = port;
 290
 291         skb->h.raw = skb->tail;
 292         __skb_pull(skb, skb->tail - skb->data);
 293
 294         if (sock_queue_err_skb(sk, skb))
 295                 kfree_skb(skb);
 296 }
 297
 298 /*
 299  *      Handle MSG_ERRQUEUE
 300  */
 301 int ip_recv_error(struct sock *sk, struct msghdr *msg, int len)
 302 {
 303         struct sock_exterr_skb *serr;
 304         struct sk_buff *skb, *skb2;
 305         struct sockaddr_in *sin;
 306         struct {
 307                 struct sock_extended_err ee;
 308                 struct sockaddr_in       offender;
 309         } errhdr;
 310         int err;
 311         int copied;
 312
 313         err = -EAGAIN;
 314         skb = skb_dequeue(&sk->sk_error_queue);
 315         if (skb == NULL)
 316                 goto out;
 317
 318         copied = skb->len;
 319         if (copied > len) {
 320                 msg->msg_flags |= MSG_TRUNC;
 321                 copied = len;
 322         }
 323         err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, copied);
 324         if (err)
 325                 goto out_free_skb;
 326
 327         sock_recv_timestamp(msg, sk, skb);
 328
 329         serr = SKB_EXT_ERR(skb);
 330
 331         sin = (struct sockaddr_in *)msg->msg_name;
 332         if (sin) {
 333                 sin->sin_family = AF_INET;
 334                 sin->sin_addr.s_addr = *(u32*)(skb->nh.raw + serr->addr_offset);
 335                 sin->sin_port = serr->port;
 336                 memset(&sin->sin_zero, 0, sizeof(sin->sin_zero));
 337         }
 338
 339         memcpy(&errhdr.ee, &serr->ee, sizeof(struct sock_extended_err));
 340         sin = &errhdr.offender;
 341         sin->sin_family = AF_UNSPEC;
 342         if (serr->ee.ee_origin == SO_EE_ORIGIN_ICMP) {
 343                 struct inet_sock *inet = inet_sk(sk);
 344
 345                 sin->sin_family = AF_INET;
 346                 sin->sin_addr.s_addr = skb->nh.iph->saddr;
 347                 sin->sin_port = 0;
 348                 memset(&sin->sin_zero, 0, sizeof(sin->sin_zero));
 349                 if (inet->cmsg_flags)
 350                         ip_cmsg_recv(msg, skb);
 351         }
 352
 353         put_cmsg(msg, SOL_IP, IP_RECVERR, sizeof(errhdr), &errhdr);
 354
 355         /* Now we could try to dump offended packet options */
 356
 357         msg->msg_flags |= MSG_ERRQUEUE;
 358         err = copied;
 359
 360         /* Reset and regenerate socket error */
 361         spin_lock_bh(&sk->sk_error_queue.lock);
 362         sk->sk_err = 0;
 363         if ((skb2 = skb_peek(&sk->sk_error_queue)) != NULL) {
 364                 sk->sk_err = SKB_EXT_ERR(skb2)->ee.ee_errno;
 365                 spin_unlock_bh(&sk->sk_error_queue.lock);
 366                 sk->sk_error_report(sk);
 367         } else
 368                 spin_unlock_bh(&sk->sk_error_queue.lock);
 369
 370 out_free_skb:
 371         kfree_skb(skb);
 372 out:
 373         return err;
 374 }
 375
 376
 377 /*
 378  *      Socket option code for IP. This is the end of the line after any TCP,UDP etc options on
 379  *      an IP socket.
 380  */
 381
 382 int ip_setsockopt(struct sock *sk, int level, int optname, char __user *optval, int optlen)
 383 {
 384         struct inet_sock *inet = inet_sk(sk);
 385         int val=0,err;
 386
 387         if (level != SOL_IP)
 388                 return -ENOPROTOOPT;
 389
 390         if (((1<<optname) & ((1<<IP_PKTINFO) | (1<<IP_RECVTTL) |
 391                             (1<<IP_RECVOPTS) | (1<<IP_RECVTOS) |
 392                             (1<<IP_RETOPTS) | (1<<IP_TOS) |
 393                             (1<<IP_TTL) | (1<<IP_HDRINCL) |
 394                             (1<<IP_MTU_DISCOVER) | (1<<IP_RECVERR) |
 395                             (1<<IP_ROUTER_ALERT) | (1<<IP_FREEBIND))) ||
 396                                 optname == IP_MULTICAST_TTL ||
 397                                 optname == IP_MULTICAST_LOOP) {
 398                 if (optlen >= sizeof(int)) {
 399                         if (get_user(val, (int __user *) optval))
 400                                 return -EFAULT;
 401                 } else if (optlen >= sizeof(char)) {
 402                         unsigned char ucval;
 403
 404                         if (get_user(ucval, (unsigned char __user *) optval))
 405                                 return -EFAULT;
 406                         val = (int) ucval;
 407                 }
 408         }
 409
 410         /* If optlen==0, it is equivalent to val == 0 */
 411
 412 #ifdef CONFIG_IP_MROUTE
 413         if (optname >= MRT_BASE && optname <= (MRT_BASE + 10))
 414                 return ip_mroute_setsockopt(sk,optname,optval,optlen);
 415 #endif
 416
 417         err = 0;
 418         lock_sock(sk);
 419
 420         switch (optname) {
 421                 case IP_OPTIONS:
 422                 {
 423                         struct ip_options * opt = NULL;
 424                         if (optlen > 40 || optlen < 0)
 425                                 goto e_inval;
 426                         err = ip_options_get_from_user(&opt, optval, optlen);
 427                         if (err)
 428                                 break;
 429                         if (inet->is_icsk) {
 430                                 struct inet_connection_sock *icsk = inet_csk(sk);
 431 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 432                                 if (sk->sk_family == PF_INET ||
 433                                     (!((1 << sk->sk_state) &
 434                                        (TCPF_LISTEN | TCPF_CLOSE)) &&
 435                                      inet->daddr != LOOPBACK4_IPV6)) {
 436 #endif
 437                                         if (inet->opt)
 438                                                 icsk->icsk_ext_hdr_len -= inet->opt->optlen;
 439                                         if (opt)
 440                                                 icsk->icsk_ext_hdr_len += opt->optlen;
 441                                         icsk->icsk_sync_mss(sk, icsk->icsk_pmtu_cookie);
 442 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 443                                 }
 444 #endif
 445                         }
 446                         opt = xchg(&inet->opt, opt);
 447                         kfree(opt);
 448                         break;
 449                 }
 450                 case IP_PKTINFO:
 451                         if (val)
 452                                 inet->cmsg_flags |= IP_CMSG_PKTINFO;
 453                         else
 454                                 inet->cmsg_flags &= ~IP_CMSG_PKTINFO;
 455                         break;
 456                 case IP_RECVTTL:
 457                         if (val)
 458                                 inet->cmsg_flags |=  IP_CMSG_TTL;
 459                         else
 460                                 inet->cmsg_flags &= ~IP_CMSG_TTL;
 461                         break;
 462                 case IP_RECVTOS:
 463                         if (val)
 464                                 inet->cmsg_flags |=  IP_CMSG_TOS;
 465                         else
 466                                 inet->cmsg_flags &= ~IP_CMSG_TOS;
 467                         break;
 468                 case IP_RECVOPTS:
 469                         if (val)
 470                                 inet->cmsg_flags |=  IP_CMSG_RECVOPTS;
 471                         else
 472                                 inet->cmsg_flags &= ~IP_CMSG_RECVOPTS;
 473                         break;
 474                 case IP_RETOPTS:
 475                         if (val)
 476                                 inet->cmsg_flags |= IP_CMSG_RETOPTS;
 477                         else
 478                                 inet->cmsg_flags &= ~IP_CMSG_RETOPTS;
 479                         break;
 480                 case IP_TOS:    /* This sets both TOS and Precedence */
 481                         if (sk->sk_type == SOCK_STREAM) {
 482                                 val &= ~3;
 483                                 val |= inet->tos & 3;
 484                         }
 485                         if (IPTOS_PREC(val) >= IPTOS_PREC_CRITIC_ECP &&
 486                             !capable(CAP_NET_ADMIN)) {
 487                                 err = -EPERM;
 488                                 break;
 489                         }
 490                         if (inet->tos != val) {
 491                                 inet->tos = val;
 492                                 sk->sk_priority = rt_tos2priority(val);
 493                                 sk_dst_reset(sk);
 494                         }
 495                         break;
 496                 case IP_TTL:
 497                         if (optlen<1)
 498                                 goto e_inval;
 499                         if (val != -1 && (val < 1 || val>255))
 500                                 goto e_inval;
 501                         inet->uc_ttl = val;
 502                         break;
 503                 case IP_HDRINCL:
 504                         if (sk->sk_type != SOCK_RAW) {
 505                                 err = -ENOPROTOOPT;
 506                                 break;
 507                         }
 508                         inet->hdrincl = val ? 1 : 0;
 509                         break;
 510                 case IP_MTU_DISCOVER:
 511                         if (val<0 || val>2)
 512                                 goto e_inval;
 513                         inet->pmtudisc = val;
 514                         break;
 515                 case IP_RECVERR:
 516                         inet->recverr = !!val;
 517                         if (!val)
 518                                 skb_queue_purge(&sk->sk_error_queue);
 519                         break;
 520                 case IP_MULTICAST_TTL:
 521                         if (sk->sk_type == SOCK_STREAM)
 522                                 goto e_inval;
 523                         if (optlen<1)
 524                                 goto e_inval;
 525                         if (val==-1)
 526                                 val = 1;
 527                         if (val < 0 || val > 255)
 528                                 goto e_inval;
 529                         inet->mc_ttl = val;
 530                         break;
 531                 case IP_MULTICAST_LOOP:
 532                         if (optlen<1)
 533                                 goto e_inval;
 534                         inet->mc_loop = !!val;
 535                         break;
 536                 case IP_MULTICAST_IF:
 537                 {
 538                         struct ip_mreqn mreq;
 539                         struct net_device *dev = NULL;
 540
 541                         if (sk->sk_type == SOCK_STREAM)
 542                                 goto e_inval;
 543                         /*
 544                          *      Check the arguments are allowable
 545                          */
 546
 547                         err = -EFAULT;
 548                         if (optlen >= sizeof(struct ip_mreqn)) {
 549                                 if (copy_from_user(&mreq,optval,sizeof(mreq)))
 550                                         break;
 551                         } else {
 552                                 memset(&mreq, 0, sizeof(mreq));
 553                                 if (optlen >= sizeof(struct in_addr) &&
 554                                     copy_from_user(&mreq.imr_address,optval,sizeof(struct in_addr)))
 555                                         break;
 556                         }
 557
 558                         if (!mreq.imr_ifindex) {
 559                                 if (mreq.imr_address.s_addr == INADDR_ANY) {
 560                                         inet->mc_index = 0;
 561                                         inet->mc_addr  = 0;
 562                                         err = 0;
 563                                         break;
 564                                 }
 565                                 dev = ip_dev_find(mreq.imr_address.s_addr);
 566                                 if (dev) {
 567                                         mreq.imr_ifindex = dev->ifindex;
 568                                         dev_put(dev);
 569                                 }
 570                         } else
 571                                 dev = __dev_get_by_index(mreq.imr_ifindex);
 572
 573
 574                         err = -EADDRNOTAVAIL;
 575                         if (!dev)
 576                                 break;
 577
 578                         err = -EINVAL;
 579                         if (sk->sk_bound_dev_if &&
 580                             mreq.imr_ifindex != sk->sk_bound_dev_if)
 581                                 break;
 582
 583                         inet->mc_index = mreq.imr_ifindex;
 584                         inet->mc_addr  = mreq.imr_address.s_addr;
 585                         err = 0;
 586                         break;
 587                 }
 588
 589                 case IP_ADD_MEMBERSHIP:
 590                 case IP_DROP_MEMBERSHIP:
 591                 {
 592                         struct ip_mreqn mreq;
 593
 594                         if (optlen < sizeof(struct ip_mreq))
 595                                 goto e_inval;
 596                         err = -EFAULT;
 597                         if (optlen >= sizeof(struct ip_mreqn)) {
 598                                 if(copy_from_user(&mreq,optval,sizeof(mreq)))
 599                                         break;
 600                         } else {
 601                                 memset(&mreq, 0, sizeof(mreq));
 602                                 if (copy_from_user(&mreq,optval,sizeof(struct ip_mreq)))
 603                                         break;
 604                         }
 605
 606                         if (optname == IP_ADD_MEMBERSHIP)
 607                                 err = ip_mc_join_group(sk, &mreq);
 608                         else
 609                                 err = ip_mc_leave_group(sk, &mreq);
 610                         break;
 611                 }
 612                 case IP_MSFILTER:
 613                 {
 614                         extern int sysctl_igmp_max_msf;
 615                         struct ip_msfilter *msf;
 616
 617                         if (optlen < IP_MSFILTER_SIZE(0))
 618                                 goto e_inval;
 619                         if (optlen > sysctl_optmem_max) {
 620                                 err = -ENOBUFS;
 621                                 break;
 622                         }
 623                         msf = (struct ip_msfilter *)kmalloc(optlen, GFP_KERNEL);
 624                         if (msf == 0) {
 625                                 err = -ENOBUFS;
 626                                 break;
 627                         }
 628                         err = -EFAULT;
 629                         if (copy_from_user(msf, optval, optlen)) {
 630                                 kfree(msf);
 631                                 break;
 632                         }
 633                         /* numsrc >= (1G-4) overflow in 32 bits */
 634                         if (msf->imsf_numsrc >= 0x3ffffffcU ||
 635                             msf->imsf_numsrc > sysctl_igmp_max_msf) {
 636                                 kfree(msf);
 637                                 err = -ENOBUFS;
 638                                 break;
 639                         }
 640                         if (IP_MSFILTER_SIZE(msf->imsf_numsrc) > optlen) {
 641                                 kfree(msf);
 642                                 err = -EINVAL;
 643                                 break;
 644                         }
 645                         err = ip_mc_msfilter(sk, msf, 0);
 646                         kfree(msf);
 647                         break;
 648                 }
 649                 case IP_BLOCK_SOURCE:
 650                 case IP_UNBLOCK_SOURCE:
 651                 case IP_ADD_SOURCE_MEMBERSHIP:
 652                 case IP_DROP_SOURCE_MEMBERSHIP:
 653                 {
 654                         struct ip_mreq_source mreqs;
 655                         int omode, add;
 656
 657                         if (optlen != sizeof(struct ip_mreq_source))
 658                                 goto e_inval;
 659                         if (copy_from_user(&mreqs, optval, sizeof(mreqs))) {
 660                                 err = -EFAULT;
 661                                 break;
 662                         }
 663                         if (optname == IP_BLOCK_SOURCE) {
 664                                 omode = MCAST_EXCLUDE;
 665                                 add = 1;
 666                         } else if (optname == IP_UNBLOCK_SOURCE) {
 667                                 omode = MCAST_EXCLUDE;
 668                                 add = 0;
 669                         } else if (optname == IP_ADD_SOURCE_MEMBERSHIP) {
 670                                 struct ip_mreqn mreq;
 671
 672                                 mreq.imr_multiaddr.s_addr = mreqs.imr_multiaddr;
 673                                 mreq.imr_address.s_addr = mreqs.imr_interface;
 674                                 mreq.imr_ifindex = 0;
 675                                 err = ip_mc_join_group(sk, &mreq);
 676                                 if (err && err != -EADDRINUSE)
 677                                         break;
 678                                 omode = MCAST_INCLUDE;
 679                                 add = 1;
 680                         } else /* IP_DROP_SOURCE_MEMBERSHIP */ {
 681                                 omode = MCAST_INCLUDE;
 682                                 add = 0;
 683                         }
 684                         err = ip_mc_source(add, omode, sk, &mreqs, 0);
 685                         break;
 686                 }
 687                 case MCAST_JOIN_GROUP:
 688                 case MCAST_LEAVE_GROUP:
 689                 {
 690                         struct group_req greq;
 691                         struct sockaddr_in *psin;
 692                         struct ip_mreqn mreq;
 693
 694                         if (optlen < sizeof(struct group_req))
 695                                 goto e_inval;
 696                         err = -EFAULT;
 697                         if(copy_from_user(&greq, optval, sizeof(greq)))
 698                                 break;
 699                         psin = (struct sockaddr_in *)&greq.gr_group;
 700                         if (psin->sin_family != AF_INET)
 701                                 goto e_inval;
 702                         memset(&mreq, 0, sizeof(mreq));
 703                         mreq.imr_multiaddr = psin->sin_addr;
 704                         mreq.imr_ifindex = greq.gr_interface;
 705
 706                         if (optname == MCAST_JOIN_GROUP)
 707                                 err = ip_mc_join_group(sk, &mreq);
 708                         else
 709                                 err = ip_mc_leave_group(sk, &mreq);
 710                         break;
 711                 }
 712                 case MCAST_JOIN_SOURCE_GROUP:
 713                 case MCAST_LEAVE_SOURCE_GROUP:
 714                 case MCAST_BLOCK_SOURCE:
 715                 case MCAST_UNBLOCK_SOURCE:
 716                 {
 717                         struct group_source_req greqs;
 718                         struct ip_mreq_source mreqs;
 719                         struct sockaddr_in *psin;
 720                         int omode, add;
 721
 722                         if (optlen != sizeof(struct group_source_req))
 723                                 goto e_inval;
 724                         if (copy_from_user(&greqs, optval, sizeof(greqs))) {
 725                                 err = -EFAULT;
 726                                 break;
 727                         }
 728                         if (greqs.gsr_group.ss_family != AF_INET ||
 729                             greqs.gsr_source.ss_family != AF_INET) {
 730                                 err = -EADDRNOTAVAIL;
 731                                 break;
 732                         }
 733                         psin = (struct sockaddr_in *)&greqs.gsr_group;
 734                         mreqs.imr_multiaddr = psin->sin_addr.s_addr;
 735                         psin = (struct sockaddr_in *)&greqs.gsr_source;
 736                         mreqs.imr_sourceaddr = psin->sin_addr.s_addr;
 737                         mreqs.imr_interface = 0; /* use index for mc_source */
 738
 739                         if (optname == MCAST_BLOCK_SOURCE) {
 740                                 omode = MCAST_EXCLUDE;
 741                                 add = 1;
 742                         } else if (optname == MCAST_UNBLOCK_SOURCE) {
 743                                 omode = MCAST_EXCLUDE;
 744                                 add = 0;
 745                         } else if (optname == MCAST_JOIN_SOURCE_GROUP) {
 746                                 struct ip_mreqn mreq;
 747
 748                                 psin = (struct sockaddr_in *)&greqs.gsr_group;
 749                                 mreq.imr_multiaddr = psin->sin_addr;
 750                                 mreq.imr_address.s_addr = 0;
 751                                 mreq.imr_ifindex = greqs.gsr_interface;
 752                                 err = ip_mc_join_group(sk, &mreq);
 753                                 if (err && err != -EADDRINUSE)
 754                                         break;
 755                                 greqs.gsr_interface = mreq.imr_ifindex;
 756                                 omode = MCAST_INCLUDE;
 757                                 add = 1;
 758                         } else /* MCAST_LEAVE_SOURCE_GROUP */ {
 759                                 omode = MCAST_INCLUDE;
 760                                 add = 0;
 761                         }
 762                         err = ip_mc_source(add, omode, sk, &mreqs,
 763                                 greqs.gsr_interface);
 764                         break;
 765                 }
 766                 case MCAST_MSFILTER:
 767                 {
 768                         extern int sysctl_igmp_max_msf;
 769                         struct sockaddr_in *psin;
 770                         struct ip_msfilter *msf = NULL;
 771                         struct group_filter *gsf = NULL;
 772                         int msize, i, ifindex;
 773
 774                         if (optlen < GROUP_FILTER_SIZE(0))
 775                                 goto e_inval;
 776                         if (optlen > sysctl_optmem_max) {
 777                                 err = -ENOBUFS;
 778                                 break;
 779                         }
 780                         gsf = (struct group_filter *)kmalloc(optlen,GFP_KERNEL);
 781                         if (gsf == 0) {
 782                                 err = -ENOBUFS;
 783                                 break;
 784                         }
 785                         err = -EFAULT;
 786                         if (copy_from_user(gsf, optval, optlen)) {
 787                                 goto mc_msf_out;
 788                         }
 789                         /* numsrc >= (4G-140)/128 overflow in 32 bits */
 790                         if (gsf->gf_numsrc >= 0x1ffffff ||
 791                             gsf->gf_numsrc > sysctl_igmp_max_msf) {
 792                                 err = -ENOBUFS;
 793                                 goto mc_msf_out;
 794                         }
 795                         if (GROUP_FILTER_SIZE(gsf->gf_numsrc) > optlen) {
 796                                 err = -EINVAL;
 797                                 goto mc_msf_out;
 798                         }
 799                         msize = IP_MSFILTER_SIZE(gsf->gf_numsrc);
 800                         msf = (struct ip_msfilter *)kmalloc(msize,GFP_KERNEL);
 801                         if (msf == 0) {
 802                                 err = -ENOBUFS;
 803                                 goto mc_msf_out;
 804                         }
 805                         ifindex = gsf->gf_interface;
 806                         psin = (struct sockaddr_in *)&gsf->gf_group;
 807                         if (psin->sin_family != AF_INET) {
 808                                 err = -EADDRNOTAVAIL;
 809                                 goto mc_msf_out;
 810                         }
 811                         msf->imsf_multiaddr = psin->sin_addr.s_addr;
 812                         msf->imsf_interface = 0;
 813                         msf->imsf_fmode = gsf->gf_fmode;
 814                         msf->imsf_numsrc = gsf->gf_numsrc;
 815                         err = -EADDRNOTAVAIL;
 816                         for (i=0; i<gsf->gf_numsrc; ++i) {
 817                                 psin = (struct sockaddr_in *)&gsf->gf_slist[i];
 818
 819                                 if (psin->sin_family != AF_INET)
 820                                         goto mc_msf_out;
 821                                 msf->imsf_slist[i] = psin->sin_addr.s_addr;
 822                         }
 823                         kfree(gsf);
 824                         gsf = NULL;
 825
 826                         err = ip_mc_msfilter(sk, msf, ifindex);
 827 mc_msf_out:
 828                         kfree(msf);
 829                         kfree(gsf);
 830                         break;
 831                 }
 832                 case IP_ROUTER_ALERT:
 833                         err = ip_ra_control(sk, val ? 1 : 0, NULL);
 834                         break;
 835
 836                 case IP_FREEBIND:
 837                         if (optlen<1)
 838                                 goto e_inval;
 839                         inet->freebind = !!val;
 840                         break;
 841
 842                 case IP_IPSEC_POLICY:
 843                 case IP_XFRM_POLICY:
 844                         err = -EPERM;
 845                         if (!capable(CAP_NET_ADMIN))
 846                                 break;
 847                         err = xfrm_user_policy(sk, optname, optval, optlen);
 848                         break;
 849
 850                 default:
 851 #ifdef CONFIG_NETFILTER
 852                         err = nf_setsockopt(sk, PF_INET, optname, optval,
 853                                             optlen);
 854 #else
 855                         err = -ENOPROTOOPT;
 856 #endif
 857                         break;
 858         }
 859         release_sock(sk);
 860         return err;
 861
 862 e_inval:
 863         release_sock(sk);
 864         return -EINVAL;
 865 }
 866
 867 /*
 868  *      Get the options. Note for future reference. The GET of IP options gets the
 869  *      _received_ ones. The set sets the _sent_ ones.
 870  */
 871
 872 int ip_getsockopt(struct sock *sk, int level, int optname, char __user *optval, int __user *optlen)
 873 {
 874         struct inet_sock *inet = inet_sk(sk);
 875         int val;
 876         int len;
 877
 878         if(level!=SOL_IP)
 879                 return -EOPNOTSUPP;
 880
 881 #ifdef CONFIG_IP_MROUTE
 882         if(optname>=MRT_BASE && optname <=MRT_BASE+10)
 883         {
 884                 return ip_mroute_getsockopt(sk,optname,optval,optlen);
 885         }
 886 #endif
 887
 888         if(get_user(len,optlen))
 889                 return -EFAULT;
 890         if(len < 0)
 891                 return -EINVAL;
 892
 893         lock_sock(sk);
 894
 895         switch(optname) {
 896                 case IP_OPTIONS:
 897                         {
 898                                 unsigned char optbuf[sizeof(struct ip_options)+40];
 899                                 struct ip_options * opt = (struct ip_options*)optbuf;
 900                                 opt->optlen = 0;
 901                                 if (inet->opt)
 902                                         memcpy(optbuf, inet->opt,
 903                                                sizeof(struct ip_options)+
 904                                                inet->opt->optlen);
 905                                 release_sock(sk);
 906
 907                                 if (opt->optlen == 0)
 908                                         return put_user(0, optlen);
 909
 910                                 ip_options_undo(opt);
 911
 912                                 len = min_t(unsigned int, len, opt->optlen);
 913                                 if(put_user(len, optlen))
 914                                         return -EFAULT;
 915                                 if(copy_to_user(optval, opt->__data, len))
 916                                         return -EFAULT;
 917                                 return 0;
 918                         }
 919                 case IP_PKTINFO:
 920                         val = (inet->cmsg_flags & IP_CMSG_PKTINFO) != 0;
 921                         break;
 922                 case IP_RECVTTL:
 923                         val = (inet->cmsg_flags & IP_CMSG_TTL) != 0;
 924                         break;
 925                 case IP_RECVTOS:
 926                         val = (inet->cmsg_flags & IP_CMSG_TOS) != 0;
 927                         break;
 928                 case IP_RECVOPTS:
 929                         val = (inet->cmsg_flags & IP_CMSG_RECVOPTS) != 0;
 930                         break;
 931                 case IP_RETOPTS:
 932                         val = (inet->cmsg_flags & IP_CMSG_RETOPTS) != 0;
 933                         break;
 934                 case IP_TOS:
 935                         val = inet->tos;
 936                         break;
 937                 case IP_TTL:
 938                         val = (inet->uc_ttl == -1 ?
 939                                sysctl_ip_default_ttl :
 940                                inet->uc_ttl);
 941                         break;
 942                 case IP_HDRINCL:
 943                         val = inet->hdrincl;
 944                         break;
 945                 case IP_MTU_DISCOVER:
 946                         val = inet->pmtudisc;
 947                         break;
 948                 case IP_MTU:
 949                 {
 950                         struct dst_entry *dst;
 951                         val = 0;
 952                         dst = sk_dst_get(sk);
 953                         if (dst) {
 954                                 val = dst_mtu(dst);
 955                                 dst_release(dst);
 956                         }
 957                         if (!val) {
 958                                 release_sock(sk);
 959                                 return -ENOTCONN;
 960                         }
 961                         break;
 962                 }
 963                 case IP_RECVERR:
 964                         val = inet->recverr;
 965                         break;
 966                 case IP_MULTICAST_TTL:
 967                         val = inet->mc_ttl;
 968                         break;
 969                 case IP_MULTICAST_LOOP:
 970                         val = inet->mc_loop;
 971                         break;
 972                 case IP_MULTICAST_IF:
 973                 {
 974                         struct in_addr addr;
 975                         len = min_t(unsigned int, len, sizeof(struct in_addr));
 976                         addr.s_addr = inet->mc_addr;
 977                         release_sock(sk);
 978
 979                         if(put_user(len, optlen))
 980                                 return -EFAULT;
 981                         if(copy_to_user(optval, &addr, len))
 982                                 return -EFAULT;
 983                         return 0;
 984                 }
 985                 case IP_MSFILTER:
 986                 {
 987                         struct ip_msfilter msf;
 988                         int err;
 989
 990                         if (len < IP_MSFILTER_SIZE(0)) {
 991                                 release_sock(sk);
 992                                 return -EINVAL;
 993                         }
 994                         if (copy_from_user(&msf, optval, IP_MSFILTER_SIZE(0))) {
 995                                 release_sock(sk);
 996                                 return -EFAULT;
 997                         }
 998                         err = ip_mc_msfget(sk, &msf,
 999                                 (struct ip_msfilter __user *)optval, optlen);
1000                         release_sock(sk);
1001                         return err;
1002                 }
1003                 case MCAST_MSFILTER:
1004                 {
1005                         struct group_filter gsf;
1006                         int err;
1007
1008                         if (len < GROUP_FILTER_SIZE(0)) {
1009                                 release_sock(sk);
1010                                 return -EINVAL;
1011                         }
1012                         if (copy_from_user(&gsf, optval, GROUP_FILTER_SIZE(0))) {
1013                                 release_sock(sk);
1014                                 return -EFAULT;
1015                         }
1016                         err = ip_mc_gsfget(sk, &gsf,
1017                                 (struct group_filter __user *)optval, optlen);
1018                         release_sock(sk);
1019                         return err;
1020                 }
1021                 case IP_PKTOPTIONS:
1022                 {
1023                         struct msghdr msg;
1024
1025                         release_sock(sk);
1026
1027                         if (sk->sk_type != SOCK_STREAM)
1028                                 return -ENOPROTOOPT;
1029
1030                         msg.msg_control = optval;
1031                         msg.msg_controllen = len;
1032                         msg.msg_flags = 0;
1033
1034                         if (inet->cmsg_flags & IP_CMSG_PKTINFO) {
1035                                 struct in_pktinfo info;
1036
1037                                 info.ipi_addr.s_addr = inet->rcv_saddr;
1038                                 info.ipi_spec_dst.s_addr = inet->rcv_saddr;
1039                                 info.ipi_ifindex = inet->mc_index;
1040                                 put_cmsg(&msg, SOL_IP, IP_PKTINFO, sizeof(info), &info);
1041                         }
1042                         if (inet->cmsg_flags & IP_CMSG_TTL) {
1043                                 int hlim = inet->mc_ttl;
1044                                 put_cmsg(&msg, SOL_IP, IP_TTL, sizeof(hlim), &hlim);
1045                         }
1046                         len -= msg.msg_controllen;
1047                         return put_user(len, optlen);
1048                 }
1049                 case IP_FREEBIND:
1050                         val = inet->freebind;
1051                         break;
1052                 default:
1053 #ifdef CONFIG_NETFILTER
1054                         val = nf_getsockopt(sk, PF_INET, optname, optval,
1055                                             &len);
1056                         release_sock(sk);
1057                         if (val >= 0)
1058                                 val = put_user(len, optlen);
1059                         return val;
1060 #else
1061                         release_sock(sk);
1062                         return -ENOPROTOOPT;
1063 #endif
1064         }
1065         release_sock(sk);
1066
1067         if (len < sizeof(int) && len > 0 && val>=0 && val<255) {
1068                 unsigned char ucval = (unsigned char)val;
1069                 len = 1;
1070                 if(put_user(len, optlen))
1071                         return -EFAULT;
1072                 if(copy_to_user(optval,&ucval,1))
1073                         return -EFAULT;
1074         } else {
1075                 len = min_t(unsigned int, sizeof(int), len);
1076                 if(put_user(len, optlen))
1077                         return -EFAULT;
1078                 if(copy_to_user(optval,&val,len))
1079                         return -EFAULT;
1080         }
1081         return 0;
1082 }
1083
1084 EXPORT_SYMBOL(ip_cmsg_recv);
1085
1086 EXPORT_SYMBOL(ip_getsockopt);
1087 EXPORT_SYMBOL(ip_setsockopt);