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.h>
  33 #include <linux/tcp.h>
  34 #include <linux/udp.h>
  35 #include <linux/igmp.h>
  36 #include <linux/netfilter.h>
  37 #include <linux/route.h>
  38 #include <linux/mroute.h>
  39 #include <net/route.h>
  40 #include <net/xfrm.h>
  41 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
  42 #include <net/transp_v6.h>
  43 #endif
  44
  45 #include <linux/errqueue.h>
  46 #include <asm/uaccess.h>
  47
  48 #define IP_CMSG_PKTINFO         1
  49 #define IP_CMSG_TTL             2
  50 #define IP_CMSG_TOS             4
  51 #define IP_CMSG_RECVOPTS        8
  52 #define IP_CMSG_RETOPTS         16
  53
  54 /*
  55  *      SOL_IP control messages.
  56  */
  57
  58 static void ip_cmsg_recv_pktinfo(struct msghdr *msg, struct sk_buff *skb)
  59 {
  60         struct in_pktinfo info;
  61         struct rtable *rt = (struct rtable *)skb->dst;
  62
  63         info.ipi_addr.s_addr = skb->nh.iph->daddr;
  64         if (rt) {
  65                 info.ipi_ifindex = rt->rt_iif;
  66                 info.ipi_spec_dst.s_addr = rt->rt_spec_dst;
  67         } else {
  68                 info.ipi_ifindex = 0;
  69                 info.ipi_spec_dst.s_addr = 0;
  70         }
  71
  72         put_cmsg(msg, SOL_IP, IP_PKTINFO, sizeof(info), &info);
  73 }
  74
  75 static void ip_cmsg_recv_ttl(struct msghdr *msg, struct sk_buff *skb)
  76 {
  77         int ttl = skb->nh.iph->ttl;
  78         put_cmsg(msg, SOL_IP, IP_TTL, sizeof(int), &ttl);
  79 }
  80
  81 static void ip_cmsg_recv_tos(struct msghdr *msg, struct sk_buff *skb)
  82 {
  83         put_cmsg(msg, SOL_IP, IP_TOS, 1, &skb->nh.iph->tos);
  84 }
  85
  86 static void ip_cmsg_recv_opts(struct msghdr *msg, struct sk_buff *skb)
  87 {
  88         if (IPCB(skb)->opt.optlen == 0)
  89                 return;
  90
  91         put_cmsg(msg, SOL_IP, IP_RECVOPTS, IPCB(skb)->opt.optlen, skb->nh.iph+1);
  92 }
  93
  94
  95 void ip_cmsg_recv_retopts(struct msghdr *msg, struct sk_buff *skb)
  96 {
  97         unsigned char optbuf[sizeof(struct ip_options) + 40];
  98         struct ip_options * opt = (struct ip_options*)optbuf;
  99
 100         if (IPCB(skb)->opt.optlen == 0)
 101                 return;
 102
 103         if (ip_options_echo(opt, skb)) {
 104                 msg->msg_flags |= MSG_CTRUNC;
 105                 return;
 106         }
 107         ip_options_undo(opt);
 108
 109         put_cmsg(msg, SOL_IP, IP_RETOPTS, opt->optlen, opt->__data);
 110 }
 111
 112
 113 void ip_cmsg_recv(struct msghdr *msg, struct sk_buff *skb)
 114 {
 115         struct inet_opt *inet = inet_sk(skb->sk);
 116         unsigned flags = inet->cmsg_flags;
 117
 118         /* Ordered by supposed usage frequency */
 119         if (flags & 1)
 120                 ip_cmsg_recv_pktinfo(msg, skb);
 121         if ((flags>>=1) == 0)
 122                 return;
 123
 124         if (flags & 1)
 125                 ip_cmsg_recv_ttl(msg, skb);
 126         if ((flags>>=1) == 0)
 127                 return;
 128
 129         if (flags & 1)
 130                 ip_cmsg_recv_tos(msg, skb);
 131         if ((flags>>=1) == 0)
 132                 return;
 133
 134         if (flags & 1)
 135                 ip_cmsg_recv_opts(msg, skb);
 136         if ((flags>>=1) == 0)
 137                 return;
 138
 139         if (flags & 1)
 140                 ip_cmsg_recv_retopts(msg, skb);
 141 }
 142
 143 int ip_cmsg_send(struct msghdr *msg, struct ipcm_cookie *ipc)
 144 {
 145         int err;
 146         struct cmsghdr *cmsg;
 147
 148         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
 149                 if (cmsg->cmsg_len < sizeof(struct cmsghdr) ||
 150                     (unsigned long)(((char*)cmsg - (char*)msg->msg_control)
 151                                     + cmsg->cmsg_len) > msg->msg_controllen) {
 152                         return -EINVAL;
 153                 }
 154                 if (cmsg->cmsg_level != SOL_IP)
 155                         continue;
 156                 switch (cmsg->cmsg_type) {
 157                 case IP_RETOPTS:
 158                         err = cmsg->cmsg_len - CMSG_ALIGN(sizeof(struct cmsghdr));
 159                         err = ip_options_get(&ipc->opt, CMSG_DATA(cmsg), err < 40 ? err : 40, 0);
 160                         if (err)
 161                                 return err;
 162                         break;
 163                 case IP_PKTINFO:
 164                 {
 165                         struct in_pktinfo *info;
 166                         if (cmsg->cmsg_len != CMSG_LEN(sizeof(struct in_pktinfo)))
 167                                 return -EINVAL;
 168                         info = (struct in_pktinfo *)CMSG_DATA(cmsg);
 169                         ipc->oif = info->ipi_ifindex;
 170                         ipc->addr = info->ipi_spec_dst.s_addr;
 171                         break;
 172                 }
 173                 default:
 174                         return -EINVAL;
 175                 }
 176         }
 177         return 0;
 178 }
 179
 180
 181 /* Special input handler for packets caught by router alert option.
 182    They are selected only by protocol field, and then processed likely
 183    local ones; but only if someone wants them! Otherwise, router
 184    not running rsvpd will kill RSVP.
 185
 186    It is user level problem, what it will make with them.
 187    I have no idea, how it will masquearde or NAT them (it is joke, joke :-)),
 188    but receiver should be enough clever f.e. to forward mtrace requests,
 189    sent to multicast group to reach destination designated router.
 190  */
 191 struct ip_ra_chain *ip_ra_chain;
 192 rwlock_t ip_ra_lock = RW_LOCK_UNLOCKED;
 193
 194 int ip_ra_control(struct sock *sk, unsigned char on, void (*destructor)(struct sock *))
 195 {
 196         struct ip_ra_chain *ra, *new_ra, **rap;
 197
 198         if (sk->sk_type != SOCK_RAW || inet_sk(sk)->num == IPPROTO_RAW)
 199                 return -EINVAL;
 200
 201         new_ra = on ? kmalloc(sizeof(*new_ra), GFP_KERNEL) : NULL;
 202
 203         write_lock_bh(&ip_ra_lock);
 204         for (rap = &ip_ra_chain; (ra=*rap) != NULL; rap = &ra->next) {
 205                 if (ra->sk == sk) {
 206                         if (on) {
 207                                 write_unlock_bh(&ip_ra_lock);
 208                                 if (new_ra)
 209                                         kfree(new_ra);
 210                                 return -EADDRINUSE;
 211                         }
 212                         *rap = ra->next;
 213                         write_unlock_bh(&ip_ra_lock);
 214
 215                         if (ra->destructor)
 216                                 ra->destructor(sk);
 217                         sock_put(sk);
 218                         kfree(ra);
 219                         return 0;
 220                 }
 221         }
 222         if (new_ra == NULL) {
 223                 write_unlock_bh(&ip_ra_lock);
 224                 return -ENOBUFS;
 225         }
 226         new_ra->sk = sk;
 227         new_ra->destructor = destructor;
 228
 229         new_ra->next = ra;
 230         *rap = new_ra;
 231         sock_hold(sk);
 232         write_unlock_bh(&ip_ra_lock);
 233
 234         return 0;
 235 }
 236
 237 void ip_icmp_error(struct sock *sk, struct sk_buff *skb, int err,
 238                    u16 port, u32 info, u8 *payload)
 239 {
 240         struct inet_opt *inet = inet_sk(sk);
 241         struct sock_exterr_skb *serr;
 242
 243         if (!inet->recverr)
 244                 return;
 245
 246         skb = skb_clone(skb, GFP_ATOMIC);
 247         if (!skb)
 248                 return;
 249
 250         serr = SKB_EXT_ERR(skb);
 251         serr->ee.ee_errno = err;
 252         serr->ee.ee_origin = SO_EE_ORIGIN_ICMP;
 253         serr->ee.ee_type = skb->h.icmph->type;
 254         serr->ee.ee_code = skb->h.icmph->code;
 255         serr->ee.ee_pad = 0;
 256         serr->ee.ee_info = info;
 257         serr->ee.ee_data = 0;
 258         serr->addr_offset = (u8*)&(((struct iphdr*)(skb->h.icmph+1))->daddr) - skb->nh.raw;
 259         serr->port = port;
 260
 261         skb->h.raw = payload;
 262         if (!skb_pull(skb, payload - skb->data) ||
 263             sock_queue_err_skb(sk, skb))
 264                 kfree_skb(skb);
 265 }
 266
 267 void ip_local_error(struct sock *sk, int err, u32 daddr, u16 port, u32 info)
 268 {
 269         struct inet_opt *inet = inet_sk(sk);
 270         struct sock_exterr_skb *serr;
 271         struct iphdr *iph;
 272         struct sk_buff *skb;
 273
 274         if (!inet->recverr)
 275                 return;
 276
 277         skb = alloc_skb(sizeof(struct iphdr), GFP_ATOMIC);
 278         if (!skb)
 279                 return;
 280
 281         iph = (struct iphdr*)skb_put(skb, sizeof(struct iphdr));
 282         skb->nh.iph = iph;
 283         iph->daddr = daddr;
 284
 285         serr = SKB_EXT_ERR(skb);
 286         serr->ee.ee_errno = err;
 287         serr->ee.ee_origin = SO_EE_ORIGIN_LOCAL;
 288         serr->ee.ee_type = 0;
 289         serr->ee.ee_code = 0;
 290         serr->ee.ee_pad = 0;
 291         serr->ee.ee_info = info;
 292         serr->ee.ee_data = 0;
 293         serr->addr_offset = (u8*)&iph->daddr - skb->nh.raw;
 294         serr->port = port;
 295
 296         skb->h.raw = skb->tail;
 297         __skb_pull(skb, skb->tail - skb->data);
 298
 299         if (sock_queue_err_skb(sk, skb))
 300                 kfree_skb(skb);
 301 }
 302
 303 /*
 304  *      Handle MSG_ERRQUEUE
 305  */
 306 int ip_recv_error(struct sock *sk, struct msghdr *msg, int len)
 307 {
 308         struct sock_exterr_skb *serr;
 309         struct sk_buff *skb, *skb2;
 310         struct sockaddr_in *sin;
 311         struct {
 312                 struct sock_extended_err ee;
 313                 struct sockaddr_in       offender;
 314         } errhdr;
 315         int err;
 316         int copied;
 317
 318         err = -EAGAIN;
 319         skb = skb_dequeue(&sk->sk_error_queue);
 320         if (skb == NULL)
 321                 goto out;
 322
 323         copied = skb->len;
 324         if (copied > len) {
 325                 msg->msg_flags |= MSG_TRUNC;
 326                 copied = len;
 327         }
 328         err = skb_copy_datagram_iovec(skb, 0, msg->msg_iov, copied);
 329         if (err)
 330                 goto out_free_skb;
 331
 332         sock_recv_timestamp(msg, sk, skb);
 333
 334         serr = SKB_EXT_ERR(skb);
 335
 336         sin = (struct sockaddr_in *)msg->msg_name;
 337         if (sin) {
 338                 sin->sin_family = AF_INET;
 339                 sin->sin_addr.s_addr = *(u32*)(skb->nh.raw + serr->addr_offset);
 340                 sin->sin_port = serr->port;
 341                 memset(&sin->sin_zero, 0, sizeof(sin->sin_zero));
 342         }
 343
 344         memcpy(&errhdr.ee, &serr->ee, sizeof(struct sock_extended_err));
 345         sin = &errhdr.offender;
 346         sin->sin_family = AF_UNSPEC;
 347         if (serr->ee.ee_origin == SO_EE_ORIGIN_ICMP) {
 348                 struct inet_opt *inet = inet_sk(sk);
 349
 350                 sin->sin_family = AF_INET;
 351                 sin->sin_addr.s_addr = skb->nh.iph->saddr;
 352                 sin->sin_port = 0;
 353                 memset(&sin->sin_zero, 0, sizeof(sin->sin_zero));
 354                 if (inet->cmsg_flags)
 355                         ip_cmsg_recv(msg, skb);
 356         }
 357
 358         put_cmsg(msg, SOL_IP, IP_RECVERR, sizeof(errhdr), &errhdr);
 359
 360         /* Now we could try to dump offended packet options */
 361
 362         msg->msg_flags |= MSG_ERRQUEUE;
 363         err = copied;
 364
 365         /* Reset and regenerate socket error */
 366         spin_lock_irq(&sk->sk_error_queue.lock);
 367         sk->sk_err = 0;
 368         if ((skb2 = skb_peek(&sk->sk_error_queue)) != NULL) {
 369                 sk->sk_err = SKB_EXT_ERR(skb2)->ee.ee_errno;
 370                 spin_unlock_irq(&sk->sk_error_queue.lock);
 371                 sk->sk_error_report(sk);
 372         } else
 373                 spin_unlock_irq(&sk->sk_error_queue.lock);
 374
 375 out_free_skb:
 376         kfree_skb(skb);
 377 out:
 378         return err;
 379 }
 380
 381
 382 /*
 383  *      Socket option code for IP. This is the end of the line after any TCP,UDP etc options on
 384  *      an IP socket.
 385  */
 386
 387 int ip_setsockopt(struct sock *sk, int level, int optname, char __user *optval, int optlen)
 388 {
 389         struct inet_opt *inet = inet_sk(sk);
 390         int val=0,err;
 391
 392         if (level != SOL_IP)
 393                 return -ENOPROTOOPT;
 394
 395         if (((1<<optname) & ((1<<IP_PKTINFO) | (1<<IP_RECVTTL) |
 396                             (1<<IP_RECVOPTS) | (1<<IP_RECVTOS) |
 397                             (1<<IP_RETOPTS) | (1<<IP_TOS) |
 398                             (1<<IP_TTL) | (1<<IP_HDRINCL) |
 399                             (1<<IP_MTU_DISCOVER) | (1<<IP_RECVERR) |
 400                             (1<<IP_ROUTER_ALERT) | (1<<IP_FREEBIND))) ||
 401                                 optname == IP_MULTICAST_TTL ||
 402                                 optname == IP_MULTICAST_LOOP) {
 403                 if (optlen >= sizeof(int)) {
 404                         if (get_user(val, (int __user *) optval))
 405                                 return -EFAULT;
 406                 } else if (optlen >= sizeof(char)) {
 407                         unsigned char ucval;
 408
 409                         if (get_user(ucval, (unsigned char __user *) optval))
 410                                 return -EFAULT;
 411                         val = (int) ucval;
 412                 }
 413         }
 414
 415         /* If optlen==0, it is equivalent to val == 0 */
 416
 417 #ifdef CONFIG_IP_MROUTE
 418         if (optname >= MRT_BASE && optname <= (MRT_BASE + 10))
 419                 return ip_mroute_setsockopt(sk,optname,optval,optlen);
 420 #endif
 421
 422         err = 0;
 423         lock_sock(sk);
 424
 425         switch (optname) {
 426                 case IP_OPTIONS:
 427                 {
 428                         struct ip_options * opt = NULL;
 429                         if (optlen > 40 || optlen < 0)
 430                                 goto e_inval;
 431                         err = ip_options_get(&opt, optval, optlen, 1);
 432                         if (err)
 433                                 break;
 434                         if (sk->sk_type == SOCK_STREAM) {
 435                                 struct tcp_opt *tp = tcp_sk(sk);
 436 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 437                                 if (sk->sk_family == PF_INET ||
 438                                     (!((1 << sk->sk_state) &
 439                                        (TCPF_LISTEN | TCPF_CLOSE)) &&
 440                                      inet->daddr != LOOPBACK4_IPV6)) {
 441 #endif
 442                                         if (inet->opt)
 443                                                 tp->ext_header_len -= inet->opt->optlen;
 444                                         if (opt)
 445                                                 tp->ext_header_len += opt->optlen;
 446                                         tcp_sync_mss(sk, tp->pmtu_cookie);
 447 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 448                                 }
 449 #endif
 450                         }
 451                         opt = xchg(&inet->opt, opt);
 452                         if (opt)
 453                                 kfree(opt);
 454                         break;
 455                 }
 456                 case IP_PKTINFO:
 457                         if (val)
 458                                 inet->cmsg_flags |= IP_CMSG_PKTINFO;
 459                         else
 460                                 inet->cmsg_flags &= ~IP_CMSG_PKTINFO;
 461                         break;
 462                 case IP_RECVTTL:
 463                         if (val)
 464                                 inet->cmsg_flags |=  IP_CMSG_TTL;
 465                         else
 466                                 inet->cmsg_flags &= ~IP_CMSG_TTL;
 467                         break;
 468                 case IP_RECVTOS:
 469                         if (val)
 470                                 inet->cmsg_flags |=  IP_CMSG_TOS;
 471                         else
 472                                 inet->cmsg_flags &= ~IP_CMSG_TOS;
 473                         break;
 474                 case IP_RECVOPTS:
 475                         if (val)
 476                                 inet->cmsg_flags |=  IP_CMSG_RECVOPTS;
 477                         else
 478                                 inet->cmsg_flags &= ~IP_CMSG_RECVOPTS;
 479                         break;
 480                 case IP_RETOPTS:
 481                         if (val)
 482                                 inet->cmsg_flags |= IP_CMSG_RETOPTS;
 483                         else
 484                                 inet->cmsg_flags &= ~IP_CMSG_RETOPTS;
 485                         break;
 486                 case IP_TOS:    /* This sets both TOS and Precedence */
 487                         if (sk->sk_type == SOCK_STREAM) {
 488                                 val &= ~3;
 489                                 val |= inet->tos & 3;
 490                         }
 491                         if (IPTOS_PREC(val) >= IPTOS_PREC_CRITIC_ECP &&
 492                             !capable(CAP_NET_ADMIN)) {
 493                                 err = -EPERM;
 494                                 break;
 495                         }
 496                         if (inet->tos != val) {
 497                                 inet->tos = val;
 498                                 sk->sk_priority = rt_tos2priority(val);
 499                                 sk_dst_reset(sk);
 500                         }
 501                         break;
 502                 case IP_TTL:
 503                         if (optlen<1)
 504                                 goto e_inval;
 505                         if (val != -1 && (val < 1 || val>255))
 506                                 goto e_inval;
 507                         inet->uc_ttl = val;
 508                         break;
 509                 case IP_HDRINCL:
 510                         if (sk->sk_type != SOCK_RAW) {
 511                                 err = -ENOPROTOOPT;
 512                                 break;
 513                         }
 514                         inet->hdrincl = val ? 1 : 0;
 515                         break;
 516                 case IP_MTU_DISCOVER:
 517                         if (val<0 || val>2)
 518                                 goto e_inval;
 519                         inet->pmtudisc = val;
 520                         break;
 521                 case IP_RECVERR:
 522                         inet->recverr = !!val;
 523                         if (!val)
 524                                 skb_queue_purge(&sk->sk_error_queue);
 525                         break;
 526                 case IP_MULTICAST_TTL:
 527                         if (sk->sk_type == SOCK_STREAM)
 528                                 goto e_inval;
 529                         if (optlen<1)
 530                                 goto e_inval;
 531                         if (val==-1)
 532                                 val = 1;
 533                         if (val < 0 || val > 255)
 534                                 goto e_inval;
 535                         inet->mc_ttl = val;
 536                         break;
 537                 case IP_MULTICAST_LOOP:
 538                         if (optlen<1)
 539                                 goto e_inval;
 540                         inet->mc_loop = !!val;
 541                         break;
 542                 case IP_MULTICAST_IF:
 543                 {
 544                         struct ip_mreqn mreq;
 545                         struct net_device *dev = NULL;
 546
 547                         if (sk->sk_type == SOCK_STREAM)
 548                                 goto e_inval;
 549                         /*
 550                          *      Check the arguments are allowable
 551                          */
 552
 553                         err = -EFAULT;
 554                         if (optlen >= sizeof(struct ip_mreqn)) {
 555                                 if (copy_from_user(&mreq,optval,sizeof(mreq)))
 556                                         break;
 557                         } else {
 558                                 memset(&mreq, 0, sizeof(mreq));
 559                                 if (optlen >= sizeof(struct in_addr) &&
 560                                     copy_from_user(&mreq.imr_address,optval,sizeof(struct in_addr)))
 561                                         break;
 562                         }
 563
 564                         if (!mreq.imr_ifindex) {
 565                                 if (mreq.imr_address.s_addr == INADDR_ANY) {
 566                                         inet->mc_index = 0;
 567                                         inet->mc_addr  = 0;
 568                                         err = 0;
 569                                         break;
 570                                 }
 571                                 dev = ip_dev_find(mreq.imr_address.s_addr);
 572                                 if (dev) {
 573                                         mreq.imr_ifindex = dev->ifindex;
 574                                         dev_put(dev);
 575                                 }
 576                         } else
 577                                 dev = __dev_get_by_index(mreq.imr_ifindex);
 578
 579
 580                         err = -EADDRNOTAVAIL;
 581                         if (!dev)
 582                                 break;
 583
 584                         err = -EINVAL;
 585                         if (sk->sk_bound_dev_if &&
 586                             mreq.imr_ifindex != sk->sk_bound_dev_if)
 587                                 break;
 588
 589                         inet->mc_index = mreq.imr_ifindex;
 590                         inet->mc_addr  = mreq.imr_address.s_addr;
 591                         err = 0;
 592                         break;
 593                 }
 594
 595                 case IP_ADD_MEMBERSHIP:
 596                 case IP_DROP_MEMBERSHIP:
 597                 {
 598                         struct ip_mreqn mreq;
 599
 600                         if (optlen < sizeof(struct ip_mreq))
 601                                 goto e_inval;
 602                         err = -EFAULT;
 603                         if (optlen >= sizeof(struct ip_mreqn)) {
 604                                 if(copy_from_user(&mreq,optval,sizeof(mreq)))
 605                                         break;
 606                         } else {
 607                                 memset(&mreq, 0, sizeof(mreq));
 608                                 if (copy_from_user(&mreq,optval,sizeof(struct ip_mreq)))
 609                                         break;
 610                         }
 611
 612                         if (optname == IP_ADD_MEMBERSHIP)
 613                                 err = ip_mc_join_group(sk, &mreq);
 614                         else
 615                                 err = ip_mc_leave_group(sk, &mreq);
 616                         break;
 617                 }
 618                 case IP_MSFILTER:
 619                 {
 620                         extern int sysctl_optmem_max;
 621                         extern int sysctl_igmp_max_msf;
 622                         struct ip_msfilter *msf;
 623
 624                         if (optlen < IP_MSFILTER_SIZE(0))
 625                                 goto e_inval;
 626                         if (optlen > sysctl_optmem_max) {
 627                                 err = -ENOBUFS;
 628                                 break;
 629                         }
 630                         msf = (struct ip_msfilter *)kmalloc(optlen, GFP_KERNEL);
 631                         if (msf == 0) {
 632                                 err = -ENOBUFS;
 633                                 break;
 634                         }
 635                         err = -EFAULT;
 636                         if (copy_from_user(msf, optval, optlen)) {
 637                                 kfree(msf);
 638                                 break;
 639                         }
 640                         /* numsrc >= (1G-4) overflow in 32 bits */
 641                         if (msf->imsf_numsrc >= 0x3ffffffcU ||
 642                             msf->imsf_numsrc > sysctl_igmp_max_msf) {
 643                                 kfree(msf);
 644                                 err = -ENOBUFS;
 645                                 break;
 646                         }
 647                         if (IP_MSFILTER_SIZE(msf->imsf_numsrc) > optlen) {
 648                                 kfree(msf);
 649                                 err = -EINVAL;
 650                                 break;
 651                         }
 652                         err = ip_mc_msfilter(sk, msf, 0);
 653                         kfree(msf);
 654                         break;
 655                 }
 656                 case IP_BLOCK_SOURCE:
 657                 case IP_UNBLOCK_SOURCE:
 658                 case IP_ADD_SOURCE_MEMBERSHIP:
 659                 case IP_DROP_SOURCE_MEMBERSHIP:
 660                 {
 661                         struct ip_mreq_source mreqs;
 662                         int omode, add;
 663
 664                         if (optlen != sizeof(struct ip_mreq_source))
 665                                 goto e_inval;
 666                         if (copy_from_user(&mreqs, optval, sizeof(mreqs))) {
 667                                 err = -EFAULT;
 668                                 break;
 669                         }
 670                         if (optname == IP_BLOCK_SOURCE) {
 671                                 omode = MCAST_EXCLUDE;
 672                                 add = 1;
 673                         } else if (optname == IP_UNBLOCK_SOURCE) {
 674                                 omode = MCAST_EXCLUDE;
 675                                 add = 0;
 676                         } else if (optname == IP_ADD_SOURCE_MEMBERSHIP) {
 677                                 struct ip_mreqn mreq;
 678
 679                                 mreq.imr_multiaddr.s_addr = mreqs.imr_multiaddr;
 680                                 mreq.imr_address.s_addr = mreqs.imr_interface;
 681                                 mreq.imr_ifindex = 0;
 682                                 err = ip_mc_join_group(sk, &mreq);
 683                                 if (err)
 684                                         break;
 685                                 omode = MCAST_INCLUDE;
 686                                 add = 1;
 687                         } else /*IP_DROP_SOURCE_MEMBERSHIP */ {
 688                                 omode = MCAST_INCLUDE;
 689                                 add = 0;
 690                         }
 691                         err = ip_mc_source(add, omode, sk, &mreqs, 0);
 692                         break;
 693                 }
 694                 case MCAST_JOIN_GROUP:
 695                 case MCAST_LEAVE_GROUP:
 696                 {
 697                         struct group_req greq;
 698                         struct sockaddr_in *psin;
 699                         struct ip_mreqn mreq;
 700
 701                         if (optlen < sizeof(struct group_req))
 702                                 goto e_inval;
 703                         err = -EFAULT;
 704                         if(copy_from_user(&greq, optval, sizeof(greq)))
 705                                 break;
 706                         psin = (struct sockaddr_in *)&greq.gr_group;
 707                         if (psin->sin_family != AF_INET)
 708                                 goto e_inval;
 709                         memset(&mreq, 0, sizeof(mreq));
 710                         mreq.imr_multiaddr = psin->sin_addr;
 711                         mreq.imr_ifindex = greq.gr_interface;
 712
 713                         if (optname == MCAST_JOIN_GROUP)
 714                                 err = ip_mc_join_group(sk, &mreq);
 715                         else
 716                                 err = ip_mc_leave_group(sk, &mreq);
 717                         break;
 718                 }
 719                 case MCAST_JOIN_SOURCE_GROUP:
 720                 case MCAST_LEAVE_SOURCE_GROUP:
 721                 case MCAST_BLOCK_SOURCE:
 722                 case MCAST_UNBLOCK_SOURCE:
 723                 {
 724                         struct group_source_req greqs;
 725                         struct ip_mreq_source mreqs;
 726                         struct sockaddr_in *psin;
 727                         int omode, add;
 728
 729                         if (optlen != sizeof(struct group_source_req))
 730                                 goto e_inval;
 731                         if (copy_from_user(&greqs, optval, sizeof(greqs))) {
 732                                 err = -EFAULT;
 733                                 break;
 734                         }
 735                         if (greqs.gsr_group.ss_family != AF_INET ||
 736                             greqs.gsr_source.ss_family != AF_INET) {
 737                                 err = -EADDRNOTAVAIL;
 738                                 break;
 739                         }
 740                         psin = (struct sockaddr_in *)&greqs.gsr_group;
 741                         mreqs.imr_multiaddr = psin->sin_addr.s_addr;
 742                         psin = (struct sockaddr_in *)&greqs.gsr_source;
 743                         mreqs.imr_sourceaddr = psin->sin_addr.s_addr;
 744                         mreqs.imr_interface = 0; /* use index for mc_source */
 745
 746                         if (optname == MCAST_BLOCK_SOURCE) {
 747                                 omode = MCAST_EXCLUDE;
 748                                 add = 1;
 749                         } else if (optname == MCAST_UNBLOCK_SOURCE) {
 750                                 omode = MCAST_EXCLUDE;
 751                                 add = 0;
 752                         } else if (optname == MCAST_JOIN_SOURCE_GROUP) {
 753                                 struct ip_mreqn mreq;
 754
 755                                 psin = (struct sockaddr_in *)&greqs.gsr_group;
 756                                 mreq.imr_multiaddr = psin->sin_addr;
 757                                 mreq.imr_address.s_addr = 0;
 758                                 mreq.imr_ifindex = greqs.gsr_interface;
 759                                 err = ip_mc_join_group(sk, &mreq);
 760                                 if (err)
 761                                         break;
 762                                 omode = MCAST_INCLUDE;
 763                                 add = 1;
 764                         } else /* MCAST_LEAVE_SOURCE_GROUP */ {
 765                                 omode = MCAST_INCLUDE;
 766                                 add = 0;
 767                         }
 768                         err = ip_mc_source(add, omode, sk, &mreqs,
 769                                 greqs.gsr_interface);
 770                         break;
 771                 }
 772                 case MCAST_MSFILTER:
 773                 {
 774                         extern int sysctl_optmem_max;
 775                         extern int sysctl_igmp_max_msf;
 776                         struct sockaddr_in *psin;
 777                         struct ip_msfilter *msf = 0;
 778                         struct group_filter *gsf = 0;
 779                         int msize, i, ifindex;
 780
 781                         if (optlen < GROUP_FILTER_SIZE(0))
 782                                 goto e_inval;
 783                         if (optlen > sysctl_optmem_max) {
 784                                 err = -ENOBUFS;
 785                                 break;
 786                         }
 787                         gsf = (struct group_filter *)kmalloc(optlen,GFP_KERNEL);
 788                         if (gsf == 0) {
 789                                 err = -ENOBUFS;
 790                                 break;
 791                         }
 792                         err = -EFAULT;
 793                         if (copy_from_user(gsf, optval, optlen)) {
 794                                 goto mc_msf_out;
 795                         }
 796                         /* numsrc >= (4G-140)/128 overflow in 32 bits */
 797                         if (gsf->gf_numsrc >= 0x1ffffff ||
 798                             gsf->gf_numsrc > sysctl_igmp_max_msf) {
 799                                 err = -ENOBUFS;
 800                                 goto mc_msf_out;
 801                         }
 802                         if (GROUP_FILTER_SIZE(gsf->gf_numsrc) > optlen) {
 803                                 err = -EINVAL;
 804                                 goto mc_msf_out;
 805                         }
 806                         msize = IP_MSFILTER_SIZE(gsf->gf_numsrc);
 807                         msf = (struct ip_msfilter *)kmalloc(msize,GFP_KERNEL);
 808                         if (msf == 0) {
 809                                 err = -ENOBUFS;
 810                                 goto mc_msf_out;
 811                         }
 812                         ifindex = gsf->gf_interface;
 813                         psin = (struct sockaddr_in *)&gsf->gf_group;
 814                         if (psin->sin_family != AF_INET) {
 815                                 err = -EADDRNOTAVAIL;
 816                                 goto mc_msf_out;
 817                         }
 818                         msf->imsf_multiaddr = psin->sin_addr.s_addr;
 819                         msf->imsf_interface = 0;
 820                         msf->imsf_fmode = gsf->gf_fmode;
 821                         msf->imsf_numsrc = gsf->gf_numsrc;
 822                         err = -EADDRNOTAVAIL;
 823                         for (i=0; i<gsf->gf_numsrc; ++i) {
 824                                 psin = (struct sockaddr_in *)&gsf->gf_slist[i];
 825
 826                                 if (psin->sin_family != AF_INET)
 827                                         goto mc_msf_out;
 828                                 msf->imsf_slist[i] = psin->sin_addr.s_addr;
 829                         }
 830                         kfree(gsf);
 831                         gsf = 0;
 832
 833                         err = ip_mc_msfilter(sk, msf, ifindex);
 834 mc_msf_out:
 835                         if (msf)
 836                                 kfree(msf);
 837                         if (gsf)
 838                                 kfree(gsf);
 839                         break;
 840                 }
 841                 case IP_ROUTER_ALERT:
 842                         err = ip_ra_control(sk, val ? 1 : 0, NULL);
 843                         break;
 844
 845                 case IP_FREEBIND:
 846                         if (optlen<1)
 847                                 goto e_inval;
 848                         inet->freebind = !!val;
 849                         break;
 850
 851                 case IP_IPSEC_POLICY:
 852                 case IP_XFRM_POLICY:
 853                         err = xfrm_user_policy(sk, optname, optval, optlen);
 854                         break;
 855
 856                 default:
 857 #ifdef CONFIG_NETFILTER
 858                         err = nf_setsockopt(sk, PF_INET, optname, optval,
 859                                             optlen);
 860 #else
 861                         err = -ENOPROTOOPT;
 862 #endif
 863                         break;
 864         }
 865         release_sock(sk);
 866         return err;
 867
 868 e_inval:
 869         release_sock(sk);
 870         return -EINVAL;
 871 }
 872
 873 /*
 874  *      Get the options. Note for future reference. The GET of IP options gets the
 875  *      _received_ ones. The set sets the _sent_ ones.
 876  */
 877
 878 int ip_getsockopt(struct sock *sk, int level, int optname, char __user *optval, int __user *optlen)
 879 {
 880         struct inet_opt *inet = inet_sk(sk);
 881         int val;
 882         int len;
 883
 884         if(level!=SOL_IP)
 885                 return -EOPNOTSUPP;
 886
 887 #ifdef CONFIG_IP_MROUTE
 888         if(optname>=MRT_BASE && optname <=MRT_BASE+10)
 889         {
 890                 return ip_mroute_getsockopt(sk,optname,optval,optlen);
 891         }
 892 #endif
 893
 894         if(get_user(len,optlen))
 895                 return -EFAULT;
 896         if(len < 0)
 897                 return -EINVAL;
 898
 899         lock_sock(sk);
 900
 901         switch(optname) {
 902                 case IP_OPTIONS:
 903                         {
 904                                 unsigned char optbuf[sizeof(struct ip_options)+40];
 905                                 struct ip_options * opt = (struct ip_options*)optbuf;
 906                                 opt->optlen = 0;
 907                                 if (inet->opt)
 908                                         memcpy(optbuf, inet->opt,
 909                                                sizeof(struct ip_options)+
 910                                                inet->opt->optlen);
 911                                 release_sock(sk);
 912
 913                                 if (opt->optlen == 0)
 914                                         return put_user(0, optlen);
 915
 916                                 ip_options_undo(opt);
 917
 918                                 len = min_t(unsigned int, len, opt->optlen);
 919                                 if(put_user(len, optlen))
 920                                         return -EFAULT;
 921                                 if(copy_to_user(optval, opt->__data, len))
 922                                         return -EFAULT;
 923                                 return 0;
 924                         }
 925                 case IP_PKTINFO:
 926                         val = (inet->cmsg_flags & IP_CMSG_PKTINFO) != 0;
 927                         break;
 928                 case IP_RECVTTL:
 929                         val = (inet->cmsg_flags & IP_CMSG_TTL) != 0;
 930                         break;
 931                 case IP_RECVTOS:
 932                         val = (inet->cmsg_flags & IP_CMSG_TOS) != 0;
 933                         break;
 934                 case IP_RECVOPTS:
 935                         val = (inet->cmsg_flags & IP_CMSG_RECVOPTS) != 0;
 936                         break;
 937                 case IP_RETOPTS:
 938                         val = (inet->cmsg_flags & IP_CMSG_RETOPTS) != 0;
 939                         break;
 940                 case IP_TOS:
 941                         val = inet->tos;
 942                         break;
 943                 case IP_TTL:
 944                         val = (inet->uc_ttl == -1 ?
 945                                sysctl_ip_default_ttl :
 946                                inet->uc_ttl);
 947                         break;
 948                 case IP_HDRINCL:
 949                         val = inet->hdrincl;
 950                         break;
 951                 case IP_MTU_DISCOVER:
 952                         val = inet->pmtudisc;
 953                         break;
 954                 case IP_MTU:
 955                 {
 956                         struct dst_entry *dst;
 957                         val = 0;
 958                         dst = sk_dst_get(sk);
 959                         if (dst) {
 960                                 val = dst_pmtu(dst) - dst->header_len;
 961                                 dst_release(dst);
 962                         }
 963                         if (!val) {
 964                                 release_sock(sk);
 965                                 return -ENOTCONN;
 966                         }
 967                         break;
 968                 }
 969                 case IP_RECVERR:
 970                         val = inet->recverr;
 971                         break;
 972                 case IP_MULTICAST_TTL:
 973                         val = inet->mc_ttl;
 974                         break;
 975                 case IP_MULTICAST_LOOP:
 976                         val = inet->mc_loop;
 977                         break;
 978                 case IP_MULTICAST_IF:
 979                 {
 980                         struct in_addr addr;
 981                         len = min_t(unsigned int, len, sizeof(struct in_addr));
 982                         addr.s_addr = inet->mc_addr;
 983                         release_sock(sk);
 984
 985                         if(put_user(len, optlen))
 986                                 return -EFAULT;
 987                         if(copy_to_user(optval, &addr, len))
 988                                 return -EFAULT;
 989                         return 0;
 990                 }
 991                 case IP_MSFILTER:
 992                 {
 993                         struct ip_msfilter msf;
 994                         int err;
 995
 996                         if (len < IP_MSFILTER_SIZE(0)) {
 997                                 release_sock(sk);
 998                                 return -EINVAL;
 999                         }
1000                         if (copy_from_user(&msf, optval, IP_MSFILTER_SIZE(0))) {
1001                                 release_sock(sk);
1002                                 return -EFAULT;
1003                         }
1004                         err = ip_mc_msfget(sk, &msf,
1005                                 (struct ip_msfilter __user *)optval, optlen);
1006                         release_sock(sk);
1007                         return err;
1008                 }
1009                 case MCAST_MSFILTER:
1010                 {
1011                         struct group_filter gsf;
1012                         int err;
1013
1014                         if (len < GROUP_FILTER_SIZE(0)) {
1015                                 release_sock(sk);
1016                                 return -EINVAL;
1017                         }
1018                         if (copy_from_user(&gsf, optval, GROUP_FILTER_SIZE(0))) {
1019                                 release_sock(sk);
1020                                 return -EFAULT;
1021                         }
1022                         err = ip_mc_gsfget(sk, &gsf,
1023                                 (struct group_filter __user *)optval, optlen);
1024                         release_sock(sk);
1025                         return err;
1026                 }
1027                 case IP_PKTOPTIONS:
1028                 {
1029                         struct msghdr msg;
1030
1031                         release_sock(sk);
1032
1033                         if (sk->sk_type != SOCK_STREAM)
1034                                 return -ENOPROTOOPT;
1035
1036                         msg.msg_control = optval;
1037                         msg.msg_controllen = len;
1038                         msg.msg_flags = 0;
1039
1040                         if (inet->cmsg_flags & IP_CMSG_PKTINFO) {
1041                                 struct in_pktinfo info;
1042
1043                                 info.ipi_addr.s_addr = inet->rcv_saddr;
1044                                 info.ipi_spec_dst.s_addr = inet->rcv_saddr;
1045                                 info.ipi_ifindex = inet->mc_index;
1046                                 put_cmsg(&msg, SOL_IP, IP_PKTINFO, sizeof(info), &info);
1047                         }
1048                         if (inet->cmsg_flags & IP_CMSG_TTL) {
1049                                 int hlim = inet->mc_ttl;
1050                                 put_cmsg(&msg, SOL_IP, IP_TTL, sizeof(hlim), &hlim);
1051                         }
1052                         len -= msg.msg_controllen;
1053                         return put_user(len, optlen);
1054                 }
1055                 case IP_FREEBIND:
1056                         val = inet->freebind;
1057                         break;
1058                 default:
1059 #ifdef CONFIG_NETFILTER
1060                         val = nf_getsockopt(sk, PF_INET, optname, optval,
1061                                             &len);
1062                         release_sock(sk);
1063                         if (val >= 0)
1064                                 val = put_user(len, optlen);
1065                         return val;
1066 #else
1067                         release_sock(sk);
1068                         return -ENOPROTOOPT;
1069 #endif
1070         }
1071         release_sock(sk);
1072
1073         if (len < sizeof(int) && len > 0 && val>=0 && val<255) {
1074                 unsigned char ucval = (unsigned char)val;
1075                 len = 1;
1076                 if(put_user(len, optlen))
1077                         return -EFAULT;
1078                 if(copy_to_user(optval,&ucval,1))
1079                         return -EFAULT;
1080         } else {
1081                 len = min_t(unsigned int, sizeof(int), len);
1082                 if(put_user(len, optlen))
1083                         return -EFAULT;
1084                 if(copy_to_user(optval,&val,len))
1085                         return -EFAULT;
1086         }
1087         return 0;
1088 }
1089
1090 EXPORT_SYMBOL(ip_cmsg_recv);
1091
1092 #ifdef CONFIG_IP_SCTP_MODULE
1093 EXPORT_SYMBOL(ip_getsockopt);
1094 EXPORT_SYMBOL(ip_setsockopt);
1095 #endif