1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * RAW - implementation of IP "raw" sockets. 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * 12 * Fixes: 13 * Alan Cox : verify_area() fixed up 14 * Alan Cox : ICMP error handling 15 * Alan Cox : EMSGSIZE if you send too big a packet 16 * Alan Cox : Now uses generic datagrams and shared 17 * skbuff library. No more peek crashes, 18 * no more backlogs 19 * Alan Cox : Checks sk->broadcast. 20 * Alan Cox : Uses skb_free_datagram/skb_copy_datagram 21 * Alan Cox : Raw passes ip options too 22 * Alan Cox : Setsocketopt added 23 * Alan Cox : Fixed error return for broadcasts 24 * Alan Cox : Removed wake_up calls 25 * Alan Cox : Use ttl/tos 26 * Alan Cox : Cleaned up old debugging 27 * Alan Cox : Use new kernel side addresses 28 * Arnt Gulbrandsen : Fixed MSG_DONTROUTE in raw sockets. 29 * Alan Cox : BSD style RAW socket demultiplexing. 30 * Alan Cox : Beginnings of mrouted support. 31 * Alan Cox : Added IP_HDRINCL option. 32 * Alan Cox : Skip broadcast check if BSDism set. 33 * David S. Miller : New socket lookup architecture. 34 */ 35 36 #include <linux/types.h> 37 #include <linux/atomic.h> 38 #include <asm/byteorder.h> 39 #include <asm/current.h> 40 #include <linux/uaccess.h> 41 #include <asm/ioctls.h> 42 #include <linux/stddef.h> 43 #include <linux/slab.h> 44 #include <linux/errno.h> 45 #include <linux/kernel.h> 46 #include <linux/export.h> 47 #include <linux/spinlock.h> 48 #include <linux/sockios.h> 49 #include <linux/socket.h> 50 #include <linux/in.h> 51 #include <linux/mroute.h> 52 #include <linux/netdevice.h> 53 #include <linux/in_route.h> 54 #include <linux/route.h> 55 #include <linux/skbuff.h> 56 #include <linux/igmp.h> 57 #include <net/net_namespace.h> 58 #include <net/dst.h> 59 #include <net/sock.h> 60 #include <linux/ip.h> 61 #include <linux/net.h> 62 #include <net/ip.h> 63 #include <net/icmp.h> 64 #include <net/udp.h> 65 #include <net/raw.h> 66 #include <net/snmp.h> 67 #include <net/tcp_states.h> 68 #include <net/inet_common.h> 69 #include <net/checksum.h> 70 #include <net/xfrm.h> 71 #include <linux/rtnetlink.h> 72 #include <linux/proc_fs.h> 73 #include <linux/seq_file.h> 74 #include <linux/netfilter.h> 75 #include <linux/netfilter_ipv4.h> 76 #include <linux/compat.h> 77 #include <linux/uio.h> 78 79 struct raw_frag_vec { 80 struct msghdr *msg; 81 union { 82 struct icmphdr icmph; 83 char c[1]; 84 } hdr; 85 int hlen; 86 }; 87 88 struct raw_hashinfo raw_v4_hashinfo; 89 EXPORT_SYMBOL_GPL(raw_v4_hashinfo); 90 91 int raw_hash_sk(struct sock *sk) 92 { 93 struct raw_hashinfo *h = sk->sk_prot->h.raw_hash; 94 struct hlist_head *hlist; 95 96 hlist = &h->ht[raw_hashfunc(sock_net(sk), inet_sk(sk)->inet_num)]; 97 98 spin_lock(&h->lock); 99 sk_add_node_rcu(sk, hlist); 100 sock_set_flag(sk, SOCK_RCU_FREE); 101 spin_unlock(&h->lock); 102 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 103 104 return 0; 105 } 106 107 void raw_unhash_sk(struct sock *sk) 108 { 109 struct raw_hashinfo *h = sk->sk_prot->h.raw_hash; 110 111 spin_lock(&h->lock); 112 if (sk_del_node_init_rcu(sk)) 113 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 114 spin_unlock(&h->lock); 115 } 116 117 bool raw_v4_match(struct net *net, const struct sock *sk, unsigned short num, 118 __be32 raddr, __be32 laddr, int dif, int sdif) 119 { 120 const struct inet_sock *inet = inet_sk(sk); 121 __be32 daddr, rcv_saddr; 122 123 if (!net_eq(sock_net(sk), net) || inet->inet_num != num) 124 return false; 125 126 daddr = READ_ONCE(inet->inet_daddr); 127 if (daddr && daddr != raddr) 128 return false; 129 130 rcv_saddr = READ_ONCE(inet->inet_rcv_saddr); 131 if (rcv_saddr && rcv_saddr != laddr) 132 return false; 133 134 return raw_sk_bound_dev_eq(net, READ_ONCE(sk->sk_bound_dev_if), 135 dif, sdif); 136 } 137 EXPORT_SYMBOL_GPL(raw_v4_match); 138 139 /* 140 * 0 - deliver 141 * 1 - block 142 */ 143 static int icmp_filter(const struct sock *sk, const struct sk_buff *skb) 144 { 145 struct icmphdr _hdr; 146 const struct icmphdr *hdr; 147 148 hdr = skb_header_pointer(skb, skb_transport_offset(skb), 149 sizeof(_hdr), &_hdr); 150 if (!hdr) 151 return 1; 152 153 if (hdr->type < 32) { 154 __u32 data = raw_sk(sk)->filter.data; 155 156 return ((1U << hdr->type) & data) != 0; 157 } 158 159 /* Do not block unknown ICMP types */ 160 return 0; 161 } 162 163 /* IP input processing comes here for RAW socket delivery. 164 * Caller owns SKB, so we must make clones. 165 * 166 * RFC 1122: SHOULD pass TOS value up to the transport layer. 167 * -> It does. And not only TOS, but all IP header. 168 */ 169 static int raw_v4_input(struct net *net, struct sk_buff *skb, 170 const struct iphdr *iph, int hash) 171 { 172 int sdif = inet_sdif(skb); 173 struct hlist_head *hlist; 174 int dif = inet_iif(skb); 175 int delivered = 0; 176 struct sock *sk; 177 178 hlist = &raw_v4_hashinfo.ht[hash]; 179 rcu_read_lock(); 180 sk_for_each_rcu(sk, hlist) { 181 if (!raw_v4_match(net, sk, iph->protocol, 182 iph->saddr, iph->daddr, dif, sdif)) 183 continue; 184 185 if (atomic_read(&sk->sk_rmem_alloc) >= 186 READ_ONCE(sk->sk_rcvbuf)) { 187 sk_drops_inc(sk); 188 continue; 189 } 190 191 delivered = 1; 192 if ((iph->protocol != IPPROTO_ICMP || !icmp_filter(sk, skb)) && 193 ip_mc_sf_allow(sk, iph->daddr, iph->saddr, 194 skb->dev->ifindex, sdif)) { 195 struct sk_buff *clone = skb_clone(skb, GFP_ATOMIC); 196 197 /* Not releasing hash table! */ 198 if (clone) 199 raw_rcv(sk, clone); 200 } 201 } 202 rcu_read_unlock(); 203 return delivered; 204 } 205 206 int raw_local_deliver(struct sk_buff *skb, int protocol) 207 { 208 struct net *net = dev_net(skb->dev); 209 210 return raw_v4_input(net, skb, ip_hdr(skb), 211 raw_hashfunc(net, protocol)); 212 } 213 214 static void raw_err(struct sock *sk, struct sk_buff *skb, u32 info) 215 { 216 struct inet_sock *inet = inet_sk(sk); 217 const int type = icmp_hdr(skb)->type; 218 const int code = icmp_hdr(skb)->code; 219 int harderr = 0; 220 bool recverr; 221 int err = 0; 222 223 if (type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED) 224 ipv4_sk_update_pmtu(skb, sk, info); 225 else if (type == ICMP_REDIRECT) { 226 ipv4_sk_redirect(skb, sk); 227 return; 228 } 229 230 /* Report error on raw socket, if: 231 1. User requested ip_recverr. 232 2. Socket is connected (otherwise the error indication 233 is useless without ip_recverr and error is hard. 234 */ 235 recverr = inet_test_bit(RECVERR, sk); 236 if (!recverr && sk->sk_state != TCP_ESTABLISHED) 237 return; 238 239 switch (type) { 240 default: 241 case ICMP_TIME_EXCEEDED: 242 err = EHOSTUNREACH; 243 break; 244 case ICMP_SOURCE_QUENCH: 245 return; 246 case ICMP_PARAMETERPROB: 247 err = EPROTO; 248 harderr = 1; 249 break; 250 case ICMP_DEST_UNREACH: 251 err = EHOSTUNREACH; 252 if (code > NR_ICMP_UNREACH) 253 break; 254 if (code == ICMP_FRAG_NEEDED) { 255 harderr = READ_ONCE(inet->pmtudisc) != IP_PMTUDISC_DONT; 256 err = EMSGSIZE; 257 } else { 258 err = icmp_err_convert[code].errno; 259 harderr = icmp_err_convert[code].fatal; 260 } 261 } 262 263 if (recverr) { 264 const struct iphdr *iph = (const struct iphdr *)skb->data; 265 u8 *payload = skb->data + (iph->ihl << 2); 266 267 if (inet_test_bit(HDRINCL, sk)) 268 payload = skb->data; 269 ip_icmp_error(sk, skb, err, 0, info, payload); 270 } 271 272 if (recverr || harderr) { 273 sk->sk_err = err; 274 sk_error_report(sk); 275 } 276 } 277 278 void raw_icmp_error(struct sk_buff *skb, int protocol, u32 info) 279 { 280 struct net *net = dev_net(skb->dev); 281 int dif = skb->dev->ifindex; 282 int sdif = inet_sdif(skb); 283 struct hlist_head *hlist; 284 const struct iphdr *iph; 285 struct sock *sk; 286 int hash; 287 288 hash = raw_hashfunc(net, protocol); 289 hlist = &raw_v4_hashinfo.ht[hash]; 290 291 rcu_read_lock(); 292 sk_for_each_rcu(sk, hlist) { 293 iph = (const struct iphdr *)skb->data; 294 if (!raw_v4_match(net, sk, iph->protocol, 295 iph->daddr, iph->saddr, dif, sdif)) 296 continue; 297 raw_err(sk, skb, info); 298 } 299 rcu_read_unlock(); 300 } 301 302 static int raw_rcv_skb(struct sock *sk, struct sk_buff *skb) 303 { 304 enum skb_drop_reason reason; 305 306 /* Charge it to the socket. */ 307 308 ipv4_pktinfo_prepare(sk, skb, true); 309 reason = sock_queue_rcv_skb_reason(sk, skb); 310 if (reason) { 311 sk_skb_reason_drop(sk, skb, reason); 312 return NET_RX_DROP; 313 } 314 315 return NET_RX_SUCCESS; 316 } 317 318 int raw_rcv(struct sock *sk, struct sk_buff *skb) 319 { 320 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb)) { 321 sk_drops_inc(sk); 322 sk_skb_reason_drop(sk, skb, SKB_DROP_REASON_XFRM_POLICY); 323 return NET_RX_DROP; 324 } 325 nf_reset_ct(skb); 326 327 skb_push(skb, -skb_network_offset(skb)); 328 329 raw_rcv_skb(sk, skb); 330 return 0; 331 } 332 333 static int raw_send_hdrinc(struct sock *sk, struct flowi4 *fl4, 334 struct msghdr *msg, size_t length, 335 struct rtable **rtp, unsigned int flags, 336 const struct sockcm_cookie *sockc) 337 { 338 struct inet_sock *inet = inet_sk(sk); 339 struct net *net = sock_net(sk); 340 struct iphdr *iph; 341 struct sk_buff *skb; 342 unsigned int iphlen; 343 int err; 344 struct rtable *rt = *rtp; 345 int hlen, tlen; 346 347 if (length > rt->dst.dev->mtu) { 348 ip_local_error(sk, EMSGSIZE, fl4->daddr, inet->inet_dport, 349 rt->dst.dev->mtu); 350 return -EMSGSIZE; 351 } 352 if (length < sizeof(struct iphdr)) 353 return -EINVAL; 354 355 if (flags&MSG_PROBE) 356 goto out; 357 358 hlen = LL_RESERVED_SPACE(rt->dst.dev); 359 tlen = rt->dst.dev->needed_tailroom; 360 skb = sock_alloc_send_skb(sk, 361 length + hlen + tlen + 15, 362 flags & MSG_DONTWAIT, &err); 363 if (!skb) 364 goto error; 365 skb_reserve(skb, hlen); 366 367 skb->protocol = htons(ETH_P_IP); 368 skb->priority = sockc->priority; 369 skb->mark = sockc->mark; 370 skb_set_delivery_type_by_clockid(skb, sockc->transmit_time, sk->sk_clockid); 371 skb_dst_set(skb, &rt->dst); 372 *rtp = NULL; 373 374 skb_reset_network_header(skb); 375 iph = ip_hdr(skb); 376 skb_put(skb, length); 377 378 skb->ip_summed = CHECKSUM_NONE; 379 380 skb_setup_tx_timestamp(skb, sockc); 381 382 if (flags & MSG_CONFIRM) 383 skb_set_dst_pending_confirm(skb, 1); 384 385 skb->transport_header = skb->network_header; 386 err = -EFAULT; 387 if (memcpy_from_msg(iph, msg, length)) 388 goto error_free; 389 390 iphlen = iph->ihl * 4; 391 392 /* 393 * We don't want to modify the ip header, but we do need to 394 * be sure that it won't cause problems later along the network 395 * stack. Specifically we want to make sure that iph->ihl is a 396 * sane value. If ihl points beyond the length of the buffer passed 397 * in, reject the frame as invalid 398 */ 399 err = -EINVAL; 400 if (iphlen > length || iphlen < sizeof(*iph)) 401 goto error_free; 402 403 if (iphlen >= sizeof(*iph)) { 404 if (!iph->saddr) 405 iph->saddr = fl4->saddr; 406 iph->check = 0; 407 iph->tot_len = htons(length); 408 if (!iph->id) 409 ip_select_ident(net, skb, NULL); 410 411 iph->check = ip_fast_csum((unsigned char *)iph, iph->ihl); 412 skb->transport_header += iphlen; 413 if (iph->protocol == IPPROTO_ICMP && 414 length >= iphlen + sizeof(struct icmphdr)) 415 icmp_out_count(net, ((struct icmphdr *) 416 skb_transport_header(skb))->type); 417 } 418 419 err = NF_HOOK(NFPROTO_IPV4, NF_INET_LOCAL_OUT, 420 net, sk, skb, NULL, rt->dst.dev, 421 dst_output); 422 if (err > 0) 423 err = net_xmit_errno(err); 424 if (err) 425 goto error; 426 out: 427 return 0; 428 429 error_free: 430 kfree_skb(skb); 431 error: 432 IP_INC_STATS(net, IPSTATS_MIB_OUTDISCARDS); 433 if (err == -ENOBUFS && !inet_test_bit(RECVERR, sk)) 434 err = 0; 435 return err; 436 } 437 438 static int raw_probe_proto_opt(struct raw_frag_vec *rfv, struct flowi4 *fl4) 439 { 440 int err; 441 442 if (fl4->flowi4_proto != IPPROTO_ICMP) 443 return 0; 444 445 /* We only need the first two bytes. */ 446 rfv->hlen = 2; 447 448 err = memcpy_from_msg(rfv->hdr.c, rfv->msg, rfv->hlen); 449 if (err) 450 return err; 451 452 fl4->fl4_icmp_type = rfv->hdr.icmph.type; 453 fl4->fl4_icmp_code = rfv->hdr.icmph.code; 454 455 return 0; 456 } 457 458 static int raw_getfrag(void *from, char *to, int offset, int len, int odd, 459 struct sk_buff *skb) 460 { 461 struct raw_frag_vec *rfv = from; 462 463 if (offset < rfv->hlen) { 464 int copy = min(rfv->hlen - offset, len); 465 466 if (skb->ip_summed == CHECKSUM_PARTIAL) 467 memcpy(to, rfv->hdr.c + offset, copy); 468 else 469 skb->csum = csum_block_add( 470 skb->csum, 471 csum_partial_copy_nocheck(rfv->hdr.c + offset, 472 to, copy), 473 odd); 474 475 odd = 0; 476 offset += copy; 477 to += copy; 478 len -= copy; 479 480 if (!len) 481 return 0; 482 } 483 484 offset -= rfv->hlen; 485 486 return ip_generic_getfrag(rfv->msg, to, offset, len, odd, skb); 487 } 488 489 static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 490 { 491 DEFINE_RAW_FLEX(struct ip_options_rcu, opt_copy, opt.__data, 492 IP_OPTIONS_DATA_FIXED_SIZE); 493 struct inet_sock *inet = inet_sk(sk); 494 struct net *net = sock_net(sk); 495 struct ipcm_cookie ipc; 496 struct rtable *rt = NULL; 497 struct flowi4 fl4; 498 u8 scope; 499 int free = 0; 500 __be32 daddr; 501 __be32 saddr; 502 int uc_index, err; 503 struct raw_frag_vec rfv; 504 int hdrincl; 505 506 err = -EMSGSIZE; 507 if (len > 0xFFFF) 508 goto out; 509 510 hdrincl = inet_test_bit(HDRINCL, sk); 511 512 /* 513 * Check the flags. 514 */ 515 516 err = -EOPNOTSUPP; 517 if (msg->msg_flags & MSG_OOB) /* Mirror BSD error message */ 518 goto out; /* compatibility */ 519 520 /* 521 * Get and verify the address. 522 */ 523 524 if (msg->msg_namelen) { 525 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name); 526 err = -EINVAL; 527 if (msg->msg_namelen < sizeof(*usin)) 528 goto out; 529 if (usin->sin_family != AF_INET) { 530 pr_info_once("%s: %s forgot to set AF_INET. Fix it!\n", 531 __func__, current->comm); 532 err = -EAFNOSUPPORT; 533 if (usin->sin_family) 534 goto out; 535 } 536 daddr = usin->sin_addr.s_addr; 537 /* ANK: I did not forget to get protocol from port field. 538 * I just do not know, who uses this weirdness. 539 * IP_HDRINCL is much more convenient. 540 */ 541 } else { 542 err = -EDESTADDRREQ; 543 if (sk->sk_state != TCP_ESTABLISHED) 544 goto out; 545 daddr = inet->inet_daddr; 546 } 547 548 ipcm_init_sk(&ipc, inet); 549 /* Keep backward compat */ 550 if (hdrincl) 551 ipc.protocol = IPPROTO_RAW; 552 553 if (msg->msg_controllen) { 554 err = ip_cmsg_send(sk, msg, &ipc, false); 555 if (unlikely(err)) { 556 kfree(ipc.opt); 557 goto out; 558 } 559 if (ipc.opt) 560 free = 1; 561 } 562 563 saddr = ipc.addr; 564 ipc.addr = daddr; 565 566 if (!ipc.opt) { 567 struct ip_options_rcu *inet_opt; 568 569 rcu_read_lock(); 570 inet_opt = rcu_dereference(inet->inet_opt); 571 if (inet_opt) { 572 memcpy(opt_copy, inet_opt, 573 sizeof(*inet_opt) + inet_opt->opt.optlen); 574 ipc.opt = opt_copy; 575 } 576 rcu_read_unlock(); 577 } 578 579 if (ipc.opt) { 580 err = -EINVAL; 581 /* Linux does not mangle headers on raw sockets, 582 * so that IP options + IP_HDRINCL is non-sense. 583 */ 584 if (hdrincl) 585 goto done; 586 if (ipc.opt->opt.srr) { 587 if (!daddr) 588 goto done; 589 daddr = ipc.opt->opt.faddr; 590 } 591 } 592 scope = ip_sendmsg_scope(inet, &ipc, msg); 593 594 uc_index = READ_ONCE(inet->uc_index); 595 if (ipv4_is_multicast(daddr)) { 596 if (!ipc.oif || netif_index_is_l3_master(sock_net(sk), ipc.oif)) 597 ipc.oif = READ_ONCE(inet->mc_index); 598 if (!saddr) 599 saddr = READ_ONCE(inet->mc_addr); 600 } else if (!ipc.oif) { 601 ipc.oif = uc_index; 602 } else if (ipv4_is_lbcast(daddr) && uc_index) { 603 /* oif is set, packet is to local broadcast 604 * and uc_index is set. oif is most likely set 605 * by sk_bound_dev_if. If uc_index != oif check if the 606 * oif is an L3 master and uc_index is an L3 slave. 607 * If so, we want to allow the send using the uc_index. 608 */ 609 if (ipc.oif != uc_index && 610 ipc.oif == l3mdev_master_ifindex_by_index(sock_net(sk), 611 uc_index)) { 612 ipc.oif = uc_index; 613 } 614 } 615 616 flowi4_init_output(&fl4, ipc.oif, ipc.sockc.mark, 617 ipc.tos & INET_DSCP_MASK, scope, 618 hdrincl ? ipc.protocol : sk->sk_protocol, 619 inet_sk_flowi_flags(sk) | 620 (hdrincl ? FLOWI_FLAG_KNOWN_NH : 0), 621 daddr, saddr, 0, 0, sk_uid(sk)); 622 623 fl4.fl4_icmp_type = 0; 624 fl4.fl4_icmp_code = 0; 625 626 if (!hdrincl) { 627 rfv.msg = msg; 628 rfv.hlen = 0; 629 630 err = raw_probe_proto_opt(&rfv, &fl4); 631 if (err) 632 goto done; 633 } 634 635 security_sk_classify_flow(sk, flowi4_to_flowi_common(&fl4)); 636 rt = ip_route_output_flow(net, &fl4, sk); 637 if (IS_ERR(rt)) { 638 err = PTR_ERR(rt); 639 rt = NULL; 640 goto done; 641 } 642 643 err = -EACCES; 644 if (rt->rt_flags & RTCF_BROADCAST && !sock_flag(sk, SOCK_BROADCAST)) 645 goto done; 646 647 if (msg->msg_flags & MSG_CONFIRM) 648 goto do_confirm; 649 back_from_confirm: 650 651 if (hdrincl) 652 err = raw_send_hdrinc(sk, &fl4, msg, len, 653 &rt, msg->msg_flags, &ipc.sockc); 654 655 else { 656 if (!ipc.addr) 657 ipc.addr = fl4.daddr; 658 lock_sock(sk); 659 err = ip_append_data(sk, &fl4, raw_getfrag, 660 &rfv, len, 0, 661 &ipc, &rt, msg->msg_flags); 662 if (err) 663 ip_flush_pending_frames(sk); 664 else if (!(msg->msg_flags & MSG_MORE)) { 665 err = ip_push_pending_frames(sk, &fl4); 666 if (err == -ENOBUFS && !inet_test_bit(RECVERR, sk)) 667 err = 0; 668 } 669 release_sock(sk); 670 } 671 done: 672 if (free) 673 kfree(ipc.opt); 674 ip_rt_put(rt); 675 676 out: 677 if (err < 0) 678 return err; 679 return len; 680 681 do_confirm: 682 if (msg->msg_flags & MSG_PROBE) 683 dst_confirm_neigh(&rt->dst, &fl4.daddr); 684 if (!(msg->msg_flags & MSG_PROBE) || len) 685 goto back_from_confirm; 686 err = 0; 687 goto done; 688 } 689 690 static void raw_close(struct sock *sk, long timeout) 691 { 692 /* 693 * Raw sockets may have direct kernel references. Kill them. 694 */ 695 ip_ra_control(sk, 0, NULL); 696 697 sk_common_release(sk); 698 } 699 700 static void raw_destroy(struct sock *sk) 701 { 702 lock_sock(sk); 703 ip_flush_pending_frames(sk); 704 release_sock(sk); 705 } 706 707 /* This gets rid of all the nasties in af_inet. -DaveM */ 708 static int raw_bind(struct sock *sk, struct sockaddr_unsized *uaddr, 709 int addr_len) 710 { 711 struct inet_sock *inet = inet_sk(sk); 712 struct sockaddr_in *addr = (struct sockaddr_in *) uaddr; 713 struct net *net = sock_net(sk); 714 u32 tb_id = RT_TABLE_LOCAL; 715 int ret = -EINVAL; 716 int chk_addr_ret; 717 718 lock_sock(sk); 719 if (sk->sk_state != TCP_CLOSE || addr_len < sizeof(struct sockaddr_in)) 720 goto out; 721 722 if (sk->sk_bound_dev_if) 723 tb_id = l3mdev_fib_table_by_index(net, 724 sk->sk_bound_dev_if) ? : tb_id; 725 726 chk_addr_ret = inet_addr_type_table(net, addr->sin_addr.s_addr, tb_id); 727 728 ret = -EADDRNOTAVAIL; 729 if (!inet_addr_valid_or_nonlocal(net, inet, addr->sin_addr.s_addr, 730 chk_addr_ret)) 731 goto out; 732 733 inet->inet_saddr = addr->sin_addr.s_addr; 734 WRITE_ONCE(inet->inet_rcv_saddr, addr->sin_addr.s_addr); 735 if (chk_addr_ret == RTN_MULTICAST || chk_addr_ret == RTN_BROADCAST) 736 inet->inet_saddr = 0; /* Use device */ 737 sk_dst_reset(sk); 738 ret = 0; 739 out: 740 release_sock(sk); 741 return ret; 742 } 743 744 /* 745 * This should be easy, if there is something there 746 * we return it, otherwise we block. 747 */ 748 749 static int raw_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 750 int flags) 751 { 752 struct inet_sock *inet = inet_sk(sk); 753 size_t copied = 0; 754 int err = -EOPNOTSUPP; 755 DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name); 756 struct sk_buff *skb; 757 758 if (flags & MSG_OOB) 759 goto out; 760 761 if (flags & MSG_ERRQUEUE) { 762 err = ip_recv_error(sk, msg, len); 763 goto out; 764 } 765 766 skb = skb_recv_datagram(sk, flags, &err); 767 if (!skb) 768 goto out; 769 770 copied = skb->len; 771 if (len < copied) { 772 msg->msg_flags |= MSG_TRUNC; 773 copied = len; 774 } 775 776 err = skb_copy_datagram_msg(skb, 0, msg, copied); 777 if (err) 778 goto done; 779 780 sock_recv_cmsgs(msg, sk, skb); 781 782 /* Copy the address. */ 783 if (sin) { 784 sin->sin_family = AF_INET; 785 sin->sin_addr.s_addr = ip_hdr(skb)->saddr; 786 sin->sin_port = 0; 787 memset(&sin->sin_zero, 0, sizeof(sin->sin_zero)); 788 msg->msg_namelen = sizeof(*sin); 789 } 790 if (inet_cmsg_flags(inet)) 791 ip_cmsg_recv(msg, skb); 792 if (flags & MSG_TRUNC) 793 copied = skb->len; 794 done: 795 skb_free_datagram(sk, skb); 796 out: 797 if (err) 798 return err; 799 return copied; 800 } 801 802 static int raw_sk_init(struct sock *sk) 803 { 804 struct raw_sock *rp = raw_sk(sk); 805 806 sk->sk_drop_counters = &rp->drop_counters; 807 if (inet_sk(sk)->inet_num == IPPROTO_ICMP) 808 memset(&rp->filter, 0, sizeof(rp->filter)); 809 return 0; 810 } 811 812 static int raw_seticmpfilter(struct sock *sk, sockptr_t optval, int optlen) 813 { 814 if (optlen > sizeof(struct icmp_filter)) 815 optlen = sizeof(struct icmp_filter); 816 if (copy_from_sockptr(&raw_sk(sk)->filter, optval, optlen)) 817 return -EFAULT; 818 return 0; 819 } 820 821 static int raw_geticmpfilter(struct sock *sk, char __user *optval, int __user *optlen) 822 { 823 int len, ret = -EFAULT; 824 825 if (get_user(len, optlen)) 826 goto out; 827 ret = -EINVAL; 828 if (len < 0) 829 goto out; 830 if (len > sizeof(struct icmp_filter)) 831 len = sizeof(struct icmp_filter); 832 ret = -EFAULT; 833 if (put_user(len, optlen) || 834 copy_to_user(optval, &raw_sk(sk)->filter, len)) 835 goto out; 836 ret = 0; 837 out: return ret; 838 } 839 840 static int do_raw_setsockopt(struct sock *sk, int optname, 841 sockptr_t optval, unsigned int optlen) 842 { 843 if (optname == ICMP_FILTER) { 844 if (inet_sk(sk)->inet_num != IPPROTO_ICMP) 845 return -EOPNOTSUPP; 846 else 847 return raw_seticmpfilter(sk, optval, optlen); 848 } 849 return -ENOPROTOOPT; 850 } 851 852 static int raw_setsockopt(struct sock *sk, int level, int optname, 853 sockptr_t optval, unsigned int optlen) 854 { 855 if (level != SOL_RAW) 856 return ip_setsockopt(sk, level, optname, optval, optlen); 857 return do_raw_setsockopt(sk, optname, optval, optlen); 858 } 859 860 static int do_raw_getsockopt(struct sock *sk, int optname, 861 char __user *optval, int __user *optlen) 862 { 863 if (optname == ICMP_FILTER) { 864 if (inet_sk(sk)->inet_num != IPPROTO_ICMP) 865 return -EOPNOTSUPP; 866 else 867 return raw_geticmpfilter(sk, optval, optlen); 868 } 869 return -ENOPROTOOPT; 870 } 871 872 static int raw_getsockopt(struct sock *sk, int level, int optname, 873 char __user *optval, int __user *optlen) 874 { 875 if (level != SOL_RAW) 876 return ip_getsockopt(sk, level, optname, optval, optlen); 877 return do_raw_getsockopt(sk, optname, optval, optlen); 878 } 879 880 static int raw_ioctl(struct sock *sk, int cmd, int *karg) 881 { 882 switch (cmd) { 883 case SIOCOUTQ: { 884 *karg = sk_wmem_alloc_get(sk); 885 return 0; 886 } 887 case SIOCINQ: { 888 struct sk_buff *skb; 889 890 spin_lock_bh(&sk->sk_receive_queue.lock); 891 skb = skb_peek(&sk->sk_receive_queue); 892 if (skb) 893 *karg = skb->len; 894 else 895 *karg = 0; 896 spin_unlock_bh(&sk->sk_receive_queue.lock); 897 return 0; 898 } 899 900 default: 901 #ifdef CONFIG_IP_MROUTE 902 return ipmr_ioctl(sk, cmd, karg); 903 #else 904 return -ENOIOCTLCMD; 905 #endif 906 } 907 } 908 909 #ifdef CONFIG_COMPAT 910 static int compat_raw_ioctl(struct sock *sk, unsigned int cmd, unsigned long arg) 911 { 912 switch (cmd) { 913 case SIOCOUTQ: 914 case SIOCINQ: 915 return -ENOIOCTLCMD; 916 default: 917 #ifdef CONFIG_IP_MROUTE 918 return ipmr_compat_ioctl(sk, cmd, compat_ptr(arg)); 919 #else 920 return -ENOIOCTLCMD; 921 #endif 922 } 923 } 924 #endif 925 926 int raw_abort(struct sock *sk, int err) 927 { 928 lock_sock(sk); 929 930 sk->sk_err = err; 931 sk_error_report(sk); 932 __udp_disconnect(sk, 0); 933 934 release_sock(sk); 935 936 return 0; 937 } 938 939 struct proto raw_prot = { 940 .name = "RAW", 941 .owner = THIS_MODULE, 942 .close = raw_close, 943 .destroy = raw_destroy, 944 .connect = ip4_datagram_connect, 945 .disconnect = __udp_disconnect, 946 .ioctl = raw_ioctl, 947 .init = raw_sk_init, 948 .setsockopt = raw_setsockopt, 949 .getsockopt = raw_getsockopt, 950 .sendmsg = raw_sendmsg, 951 .recvmsg = raw_recvmsg, 952 .bind = raw_bind, 953 .backlog_rcv = raw_rcv_skb, 954 .release_cb = ip4_datagram_release_cb, 955 .hash = raw_hash_sk, 956 .unhash = raw_unhash_sk, 957 .obj_size = sizeof(struct raw_sock), 958 .useroffset = offsetof(struct raw_sock, filter), 959 .usersize = sizeof_field(struct raw_sock, filter), 960 .h.raw_hash = &raw_v4_hashinfo, 961 #ifdef CONFIG_COMPAT 962 .compat_ioctl = compat_raw_ioctl, 963 #endif 964 .diag_destroy = raw_abort, 965 }; 966 967 #ifdef CONFIG_PROC_FS 968 static struct sock *raw_get_first(struct seq_file *seq, int bucket) 969 { 970 struct raw_hashinfo *h = pde_data(file_inode(seq->file)); 971 struct raw_iter_state *state = raw_seq_private(seq); 972 struct hlist_head *hlist; 973 struct sock *sk; 974 975 for (state->bucket = bucket; state->bucket < RAW_HTABLE_SIZE; 976 ++state->bucket) { 977 hlist = &h->ht[state->bucket]; 978 sk_for_each(sk, hlist) { 979 if (sock_net(sk) == seq_file_net(seq)) 980 return sk; 981 } 982 } 983 return NULL; 984 } 985 986 static struct sock *raw_get_next(struct seq_file *seq, struct sock *sk) 987 { 988 struct raw_iter_state *state = raw_seq_private(seq); 989 990 do { 991 sk = sk_next(sk); 992 } while (sk && sock_net(sk) != seq_file_net(seq)); 993 994 if (!sk) 995 return raw_get_first(seq, state->bucket + 1); 996 return sk; 997 } 998 999 static struct sock *raw_get_idx(struct seq_file *seq, loff_t pos) 1000 { 1001 struct sock *sk = raw_get_first(seq, 0); 1002 1003 if (sk) 1004 while (pos && (sk = raw_get_next(seq, sk)) != NULL) 1005 --pos; 1006 return pos ? NULL : sk; 1007 } 1008 1009 void *raw_seq_start(struct seq_file *seq, loff_t *pos) 1010 __acquires(&h->lock) 1011 { 1012 struct raw_hashinfo *h = pde_data(file_inode(seq->file)); 1013 1014 spin_lock(&h->lock); 1015 1016 return *pos ? raw_get_idx(seq, *pos - 1) : SEQ_START_TOKEN; 1017 } 1018 1019 void *raw_seq_next(struct seq_file *seq, void *v, loff_t *pos) 1020 { 1021 struct sock *sk; 1022 1023 if (v == SEQ_START_TOKEN) 1024 sk = raw_get_first(seq, 0); 1025 else 1026 sk = raw_get_next(seq, v); 1027 ++*pos; 1028 return sk; 1029 } 1030 1031 void raw_seq_stop(struct seq_file *seq, void *v) 1032 __releases(&h->lock) 1033 { 1034 struct raw_hashinfo *h = pde_data(file_inode(seq->file)); 1035 1036 spin_unlock(&h->lock); 1037 } 1038 1039 static void raw_sock_seq_show(struct seq_file *seq, struct sock *sp, int i) 1040 { 1041 struct inet_sock *inet = inet_sk(sp); 1042 __be32 dest = inet->inet_daddr, 1043 src = inet->inet_rcv_saddr; 1044 __u16 destp = 0, 1045 srcp = inet->inet_num; 1046 1047 seq_printf(seq, "%4d: %08X:%04X %08X:%04X" 1048 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %llu %d %pK %u\n", 1049 i, src, srcp, dest, destp, sp->sk_state, 1050 sk_wmem_alloc_get(sp), 1051 sk_rmem_alloc_get(sp), 1052 0, 0L, 0, 1053 from_kuid_munged(seq_user_ns(seq), sk_uid(sp)), 1054 0, sock_i_ino(sp), 1055 refcount_read(&sp->sk_refcnt), sp, sk_drops_read(sp)); 1056 } 1057 1058 static int raw_seq_show(struct seq_file *seq, void *v) 1059 { 1060 if (v == SEQ_START_TOKEN) 1061 seq_printf(seq, " sl local_address rem_address st tx_queue " 1062 "rx_queue tr tm->when retrnsmt uid timeout " 1063 "inode ref pointer drops\n"); 1064 else 1065 raw_sock_seq_show(seq, v, raw_seq_private(seq)->bucket); 1066 return 0; 1067 } 1068 1069 static const struct seq_operations raw_seq_ops = { 1070 .start = raw_seq_start, 1071 .next = raw_seq_next, 1072 .stop = raw_seq_stop, 1073 .show = raw_seq_show, 1074 }; 1075 1076 static __net_init int raw_init_net(struct net *net) 1077 { 1078 if (!proc_create_net_data("raw", 0444, net->proc_net, &raw_seq_ops, 1079 sizeof(struct raw_iter_state), &raw_v4_hashinfo)) 1080 return -ENOMEM; 1081 1082 return 0; 1083 } 1084 1085 static __net_exit void raw_exit_net(struct net *net) 1086 { 1087 remove_proc_entry("raw", net->proc_net); 1088 } 1089 1090 static __net_initdata struct pernet_operations raw_net_ops = { 1091 .init = raw_init_net, 1092 .exit = raw_exit_net, 1093 }; 1094 1095 int __init raw_proc_init(void) 1096 { 1097 1098 return register_pernet_subsys(&raw_net_ops); 1099 } 1100 1101 void __init raw_proc_exit(void) 1102 { 1103 unregister_pernet_subsys(&raw_net_ops); 1104 } 1105 #endif /* CONFIG_PROC_FS */ 1106 1107 static void raw_sysctl_init_net(struct net *net) 1108 { 1109 #ifdef CONFIG_NET_L3_MASTER_DEV 1110 net->ipv4.sysctl_raw_l3mdev_accept = 1; 1111 #endif 1112 } 1113 1114 static int __net_init raw_sysctl_init(struct net *net) 1115 { 1116 raw_sysctl_init_net(net); 1117 return 0; 1118 } 1119 1120 static struct pernet_operations __net_initdata raw_sysctl_ops = { 1121 .init = raw_sysctl_init, 1122 }; 1123 1124 void __init raw_init(void) 1125 { 1126 raw_sysctl_init_net(&init_net); 1127 if (register_pernet_subsys(&raw_sysctl_ops)) 1128 panic("RAW: failed to init sysctl parameters.\n"); 1129 } 1130