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 * The User Datagram Protocol (UDP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 12 * Alan Cox, <alan@lxorguk.ukuu.org.uk> 13 * Hirokazu Takahashi, <taka@valinux.co.jp> 14 * 15 * Fixes: 16 * Alan Cox : verify_area() calls 17 * Alan Cox : stopped close while in use off icmp 18 * messages. Not a fix but a botch that 19 * for udp at least is 'valid'. 20 * Alan Cox : Fixed icmp handling properly 21 * Alan Cox : Correct error for oversized datagrams 22 * Alan Cox : Tidied select() semantics. 23 * Alan Cox : udp_err() fixed properly, also now 24 * select and read wake correctly on errors 25 * Alan Cox : udp_send verify_area moved to avoid mem leak 26 * Alan Cox : UDP can count its memory 27 * Alan Cox : send to an unknown connection causes 28 * an ECONNREFUSED off the icmp, but 29 * does NOT close. 30 * Alan Cox : Switched to new sk_buff handlers. No more backlog! 31 * Alan Cox : Using generic datagram code. Even smaller and the PEEK 32 * bug no longer crashes it. 33 * Fred Van Kempen : Net2e support for sk->broadcast. 34 * Alan Cox : Uses skb_free_datagram 35 * Alan Cox : Added get/set sockopt support. 36 * Alan Cox : Broadcasting without option set returns EACCES. 37 * Alan Cox : No wakeup calls. Instead we now use the callbacks. 38 * Alan Cox : Use ip_tos and ip_ttl 39 * Alan Cox : SNMP Mibs 40 * Alan Cox : MSG_DONTROUTE, and 0.0.0.0 support. 41 * Matt Dillon : UDP length checks. 42 * Alan Cox : Smarter af_inet used properly. 43 * Alan Cox : Use new kernel side addressing. 44 * Alan Cox : Incorrect return on truncated datagram receive. 45 * Arnt Gulbrandsen : New udp_send and stuff 46 * Alan Cox : Cache last socket 47 * Alan Cox : Route cache 48 * Jon Peatfield : Minor efficiency fix to sendto(). 49 * Mike Shaver : RFC1122 checks. 50 * Alan Cox : Nonblocking error fix. 51 * Willy Konynenberg : Transparent proxying support. 52 * Mike McLagan : Routing by source 53 * David S. Miller : New socket lookup architecture. 54 * Last socket cache retained as it 55 * does have a high hit rate. 56 * Olaf Kirch : Don't linearise iovec on sendmsg. 57 * Andi Kleen : Some cleanups, cache destination entry 58 * for connect. 59 * Vitaly E. Lavrov : Transparent proxy revived after year coma. 60 * Melvin Smith : Check msg_name not msg_namelen in sendto(), 61 * return ENOTCONN for unconnected sockets (POSIX) 62 * Janos Farkas : don't deliver multi/broadcasts to a different 63 * bound-to-device socket 64 * Hirokazu Takahashi : HW checksumming for outgoing UDP 65 * datagrams. 66 * Hirokazu Takahashi : sendfile() on UDP works now. 67 * Arnaldo C. Melo : convert /proc/net/udp to seq_file 68 * YOSHIFUJI Hideaki @USAGI and: Support IPV6_V6ONLY socket option, which 69 * Alexey Kuznetsov: allow both IPv4 and IPv6 sockets to bind 70 * a single port at the same time. 71 * Derek Atkins <derek@ihtfp.com>: Add Encapsulation Support 72 * James Chapman : Add L2TP encapsulation type. 73 */ 74 75 #define pr_fmt(fmt) "UDP: " fmt 76 77 #include <linux/bpf-cgroup.h> 78 #include <linux/uaccess.h> 79 #include <linux/uio.h> 80 #include <asm/ioctls.h> 81 #include <linux/memblock.h> 82 #include <linux/highmem.h> 83 #include <linux/types.h> 84 #include <linux/fcntl.h> 85 #include <linux/module.h> 86 #include <linux/socket.h> 87 #include <linux/sockios.h> 88 #include <linux/igmp.h> 89 #include <linux/inetdevice.h> 90 #include <linux/in.h> 91 #include <linux/errno.h> 92 #include <linux/timer.h> 93 #include <linux/mm.h> 94 #include <linux/inet.h> 95 #include <linux/netdevice.h> 96 #include <linux/slab.h> 97 #include <linux/sock_diag.h> 98 #include <net/tcp_states.h> 99 #include <linux/skbuff.h> 100 #include <linux/proc_fs.h> 101 #include <linux/seq_file.h> 102 #include <net/aligned_data.h> 103 #include <net/net_namespace.h> 104 #include <net/icmp.h> 105 #include <net/inet_common.h> 106 #include <net/inet_hashtables.h> 107 #include <net/ip.h> 108 #include <net/ip_tunnels.h> 109 #include <net/route.h> 110 #include <net/checksum.h> 111 #include <net/gso.h> 112 #include <net/xfrm.h> 113 #include <trace/events/udp.h> 114 #include <linux/static_key.h> 115 #include <linux/btf_ids.h> 116 #include <trace/events/skb.h> 117 #include <net/busy_poll.h> 118 #include <net/sock_reuseport.h> 119 #include <net/addrconf.h> 120 #include <net/udp_tunnel.h> 121 #include <net/gro.h> 122 #include <net/rps.h> 123 124 struct udp_table udp_table __read_mostly; 125 126 long sysctl_udp_mem[3] __read_mostly; 127 128 DEFINE_PER_CPU(int, udp_memory_per_cpu_fw_alloc); 129 EXPORT_PER_CPU_SYMBOL_GPL(udp_memory_per_cpu_fw_alloc); 130 131 #define MAX_UDP_PORTS 65536 132 #define PORTS_PER_CHAIN (MAX_UDP_PORTS / UDP_HTABLE_SIZE_MIN_PERNET) 133 134 static int udp_lib_lport_inuse(struct net *net, __u16 num, 135 const struct udp_hslot *hslot, 136 unsigned long *bitmap, 137 struct sock *sk, unsigned int log) 138 { 139 kuid_t uid = sk_uid(sk); 140 struct sock *sk2; 141 142 sk_for_each(sk2, &hslot->head) { 143 if (net_eq(sock_net(sk2), net) && 144 sk2 != sk && 145 (bitmap || udp_sk(sk2)->udp_port_hash == num) && 146 (!sk2->sk_reuse || !sk->sk_reuse) && 147 (!sk2->sk_bound_dev_if || !sk->sk_bound_dev_if || 148 sk2->sk_bound_dev_if == sk->sk_bound_dev_if) && 149 inet_rcv_saddr_equal(sk, sk2, true)) { 150 if (sk2->sk_reuseport && sk->sk_reuseport && 151 !rcu_access_pointer(sk->sk_reuseport_cb) && 152 uid_eq(uid, sk_uid(sk2))) { 153 if (!bitmap) 154 return 0; 155 } else { 156 if (!bitmap) 157 return 1; 158 __set_bit(udp_sk(sk2)->udp_port_hash >> log, 159 bitmap); 160 } 161 } 162 } 163 return 0; 164 } 165 166 /* 167 * Note: we still hold spinlock of primary hash chain, so no other writer 168 * can insert/delete a socket with local_port == num 169 */ 170 static int udp_lib_lport_inuse2(struct net *net, __u16 num, 171 struct udp_hslot *hslot2, 172 struct sock *sk) 173 { 174 kuid_t uid = sk_uid(sk); 175 struct sock *sk2; 176 int res = 0; 177 178 spin_lock(&hslot2->lock); 179 udp_portaddr_for_each_entry(sk2, &hslot2->head) { 180 if (net_eq(sock_net(sk2), net) && 181 sk2 != sk && 182 (udp_sk(sk2)->udp_port_hash == num) && 183 (!sk2->sk_reuse || !sk->sk_reuse) && 184 (!sk2->sk_bound_dev_if || !sk->sk_bound_dev_if || 185 sk2->sk_bound_dev_if == sk->sk_bound_dev_if) && 186 inet_rcv_saddr_equal(sk, sk2, true)) { 187 if (sk2->sk_reuseport && sk->sk_reuseport && 188 !rcu_access_pointer(sk->sk_reuseport_cb) && 189 uid_eq(uid, sk_uid(sk2))) { 190 res = 0; 191 } else { 192 res = 1; 193 } 194 break; 195 } 196 } 197 spin_unlock(&hslot2->lock); 198 return res; 199 } 200 201 static int udp_reuseport_add_sock(struct sock *sk, struct udp_hslot *hslot) 202 { 203 struct net *net = sock_net(sk); 204 kuid_t uid = sk_uid(sk); 205 struct sock *sk2; 206 207 sk_for_each(sk2, &hslot->head) { 208 if (net_eq(sock_net(sk2), net) && 209 sk2 != sk && 210 sk2->sk_family == sk->sk_family && 211 ipv6_only_sock(sk2) == ipv6_only_sock(sk) && 212 (udp_sk(sk2)->udp_port_hash == udp_sk(sk)->udp_port_hash) && 213 (sk2->sk_bound_dev_if == sk->sk_bound_dev_if) && 214 sk2->sk_reuseport && uid_eq(uid, sk_uid(sk2)) && 215 inet_rcv_saddr_equal(sk, sk2, false)) { 216 return reuseport_add_sock(sk, sk2, 217 inet_rcv_saddr_any(sk)); 218 } 219 } 220 221 return reuseport_alloc(sk, inet_rcv_saddr_any(sk)); 222 } 223 224 /** 225 * udp_lib_get_port - UDP port lookup for IPv4 and IPv6 226 * 227 * @sk: socket struct in question 228 * @snum: port number to look up 229 * @hash2_nulladdr: AF-dependent hash value in secondary hash chains, 230 * with NULL address 231 */ 232 int udp_lib_get_port(struct sock *sk, unsigned short snum, 233 unsigned int hash2_nulladdr) 234 { 235 struct udp_hslot *hslot, *hslot2; 236 struct net *net = sock_net(sk); 237 struct udp_table *udptable; 238 int error = -EADDRINUSE; 239 240 udptable = net->ipv4.udp_table; 241 242 if (!snum) { 243 DECLARE_BITMAP(bitmap, PORTS_PER_CHAIN); 244 unsigned short first, last; 245 int low, high, remaining; 246 unsigned int rand; 247 248 inet_sk_get_local_port_range(sk, &low, &high); 249 remaining = (high - low) + 1; 250 251 rand = get_random_u32(); 252 first = reciprocal_scale(rand, remaining) + low; 253 /* 254 * force rand to be an odd multiple of UDP_HTABLE_SIZE 255 */ 256 rand = (rand | 1) * (udptable->mask + 1); 257 last = first + udptable->mask + 1; 258 do { 259 hslot = udp_hashslot(udptable, net, first); 260 bitmap_zero(bitmap, PORTS_PER_CHAIN); 261 spin_lock_bh(&hslot->lock); 262 udp_lib_lport_inuse(net, snum, hslot, bitmap, sk, 263 udptable->log); 264 265 snum = first; 266 /* 267 * Iterate on all possible values of snum for this hash. 268 * Using steps of an odd multiple of UDP_HTABLE_SIZE 269 * give us randomization and full range coverage. 270 */ 271 do { 272 if (low <= snum && snum <= high && 273 !test_bit(snum >> udptable->log, bitmap) && 274 !inet_is_local_reserved_port(net, snum)) 275 goto found; 276 snum += rand; 277 } while (snum != first); 278 spin_unlock_bh(&hslot->lock); 279 cond_resched(); 280 } while (++first != last); 281 goto fail; 282 } else { 283 hslot = udp_hashslot(udptable, net, snum); 284 spin_lock_bh(&hslot->lock); 285 if (inet_use_hash2_on_bind(sk) && hslot->count > 10) { 286 int exist; 287 unsigned int slot2 = udp_sk(sk)->udp_portaddr_hash ^ snum; 288 289 slot2 &= udptable->mask; 290 hash2_nulladdr &= udptable->mask; 291 292 hslot2 = udp_hashslot2(udptable, slot2); 293 if (hslot->count < hslot2->count) 294 goto scan_primary_hash; 295 296 exist = udp_lib_lport_inuse2(net, snum, hslot2, sk); 297 if (!exist && (hash2_nulladdr != slot2)) { 298 hslot2 = udp_hashslot2(udptable, hash2_nulladdr); 299 exist = udp_lib_lport_inuse2(net, snum, hslot2, 300 sk); 301 } 302 if (exist) 303 goto fail_unlock; 304 else 305 goto found; 306 } 307 scan_primary_hash: 308 if (udp_lib_lport_inuse(net, snum, hslot, NULL, sk, 0)) 309 goto fail_unlock; 310 } 311 found: 312 inet_sk(sk)->inet_num = snum; 313 udp_sk(sk)->udp_port_hash = snum; 314 udp_sk(sk)->udp_portaddr_hash ^= snum; 315 if (sk_unhashed(sk)) { 316 if (sk->sk_reuseport && 317 udp_reuseport_add_sock(sk, hslot)) { 318 inet_sk(sk)->inet_num = 0; 319 udp_sk(sk)->udp_port_hash = 0; 320 udp_sk(sk)->udp_portaddr_hash ^= snum; 321 goto fail_unlock; 322 } 323 324 sock_set_flag(sk, SOCK_RCU_FREE); 325 326 sk_add_node_rcu(sk, &hslot->head); 327 hslot->count++; 328 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 329 330 hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash); 331 spin_lock(&hslot2->lock); 332 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport && 333 sk->sk_family == AF_INET6) 334 hlist_add_tail_rcu(&udp_sk(sk)->udp_portaddr_node, 335 &hslot2->head); 336 else 337 hlist_add_head_rcu(&udp_sk(sk)->udp_portaddr_node, 338 &hslot2->head); 339 hslot2->count++; 340 spin_unlock(&hslot2->lock); 341 } 342 343 error = 0; 344 fail_unlock: 345 spin_unlock_bh(&hslot->lock); 346 fail: 347 return error; 348 } 349 350 static int udp_v4_get_port(struct sock *sk, unsigned short snum) 351 { 352 unsigned int hash2_nulladdr = 353 ipv4_portaddr_hash(sock_net(sk), htonl(INADDR_ANY), snum); 354 unsigned int hash2_partial = 355 ipv4_portaddr_hash(sock_net(sk), inet_sk(sk)->inet_rcv_saddr, 0); 356 357 /* precompute partial secondary hash */ 358 udp_sk(sk)->udp_portaddr_hash = hash2_partial; 359 return udp_lib_get_port(sk, snum, hash2_nulladdr); 360 } 361 362 static __always_inline int 363 compute_score(struct sock *sk, const struct net *net, 364 __be32 saddr, __be16 sport, __be32 daddr, 365 unsigned short hnum, int dif, int sdif) 366 { 367 int score; 368 struct inet_sock *inet; 369 bool dev_match; 370 371 if (!net_eq(sock_net(sk), net) || 372 udp_sk(sk)->udp_port_hash != hnum || 373 ipv6_only_sock(sk)) 374 return -1; 375 376 if (sk->sk_rcv_saddr != daddr) 377 return -1; 378 379 score = (sk->sk_family == PF_INET) ? 2 : 1; 380 381 inet = inet_sk(sk); 382 if (inet->inet_daddr) { 383 if (inet->inet_daddr != saddr) 384 return -1; 385 score += 4; 386 } 387 388 if (inet->inet_dport) { 389 if (inet->inet_dport != sport) 390 return -1; 391 score += 4; 392 } 393 394 dev_match = udp_sk_bound_dev_eq(net, sk->sk_bound_dev_if, 395 dif, sdif); 396 if (!dev_match) 397 return -1; 398 if (sk->sk_bound_dev_if) 399 score += 4; 400 401 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id()) 402 score++; 403 return score; 404 } 405 406 u32 udp_ehashfn(const struct net *net, const __be32 laddr, const __u16 lport, 407 const __be32 faddr, const __be16 fport) 408 { 409 net_get_random_once(&udp_ehash_secret, sizeof(udp_ehash_secret)); 410 411 return __inet_ehashfn(laddr, lport, faddr, fport, 412 udp_ehash_secret + net_hash_mix(net)); 413 } 414 415 /** 416 * udp4_lib_lookup1() - Simplified lookup using primary hash (destination port) 417 * @net: Network namespace 418 * @saddr: Source address, network order 419 * @sport: Source port, network order 420 * @daddr: Destination address, network order 421 * @hnum: Destination port, host order 422 * @dif: Destination interface index 423 * @sdif: Destination bridge port index, if relevant 424 * @udptable: Set of UDP hash tables 425 * 426 * Simplified lookup to be used as fallback if no sockets are found due to a 427 * potential race between (receive) address change, and lookup happening before 428 * the rehash operation. This function ignores SO_REUSEPORT groups while scoring 429 * result sockets, because if we have one, we don't need the fallback at all. 430 * 431 * Called under rcu_read_lock(). 432 * 433 * Return: socket with highest matching score if any, NULL if none 434 */ 435 static struct sock *udp4_lib_lookup1(const struct net *net, 436 __be32 saddr, __be16 sport, 437 __be32 daddr, unsigned int hnum, 438 int dif, int sdif, 439 const struct udp_table *udptable) 440 { 441 unsigned int slot = udp_hashfn(net, hnum, udptable->mask); 442 struct udp_hslot *hslot = &udptable->hash[slot]; 443 struct sock *sk, *result = NULL; 444 int score, badness = 0; 445 446 sk_for_each_rcu(sk, &hslot->head) { 447 score = compute_score(sk, net, 448 saddr, sport, daddr, hnum, dif, sdif); 449 if (score > badness) { 450 result = sk; 451 badness = score; 452 } 453 } 454 455 return result; 456 } 457 458 /* called with rcu_read_lock() */ 459 static struct sock *udp4_lib_lookup2(const struct net *net, 460 __be32 saddr, __be16 sport, 461 __be32 daddr, unsigned int hnum, 462 int dif, int sdif, 463 struct udp_hslot *hslot2, 464 struct sk_buff *skb) 465 { 466 struct sock *sk, *result; 467 int score, badness; 468 bool need_rescore; 469 470 result = NULL; 471 badness = 0; 472 udp_portaddr_for_each_entry_rcu(sk, &hslot2->head) { 473 need_rescore = false; 474 rescore: 475 score = compute_score(need_rescore ? result : sk, net, saddr, 476 sport, daddr, hnum, dif, sdif); 477 if (score > badness) { 478 badness = score; 479 480 if (need_rescore) 481 continue; 482 483 if (sk->sk_state == TCP_ESTABLISHED) { 484 result = sk; 485 continue; 486 } 487 488 result = inet_lookup_reuseport(net, sk, skb, sizeof(struct udphdr), 489 saddr, sport, daddr, hnum, udp_ehashfn); 490 if (!result) { 491 result = sk; 492 continue; 493 } 494 495 /* Fall back to scoring if group has connections */ 496 if (!reuseport_has_conns(sk)) 497 return result; 498 499 /* Reuseport logic returned an error, keep original score. */ 500 if (IS_ERR(result)) 501 continue; 502 503 /* compute_score is too long of a function to be 504 * inlined twice here, and calling it uninlined 505 * here yields measurable overhead for some 506 * workloads. Work around it by jumping 507 * backwards to rescore 'result'. 508 */ 509 need_rescore = true; 510 goto rescore; 511 } 512 } 513 return result; 514 } 515 516 #if IS_ENABLED(CONFIG_BASE_SMALL) 517 static struct sock *udp4_lib_lookup4(const struct net *net, 518 __be32 saddr, __be16 sport, 519 __be32 daddr, unsigned int hnum, 520 int dif, int sdif, 521 struct udp_table *udptable) 522 { 523 return NULL; 524 } 525 526 static void udp_rehash4(struct udp_table *udptable, struct sock *sk, 527 u16 newhash4) 528 { 529 } 530 531 static void udp_unhash4(struct udp_table *udptable, struct sock *sk) 532 { 533 } 534 #else /* !CONFIG_BASE_SMALL */ 535 static struct sock *udp4_lib_lookup4(const struct net *net, 536 __be32 saddr, __be16 sport, 537 __be32 daddr, unsigned int hnum, 538 int dif, int sdif, 539 struct udp_table *udptable) 540 { 541 const __portpair ports = INET_COMBINED_PORTS(sport, hnum); 542 const struct hlist_nulls_node *node; 543 struct udp_hslot *hslot4; 544 unsigned int hash4, slot; 545 struct udp_sock *up; 546 struct sock *sk; 547 548 hash4 = udp_ehashfn(net, daddr, hnum, saddr, sport); 549 slot = hash4 & udptable->mask; 550 hslot4 = &udptable->hash4[slot]; 551 INET_ADDR_COOKIE(acookie, saddr, daddr); 552 553 begin: 554 /* SLAB_TYPESAFE_BY_RCU not used, so we don't need to touch sk_refcnt */ 555 udp_lrpa_for_each_entry_rcu(up, node, &hslot4->nulls_head) { 556 sk = (struct sock *)up; 557 if (inet_match(net, sk, acookie, ports, dif, sdif)) 558 return sk; 559 } 560 561 /* if the nulls value we got at the end of this lookup is not the 562 * expected one, we must restart lookup. We probably met an item that 563 * was moved to another chain due to rehash. 564 */ 565 if (get_nulls_value(node) != slot) 566 goto begin; 567 568 return NULL; 569 } 570 571 /* udp_rehash4() only checks hslot4, and hash4_cnt is not processed. */ 572 static void udp_rehash4(struct udp_table *udptable, struct sock *sk, 573 u16 newhash4) 574 { 575 struct udp_hslot *hslot4, *nhslot4; 576 577 hslot4 = udp_hashslot4(udptable, udp_sk(sk)->udp_lrpa_hash); 578 nhslot4 = udp_hashslot4(udptable, newhash4); 579 udp_sk(sk)->udp_lrpa_hash = newhash4; 580 581 if (hslot4 != nhslot4) { 582 spin_lock_bh(&hslot4->lock); 583 hlist_nulls_del_init_rcu(&udp_sk(sk)->udp_lrpa_node); 584 hslot4->count--; 585 spin_unlock_bh(&hslot4->lock); 586 587 spin_lock_bh(&nhslot4->lock); 588 hlist_nulls_add_head_rcu(&udp_sk(sk)->udp_lrpa_node, 589 &nhslot4->nulls_head); 590 nhslot4->count++; 591 spin_unlock_bh(&nhslot4->lock); 592 } 593 } 594 595 static void udp_unhash4(struct udp_table *udptable, struct sock *sk) 596 { 597 struct udp_hslot *hslot2, *hslot4; 598 599 if (udp_hashed4(sk)) { 600 hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash); 601 hslot4 = udp_hashslot4(udptable, udp_sk(sk)->udp_lrpa_hash); 602 603 spin_lock(&hslot4->lock); 604 hlist_nulls_del_init_rcu(&udp_sk(sk)->udp_lrpa_node); 605 hslot4->count--; 606 spin_unlock(&hslot4->lock); 607 608 spin_lock(&hslot2->lock); 609 udp_hash4_dec(hslot2); 610 spin_unlock(&hslot2->lock); 611 } 612 } 613 614 void udp_lib_hash4(struct sock *sk, u16 hash) 615 { 616 struct udp_hslot *hslot, *hslot2, *hslot4; 617 struct net *net = sock_net(sk); 618 struct udp_table *udptable; 619 620 /* Connected udp socket can re-connect to another remote address, which 621 * will be handled by rehash. Thus no need to redo hash4 here. 622 */ 623 if (udp_hashed4(sk)) 624 return; 625 626 udptable = net->ipv4.udp_table; 627 hslot = udp_hashslot(udptable, net, udp_sk(sk)->udp_port_hash); 628 hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash); 629 hslot4 = udp_hashslot4(udptable, hash); 630 udp_sk(sk)->udp_lrpa_hash = hash; 631 632 spin_lock_bh(&hslot->lock); 633 if (rcu_access_pointer(sk->sk_reuseport_cb)) 634 reuseport_detach_sock(sk); 635 636 spin_lock(&hslot4->lock); 637 hlist_nulls_add_head_rcu(&udp_sk(sk)->udp_lrpa_node, 638 &hslot4->nulls_head); 639 hslot4->count++; 640 spin_unlock(&hslot4->lock); 641 642 spin_lock(&hslot2->lock); 643 udp_hash4_inc(hslot2); 644 spin_unlock(&hslot2->lock); 645 646 spin_unlock_bh(&hslot->lock); 647 } 648 649 /* call with sock lock */ 650 void udp4_hash4(struct sock *sk) 651 { 652 struct net *net = sock_net(sk); 653 unsigned int hash; 654 655 if (sk_unhashed(sk) || sk->sk_rcv_saddr == htonl(INADDR_ANY)) 656 return; 657 658 hash = udp_ehashfn(net, sk->sk_rcv_saddr, sk->sk_num, 659 sk->sk_daddr, sk->sk_dport); 660 661 udp_lib_hash4(sk, hash); 662 } 663 #endif /* CONFIG_BASE_SMALL */ 664 665 /* UDP is nearly always wildcards out the wazoo, it makes no sense to try 666 * harder than this. -DaveM 667 */ 668 struct sock *__udp4_lib_lookup(const struct net *net, __be32 saddr, 669 __be16 sport, __be32 daddr, __be16 dport, 670 int dif, int sdif, struct sk_buff *skb) 671 { 672 struct udp_table *udptable = net->ipv4.udp_table; 673 unsigned short hnum = ntohs(dport); 674 struct udp_hslot *hslot2; 675 struct sock *result, *sk; 676 unsigned int hash2; 677 678 hash2 = ipv4_portaddr_hash(net, daddr, hnum); 679 hslot2 = udp_hashslot2(udptable, hash2); 680 681 if (udp_has_hash4(hslot2)) { 682 result = udp4_lib_lookup4(net, saddr, sport, daddr, hnum, 683 dif, sdif, udptable); 684 if (result) /* udp4_lib_lookup4 return sk or NULL */ 685 return result; 686 } 687 688 /* Lookup connected or non-wildcard socket */ 689 result = udp4_lib_lookup2(net, saddr, sport, 690 daddr, hnum, dif, sdif, 691 hslot2, skb); 692 if (!IS_ERR_OR_NULL(result) && result->sk_state == TCP_ESTABLISHED) 693 goto done; 694 695 /* Lookup redirect from BPF */ 696 if (static_branch_unlikely(&bpf_sk_lookup_enabled)) { 697 sk = inet_lookup_run_sk_lookup(net, IPPROTO_UDP, skb, sizeof(struct udphdr), 698 saddr, sport, daddr, hnum, dif, 699 udp_ehashfn); 700 if (sk) { 701 result = sk; 702 goto done; 703 } 704 } 705 706 /* Got non-wildcard socket or error on first lookup */ 707 if (result) 708 goto done; 709 710 /* Lookup wildcard sockets */ 711 hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum); 712 hslot2 = udp_hashslot2(udptable, hash2); 713 714 result = udp4_lib_lookup2(net, saddr, sport, 715 htonl(INADDR_ANY), hnum, dif, sdif, 716 hslot2, skb); 717 if (!IS_ERR_OR_NULL(result)) 718 goto done; 719 720 /* Primary hash (destination port) lookup as fallback for this race: 721 * 1. __ip4_datagram_connect() sets sk_rcv_saddr 722 * 2. lookup (this function): new sk_rcv_saddr, hashes not updated yet 723 * 3. rehash operation updating _secondary and four-tuple_ hashes 724 * The primary hash doesn't need an update after 1., so, thanks to this 725 * further step, 1. and 3. don't need to be atomic against the lookup. 726 */ 727 result = udp4_lib_lookup1(net, saddr, sport, daddr, hnum, dif, sdif, 728 udptable); 729 730 done: 731 if (IS_ERR(result)) 732 return NULL; 733 return result; 734 } 735 EXPORT_SYMBOL_GPL(__udp4_lib_lookup); 736 737 static inline struct sock *__udp4_lib_lookup_skb(struct sk_buff *skb, 738 __be16 sport, __be16 dport) 739 { 740 const struct iphdr *iph = ip_hdr(skb); 741 742 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport, 743 iph->daddr, dport, inet_iif(skb), 744 inet_sdif(skb), skb); 745 } 746 747 struct sock *udp4_lib_lookup_skb(const struct sk_buff *skb, 748 __be16 sport, __be16 dport) 749 { 750 const u16 offset = NAPI_GRO_CB(skb)->network_offsets[skb->encapsulation]; 751 const struct iphdr *iph = (struct iphdr *)(skb->data + offset); 752 int iif, sdif; 753 754 inet_get_iif_sdif(skb, &iif, &sdif); 755 756 return __udp4_lib_lookup(dev_net(skb->dev), iph->saddr, sport, 757 iph->daddr, dport, iif, sdif, NULL); 758 } 759 760 /* Must be called under rcu_read_lock(). 761 * Does increment socket refcount. 762 */ 763 #if IS_ENABLED(CONFIG_NF_TPROXY_IPV4) || IS_ENABLED(CONFIG_NF_SOCKET_IPV4) 764 struct sock *udp4_lib_lookup(const struct net *net, __be32 saddr, __be16 sport, 765 __be32 daddr, __be16 dport, int dif) 766 { 767 struct sock *sk; 768 769 sk = __udp4_lib_lookup(net, saddr, sport, daddr, dport, dif, 0, NULL); 770 if (sk && !refcount_inc_not_zero(&sk->sk_refcnt)) 771 sk = NULL; 772 return sk; 773 } 774 EXPORT_SYMBOL_GPL(udp4_lib_lookup); 775 #endif 776 777 static inline bool __udp_is_mcast_sock(struct net *net, const struct sock *sk, 778 __be16 loc_port, __be32 loc_addr, 779 __be16 rmt_port, __be32 rmt_addr, 780 int dif, int sdif, unsigned short hnum) 781 { 782 const struct inet_sock *inet = inet_sk(sk); 783 784 if (!net_eq(sock_net(sk), net) || 785 udp_sk(sk)->udp_port_hash != hnum || 786 (inet->inet_daddr && inet->inet_daddr != rmt_addr) || 787 (inet->inet_dport != rmt_port && inet->inet_dport) || 788 (inet->inet_rcv_saddr && inet->inet_rcv_saddr != loc_addr) || 789 ipv6_only_sock(sk) || 790 !udp_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif)) 791 return false; 792 if (!ip_mc_sf_allow(sk, loc_addr, rmt_addr, dif, sdif)) 793 return false; 794 return true; 795 } 796 797 DEFINE_STATIC_KEY_FALSE(udp_encap_needed_key); 798 799 #if IS_ENABLED(CONFIG_IPV6) 800 DEFINE_STATIC_KEY_FALSE(udpv6_encap_needed_key); 801 #endif 802 803 void udp_encap_enable(void) 804 { 805 static_branch_inc(&udp_encap_needed_key); 806 } 807 EXPORT_SYMBOL(udp_encap_enable); 808 809 void udp_encap_disable(void) 810 { 811 static_branch_dec(&udp_encap_needed_key); 812 } 813 EXPORT_SYMBOL(udp_encap_disable); 814 815 /* Handler for tunnels with arbitrary destination ports: no socket lookup, go 816 * through error handlers in encapsulations looking for a match. 817 */ 818 static int __udp4_lib_err_encap_no_sk(struct sk_buff *skb, u32 info) 819 { 820 int i; 821 822 for (i = 0; i < MAX_IPTUN_ENCAP_OPS; i++) { 823 int (*handler)(struct sk_buff *skb, u32 info); 824 const struct ip_tunnel_encap_ops *encap; 825 826 encap = rcu_dereference(iptun_encaps[i]); 827 if (!encap) 828 continue; 829 handler = encap->err_handler; 830 if (handler && !handler(skb, info)) 831 return 0; 832 } 833 834 return -ENOENT; 835 } 836 837 /* Try to match ICMP errors to UDP tunnels by looking up a socket without 838 * reversing source and destination port: this will match tunnels that force the 839 * same destination port on both endpoints (e.g. VXLAN, GENEVE). Note that 840 * lwtunnels might actually break this assumption by being configured with 841 * different destination ports on endpoints, in this case we won't be able to 842 * trace ICMP messages back to them. 843 * 844 * If this doesn't match any socket, probe tunnels with arbitrary destination 845 * ports (e.g. FoU, GUE): there, the receiving socket is useless, as the port 846 * we've sent packets to won't necessarily match the local destination port. 847 * 848 * Then ask the tunnel implementation to match the error against a valid 849 * association. 850 * 851 * Return an error if we can't find a match, the socket if we need further 852 * processing, zero otherwise. 853 */ 854 static struct sock *__udp4_lib_err_encap(struct net *net, 855 const struct iphdr *iph, 856 struct udphdr *uh, 857 struct sock *sk, 858 struct sk_buff *skb, u32 info) 859 { 860 int (*lookup)(struct sock *sk, struct sk_buff *skb); 861 int network_offset, transport_offset; 862 struct udp_sock *up; 863 864 network_offset = skb_network_offset(skb); 865 transport_offset = skb_transport_offset(skb); 866 867 /* Network header needs to point to the outer IPv4 header inside ICMP */ 868 skb_reset_network_header(skb); 869 870 /* Transport header needs to point to the UDP header */ 871 skb_set_transport_header(skb, iph->ihl << 2); 872 873 if (sk) { 874 up = udp_sk(sk); 875 876 lookup = READ_ONCE(up->encap_err_lookup); 877 if (lookup && lookup(sk, skb)) 878 sk = NULL; 879 880 goto out; 881 } 882 883 sk = __udp4_lib_lookup(net, iph->daddr, uh->source, 884 iph->saddr, uh->dest, skb->dev->ifindex, 0, NULL); 885 if (sk) { 886 up = udp_sk(sk); 887 888 lookup = READ_ONCE(up->encap_err_lookup); 889 if (!lookup || lookup(sk, skb)) 890 sk = NULL; 891 } 892 893 out: 894 if (!sk) 895 sk = ERR_PTR(__udp4_lib_err_encap_no_sk(skb, info)); 896 897 skb_set_transport_header(skb, transport_offset); 898 skb_set_network_header(skb, network_offset); 899 900 return sk; 901 } 902 903 /* 904 * This routine is called by the ICMP module when it gets some 905 * sort of error condition. If err < 0 then the socket should 906 * be closed and the error returned to the user. If err > 0 907 * it's just the icmp type << 8 | icmp code. 908 * Header points to the ip header of the error packet. We move 909 * on past this. Then (as it used to claim before adjustment) 910 * header points to the first 8 bytes of the udp header. We need 911 * to find the appropriate port. 912 */ 913 int udp_err(struct sk_buff *skb, u32 info) 914 { 915 const struct iphdr *iph = (const struct iphdr *)skb->data; 916 const int type = icmp_hdr(skb)->type; 917 const int code = icmp_hdr(skb)->code; 918 struct net *net = dev_net(skb->dev); 919 struct inet_sock *inet; 920 bool tunnel = false; 921 struct udphdr *uh; 922 struct sock *sk; 923 int harderr; 924 int err; 925 926 uh = (struct udphdr *)(skb->data + (iph->ihl << 2)); 927 sk = __udp4_lib_lookup(net, iph->daddr, uh->dest, 928 iph->saddr, uh->source, skb->dev->ifindex, 929 inet_sdif(skb), NULL); 930 931 if (!sk || READ_ONCE(udp_sk(sk)->encap_type)) { 932 /* No socket for error: try tunnels before discarding */ 933 if (static_branch_unlikely(&udp_encap_needed_key)) { 934 sk = __udp4_lib_err_encap(net, iph, uh, sk, skb, info); 935 if (!sk) 936 return 0; 937 } else 938 sk = ERR_PTR(-ENOENT); 939 940 if (IS_ERR(sk)) { 941 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS); 942 return PTR_ERR(sk); 943 } 944 945 tunnel = true; 946 } 947 948 err = 0; 949 harderr = 0; 950 inet = inet_sk(sk); 951 952 switch (type) { 953 default: 954 case ICMP_TIME_EXCEEDED: 955 err = EHOSTUNREACH; 956 break; 957 case ICMP_SOURCE_QUENCH: 958 goto out; 959 case ICMP_PARAMETERPROB: 960 err = EPROTO; 961 harderr = 1; 962 break; 963 case ICMP_DEST_UNREACH: 964 if (code == ICMP_FRAG_NEEDED) { /* Path MTU discovery */ 965 ipv4_sk_update_pmtu(skb, sk, info); 966 if (READ_ONCE(inet->pmtudisc) != IP_PMTUDISC_DONT) { 967 err = EMSGSIZE; 968 harderr = 1; 969 break; 970 } 971 goto out; 972 } 973 err = EHOSTUNREACH; 974 if (code <= NR_ICMP_UNREACH) { 975 harderr = icmp_err_convert[code].fatal; 976 err = icmp_err_convert[code].errno; 977 } 978 break; 979 case ICMP_REDIRECT: 980 ipv4_sk_redirect(skb, sk); 981 goto out; 982 } 983 984 /* 985 * RFC1122: OK. Passes ICMP errors back to application, as per 986 * 4.1.3.3. 987 */ 988 if (tunnel) { 989 /* ...not for tunnels though: we don't have a sending socket */ 990 if (udp_sk(sk)->encap_err_rcv) 991 udp_sk(sk)->encap_err_rcv(sk, skb, err, uh->dest, info, 992 (u8 *)(uh+1)); 993 goto out; 994 } 995 if (!inet_test_bit(RECVERR, sk)) { 996 if (!harderr || sk->sk_state != TCP_ESTABLISHED) 997 goto out; 998 } else 999 ip_icmp_error(sk, skb, err, uh->dest, info, (u8 *)(uh+1)); 1000 1001 sk->sk_err = err; 1002 sk_error_report(sk); 1003 out: 1004 return 0; 1005 } 1006 1007 /* 1008 * Throw away all pending data and cancel the corking. Socket is locked. 1009 */ 1010 void udp_flush_pending_frames(struct sock *sk) 1011 { 1012 struct udp_sock *up = udp_sk(sk); 1013 1014 if (up->pending) { 1015 up->len = 0; 1016 WRITE_ONCE(up->pending, 0); 1017 ip_flush_pending_frames(sk); 1018 } 1019 } 1020 1021 /** 1022 * udp4_hwcsum - handle outgoing HW checksumming 1023 * @skb: sk_buff containing the filled-in UDP header 1024 * (checksum field must be zeroed out) 1025 * @src: source IP address 1026 * @dst: destination IP address 1027 */ 1028 void udp4_hwcsum(struct sk_buff *skb, __be32 src, __be32 dst) 1029 { 1030 struct udphdr *uh = udp_hdr(skb); 1031 int offset = skb_transport_offset(skb); 1032 int len = skb->len - offset; 1033 int hlen = len; 1034 __wsum csum = 0; 1035 1036 if (!skb_has_frag_list(skb)) { 1037 /* 1038 * Only one fragment on the socket. 1039 */ 1040 skb->csum_start = skb_transport_header(skb) - skb->head; 1041 skb->csum_offset = offsetof(struct udphdr, check); 1042 uh->check = ~csum_tcpudp_magic(src, dst, len, 1043 IPPROTO_UDP, 0); 1044 } else { 1045 struct sk_buff *frags; 1046 1047 /* 1048 * HW-checksum won't work as there are two or more 1049 * fragments on the socket so that all csums of sk_buffs 1050 * should be together 1051 */ 1052 skb_walk_frags(skb, frags) { 1053 csum = csum_add(csum, frags->csum); 1054 hlen -= frags->len; 1055 } 1056 1057 csum = skb_checksum(skb, offset, hlen, csum); 1058 skb->ip_summed = CHECKSUM_NONE; 1059 1060 uh->check = csum_tcpudp_magic(src, dst, len, IPPROTO_UDP, csum); 1061 if (uh->check == 0) 1062 uh->check = CSUM_MANGLED_0; 1063 } 1064 } 1065 EXPORT_SYMBOL_GPL(udp4_hwcsum); 1066 1067 /* Function to set UDP checksum for an IPv4 UDP packet. This is intended 1068 * for the simple case like when setting the checksum for a UDP tunnel. 1069 */ 1070 void udp_set_csum(bool nocheck, struct sk_buff *skb, 1071 __be32 saddr, __be32 daddr, int len) 1072 { 1073 struct udphdr *uh = udp_hdr(skb); 1074 1075 if (nocheck) { 1076 uh->check = 0; 1077 } else if (skb_is_gso(skb)) { 1078 uh->check = ~udp_v4_check(len, saddr, daddr, 0); 1079 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 1080 uh->check = 0; 1081 uh->check = udp_v4_check(len, saddr, daddr, lco_csum(skb)); 1082 if (uh->check == 0) 1083 uh->check = CSUM_MANGLED_0; 1084 } else { 1085 skb->ip_summed = CHECKSUM_PARTIAL; 1086 skb->csum_start = skb_transport_header(skb) - skb->head; 1087 skb->csum_offset = offsetof(struct udphdr, check); 1088 uh->check = ~udp_v4_check(len, saddr, daddr, 0); 1089 } 1090 } 1091 EXPORT_SYMBOL(udp_set_csum); 1092 1093 static int udp_send_skb(struct sk_buff *skb, struct flowi4 *fl4, 1094 struct inet_cork *cork) 1095 { 1096 struct sock *sk = skb->sk; 1097 int offset, len, datalen; 1098 struct udphdr *uh; 1099 int err; 1100 1101 offset = skb_transport_offset(skb); 1102 len = skb->len - offset; 1103 datalen = len - sizeof(*uh); 1104 1105 /* 1106 * Create a UDP header 1107 */ 1108 uh = udp_hdr(skb); 1109 uh->source = inet_sk(sk)->inet_sport; 1110 uh->dest = fl4->fl4_dport; 1111 /* Datagram length checked in udp_sendmsg. */ 1112 udp_set_len_short(uh, len); 1113 uh->check = 0; 1114 1115 if (cork->gso_size) { 1116 const int hlen = skb_network_header_len(skb) + 1117 sizeof(struct udphdr); 1118 1119 if (hlen + min(datalen, cork->gso_size) > cork->fragsize) { 1120 kfree_skb(skb); 1121 return -EMSGSIZE; 1122 } 1123 if (datalen > cork->gso_size * UDP_MAX_SEGMENTS) { 1124 kfree_skb(skb); 1125 return -EINVAL; 1126 } 1127 if (sk->sk_no_check_tx) { 1128 kfree_skb(skb); 1129 return -EINVAL; 1130 } 1131 if (dst_xfrm(skb_dst(skb))) { 1132 kfree_skb(skb); 1133 return -EIO; 1134 } 1135 1136 if (datalen > cork->gso_size) { 1137 skb_shinfo(skb)->gso_size = cork->gso_size; 1138 skb_shinfo(skb)->gso_type = SKB_GSO_UDP_L4; 1139 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(datalen, 1140 cork->gso_size); 1141 1142 /* Don't checksum the payload, skb will get segmented */ 1143 goto csum_partial; 1144 } 1145 } 1146 1147 if (sk->sk_no_check_tx) { /* UDP csum off */ 1148 skb->ip_summed = CHECKSUM_NONE; 1149 goto send; 1150 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { /* UDP hardware csum */ 1151 csum_partial: 1152 udp4_hwcsum(skb, fl4->saddr, fl4->daddr); 1153 goto send; 1154 } 1155 1156 /* add protocol-dependent pseudo-header */ 1157 uh->check = csum_tcpudp_magic(fl4->saddr, fl4->daddr, len, 1158 IPPROTO_UDP, udp_csum(skb)); 1159 if (uh->check == 0) 1160 uh->check = CSUM_MANGLED_0; 1161 1162 send: 1163 err = ip_send_skb(sock_net(sk), skb); 1164 if (unlikely(err)) { 1165 if (err == -ENOBUFS && 1166 !inet_test_bit(RECVERR, sk)) { 1167 UDP_INC_STATS(sock_net(sk), UDP_MIB_SNDBUFERRORS); 1168 err = 0; 1169 } 1170 } else { 1171 UDP_INC_STATS(sock_net(sk), UDP_MIB_OUTDATAGRAMS); 1172 } 1173 return err; 1174 } 1175 1176 /* 1177 * Push out all pending data as one UDP datagram. Socket is locked. 1178 */ 1179 int udp_push_pending_frames(struct sock *sk) 1180 { 1181 struct udp_sock *up = udp_sk(sk); 1182 struct inet_sock *inet = inet_sk(sk); 1183 struct flowi4 *fl4 = &inet->cork.fl.u.ip4; 1184 struct sk_buff *skb; 1185 int err = 0; 1186 1187 skb = ip_finish_skb(sk, fl4); 1188 if (!skb) 1189 goto out; 1190 1191 err = udp_send_skb(skb, fl4, &inet->cork.base); 1192 1193 out: 1194 up->len = 0; 1195 WRITE_ONCE(up->pending, 0); 1196 return err; 1197 } 1198 1199 static int __udp_cmsg_send(struct cmsghdr *cmsg, u16 *gso_size) 1200 { 1201 switch (cmsg->cmsg_type) { 1202 case UDP_SEGMENT: 1203 if (cmsg->cmsg_len != CMSG_LEN(sizeof(__u16))) 1204 return -EINVAL; 1205 *gso_size = *(__u16 *)CMSG_DATA(cmsg); 1206 return 0; 1207 default: 1208 return -EINVAL; 1209 } 1210 } 1211 1212 int udp_cmsg_send(struct sock *sk, struct msghdr *msg, u16 *gso_size) 1213 { 1214 struct cmsghdr *cmsg; 1215 bool need_ip = false; 1216 int err; 1217 1218 for_each_cmsghdr(cmsg, msg) { 1219 if (!CMSG_OK(msg, cmsg)) 1220 return -EINVAL; 1221 1222 if (cmsg->cmsg_level != SOL_UDP) { 1223 need_ip = true; 1224 continue; 1225 } 1226 1227 err = __udp_cmsg_send(cmsg, gso_size); 1228 if (err) 1229 return err; 1230 } 1231 1232 return need_ip; 1233 } 1234 1235 int udp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1236 { 1237 int corkreq = udp_test_bit(CORK, sk) || msg->msg_flags & MSG_MORE; 1238 DEFINE_RAW_FLEX(struct ip_options_rcu, opt_copy, opt.__data, 1239 IP_OPTIONS_DATA_FIXED_SIZE); 1240 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name); 1241 int ulen = len, free = 0, connected = 0; 1242 struct inet_sock *inet = inet_sk(sk); 1243 struct udp_sock *up = udp_sk(sk); 1244 __be32 daddr, faddr, saddr; 1245 struct rtable *rt = NULL; 1246 struct flowi4 fl4_stack; 1247 struct ipcm_cookie ipc; 1248 struct sk_buff *skb; 1249 struct flowi4 *fl4; 1250 __be16 dport; 1251 int uc_index; 1252 u8 scope; 1253 int err; 1254 1255 if (len > 0xFFFF) 1256 return -EMSGSIZE; 1257 1258 /* 1259 * Check the flags. 1260 */ 1261 1262 if (msg->msg_flags & MSG_OOB) /* Mirror BSD error message compatibility */ 1263 return -EOPNOTSUPP; 1264 1265 fl4 = &inet->cork.fl.u.ip4; 1266 if (READ_ONCE(up->pending)) { 1267 /* 1268 * There are pending frames. 1269 * The socket lock must be held while it's corked. 1270 */ 1271 lock_sock(sk); 1272 if (likely(up->pending)) { 1273 if (unlikely(up->pending != AF_INET)) { 1274 release_sock(sk); 1275 return -EINVAL; 1276 } 1277 goto do_append_data; 1278 } 1279 release_sock(sk); 1280 } 1281 ulen += sizeof(struct udphdr); 1282 1283 /* 1284 * Get and verify the address. 1285 */ 1286 if (usin) { 1287 if (msg->msg_namelen < sizeof(*usin)) 1288 return -EINVAL; 1289 if (usin->sin_family != AF_INET) { 1290 if (usin->sin_family != AF_UNSPEC) 1291 return -EAFNOSUPPORT; 1292 } 1293 1294 daddr = usin->sin_addr.s_addr; 1295 dport = usin->sin_port; 1296 if (dport == 0) 1297 return -EINVAL; 1298 } else { 1299 if (sk->sk_state != TCP_ESTABLISHED) 1300 return -EDESTADDRREQ; 1301 daddr = inet->inet_daddr; 1302 dport = inet->inet_dport; 1303 /* Open fast path for connected socket. 1304 Route will not be used, if at least one option is set. 1305 */ 1306 connected = 1; 1307 } 1308 1309 ipcm_init_sk(&ipc, inet); 1310 ipc.gso_size = READ_ONCE(up->gso_size); 1311 1312 if (msg->msg_controllen) { 1313 err = udp_cmsg_send(sk, msg, &ipc.gso_size); 1314 if (err > 0) { 1315 err = ip_cmsg_send(sk, msg, &ipc, 1316 sk->sk_family == AF_INET6); 1317 connected = 0; 1318 } 1319 if (unlikely(err < 0)) { 1320 kfree(ipc.opt); 1321 return err; 1322 } 1323 if (ipc.opt) 1324 free = 1; 1325 } 1326 if (!ipc.opt) { 1327 struct ip_options_rcu *inet_opt; 1328 1329 rcu_read_lock(); 1330 inet_opt = rcu_dereference(inet->inet_opt); 1331 if (inet_opt) { 1332 memcpy(opt_copy, inet_opt, 1333 sizeof(*inet_opt) + inet_opt->opt.optlen); 1334 ipc.opt = opt_copy; 1335 } 1336 rcu_read_unlock(); 1337 } 1338 1339 if (cgroup_bpf_enabled(CGROUP_UDP4_SENDMSG) && !connected) { 1340 err = BPF_CGROUP_RUN_PROG_UDP4_SENDMSG_LOCK(sk, 1341 (struct sockaddr *)usin, 1342 &msg->msg_namelen, 1343 &ipc.addr); 1344 if (err) 1345 goto out_free; 1346 if (usin) { 1347 if (usin->sin_port == 0) { 1348 /* BPF program set invalid port. Reject it. */ 1349 err = -EINVAL; 1350 goto out_free; 1351 } 1352 daddr = usin->sin_addr.s_addr; 1353 dport = usin->sin_port; 1354 } 1355 } 1356 1357 saddr = ipc.addr; 1358 ipc.addr = faddr = daddr; 1359 1360 if (ipc.opt && ipc.opt->opt.srr) { 1361 if (!daddr) { 1362 err = -EINVAL; 1363 goto out_free; 1364 } 1365 faddr = ipc.opt->opt.faddr; 1366 connected = 0; 1367 } 1368 scope = ip_sendmsg_scope(inet, &ipc, msg); 1369 if (scope == RT_SCOPE_LINK) 1370 connected = 0; 1371 1372 uc_index = READ_ONCE(inet->uc_index); 1373 if (ipv4_is_multicast(daddr)) { 1374 if (!ipc.oif || netif_index_is_l3_master(sock_net(sk), ipc.oif)) 1375 ipc.oif = READ_ONCE(inet->mc_index); 1376 if (!saddr) 1377 saddr = READ_ONCE(inet->mc_addr); 1378 connected = 0; 1379 } else if (!ipc.oif) { 1380 ipc.oif = uc_index; 1381 } else if (ipv4_is_lbcast(daddr) && uc_index) { 1382 /* oif is set, packet is to local broadcast and 1383 * uc_index is set. oif is most likely set 1384 * by sk_bound_dev_if. If uc_index != oif check if the 1385 * oif is an L3 master and uc_index is an L3 slave. 1386 * If so, we want to allow the send using the uc_index. 1387 */ 1388 if (ipc.oif != uc_index && 1389 ipc.oif == l3mdev_master_ifindex_by_index(sock_net(sk), 1390 uc_index)) { 1391 ipc.oif = uc_index; 1392 } 1393 } 1394 1395 if (connected) 1396 rt = dst_rtable(sk_dst_check(sk, 0)); 1397 1398 if (!rt) { 1399 struct net *net = sock_net(sk); 1400 __u8 flow_flags = inet_sk_flowi_flags(sk); 1401 1402 fl4 = &fl4_stack; 1403 1404 flowi4_init_output(fl4, ipc.oif, ipc.sockc.mark, 1405 ipc.tos & INET_DSCP_MASK, scope, 1406 IPPROTO_UDP, flow_flags, faddr, saddr, 1407 dport, inet->inet_sport, 1408 sk_uid(sk)); 1409 1410 security_sk_classify_flow(sk, flowi4_to_flowi_common(fl4)); 1411 rt = ip_route_output_flow(net, fl4, sk); 1412 if (IS_ERR(rt)) { 1413 err = PTR_ERR(rt); 1414 rt = NULL; 1415 if (err == -ENETUNREACH) 1416 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES); 1417 goto out; 1418 } 1419 1420 err = -EACCES; 1421 if ((rt->rt_flags & RTCF_BROADCAST) && 1422 !sock_flag(sk, SOCK_BROADCAST)) 1423 goto out; 1424 if (connected) 1425 sk_dst_set(sk, dst_clone(&rt->dst)); 1426 } 1427 1428 if (msg->msg_flags&MSG_CONFIRM) 1429 goto do_confirm; 1430 back_from_confirm: 1431 1432 saddr = fl4->saddr; 1433 if (!ipc.addr) 1434 daddr = ipc.addr = fl4->daddr; 1435 1436 /* Lockless fast path for the non-corking case. */ 1437 if (!corkreq) { 1438 struct inet_cork cork; 1439 1440 skb = ip_make_skb(sk, fl4, ip_generic_getfrag, msg, ulen, 1441 sizeof(struct udphdr), &ipc, &rt, 1442 &cork, msg->msg_flags); 1443 err = PTR_ERR(skb); 1444 if (!IS_ERR_OR_NULL(skb)) 1445 err = udp_send_skb(skb, fl4, &cork); 1446 goto out; 1447 } 1448 1449 lock_sock(sk); 1450 if (unlikely(up->pending)) { 1451 /* The socket is already corked while preparing it. */ 1452 /* ... which is an evident application bug. --ANK */ 1453 release_sock(sk); 1454 1455 net_dbg_ratelimited("socket already corked\n"); 1456 err = -EINVAL; 1457 goto out; 1458 } 1459 /* 1460 * Now cork the socket to pend data. 1461 */ 1462 fl4 = &inet->cork.fl.u.ip4; 1463 fl4->daddr = daddr; 1464 fl4->saddr = saddr; 1465 fl4->fl4_dport = dport; 1466 fl4->fl4_sport = inet->inet_sport; 1467 WRITE_ONCE(up->pending, AF_INET); 1468 1469 do_append_data: 1470 up->len += ulen; 1471 err = ip_append_data(sk, fl4, ip_generic_getfrag, msg, ulen, 1472 sizeof(struct udphdr), &ipc, &rt, 1473 corkreq ? msg->msg_flags|MSG_MORE : msg->msg_flags); 1474 if (err) 1475 udp_flush_pending_frames(sk); 1476 else if (!corkreq) 1477 err = udp_push_pending_frames(sk); 1478 else if (unlikely(skb_queue_empty(&sk->sk_write_queue))) 1479 WRITE_ONCE(up->pending, 0); 1480 release_sock(sk); 1481 1482 out: 1483 ip_rt_put(rt); 1484 out_free: 1485 if (free) 1486 kfree(ipc.opt); 1487 if (!err) 1488 return len; 1489 /* 1490 * ENOBUFS = no kernel mem, SOCK_NOSPACE = no sndbuf space. Reporting 1491 * ENOBUFS might not be good (it's not tunable per se), but otherwise 1492 * we don't have a good statistic (IpOutDiscards but it can be too many 1493 * things). We could add another new stat but at least for now that 1494 * seems like overkill. 1495 */ 1496 if (err == -ENOBUFS || test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 1497 UDP_INC_STATS(sock_net(sk), UDP_MIB_SNDBUFERRORS); 1498 1499 return err; 1500 1501 do_confirm: 1502 if (msg->msg_flags & MSG_PROBE) 1503 dst_confirm_neigh(&rt->dst, &fl4->daddr); 1504 if (!(msg->msg_flags&MSG_PROBE) || len) 1505 goto back_from_confirm; 1506 err = 0; 1507 goto out; 1508 } 1509 EXPORT_SYMBOL(udp_sendmsg); 1510 1511 void udp_splice_eof(struct socket *sock) 1512 { 1513 struct sock *sk = sock->sk; 1514 struct udp_sock *up = udp_sk(sk); 1515 1516 if (!READ_ONCE(up->pending) || udp_test_bit(CORK, sk)) 1517 return; 1518 1519 lock_sock(sk); 1520 if (up->pending && !udp_test_bit(CORK, sk)) 1521 udp_push_pending_frames(sk); 1522 release_sock(sk); 1523 } 1524 1525 #define UDP_SKB_IS_STATELESS 0x80000000 1526 1527 /* all head states (dst, sk, nf conntrack) except skb extensions are 1528 * cleared by udp_rcv(). 1529 * 1530 * We need to preserve secpath, if present, to eventually process 1531 * IP_CMSG_PASSSEC at recvmsg() time. 1532 * 1533 * Other extensions can be cleared. 1534 */ 1535 static bool udp_try_make_stateless(struct sk_buff *skb) 1536 { 1537 if (!skb_has_extensions(skb)) 1538 return true; 1539 1540 if (!secpath_exists(skb)) { 1541 skb_ext_reset(skb); 1542 return true; 1543 } 1544 1545 return false; 1546 } 1547 1548 static void udp_set_dev_scratch(struct sk_buff *skb) 1549 { 1550 struct udp_dev_scratch *scratch = udp_skb_scratch(skb); 1551 1552 BUILD_BUG_ON(sizeof(struct udp_dev_scratch) > sizeof(long)); 1553 scratch->_tsize_state = skb->truesize; 1554 #if BITS_PER_LONG == 64 1555 scratch->len = skb->len; 1556 scratch->csum_unnecessary = !!skb_csum_unnecessary(skb); 1557 scratch->is_linear = !skb_is_nonlinear(skb); 1558 #endif 1559 if (udp_try_make_stateless(skb)) 1560 scratch->_tsize_state |= UDP_SKB_IS_STATELESS; 1561 } 1562 1563 static void udp_skb_csum_unnecessary_set(struct sk_buff *skb) 1564 { 1565 /* We come here after udp_lib_checksum_complete() returned 0. 1566 * This means that __skb_checksum_complete() might have 1567 * set skb->csum_valid to 1. 1568 * On 64bit platforms, we can set csum_unnecessary 1569 * to true, but only if the skb is not shared. 1570 */ 1571 #if BITS_PER_LONG == 64 1572 if (!skb_shared(skb)) 1573 udp_skb_scratch(skb)->csum_unnecessary = true; 1574 #endif 1575 } 1576 1577 static int udp_skb_truesize(struct sk_buff *skb) 1578 { 1579 return udp_skb_scratch(skb)->_tsize_state & ~UDP_SKB_IS_STATELESS; 1580 } 1581 1582 static bool udp_skb_has_head_state(struct sk_buff *skb) 1583 { 1584 return !(udp_skb_scratch(skb)->_tsize_state & UDP_SKB_IS_STATELESS); 1585 } 1586 1587 /* fully reclaim rmem/fwd memory allocated for skb */ 1588 static void udp_rmem_release(struct sock *sk, unsigned int size, 1589 int partial, bool rx_queue_lock_held) 1590 { 1591 struct udp_sock *up = udp_sk(sk); 1592 struct sk_buff_head *sk_queue; 1593 unsigned int amt; 1594 1595 if (likely(partial)) { 1596 up->forward_deficit += size; 1597 size = up->forward_deficit; 1598 if (size < READ_ONCE(up->forward_threshold) && 1599 !skb_queue_empty(&up->reader_queue)) 1600 return; 1601 } else { 1602 size += up->forward_deficit; 1603 } 1604 up->forward_deficit = 0; 1605 1606 /* acquire the sk_receive_queue for fwd allocated memory scheduling, 1607 * if the called don't held it already 1608 */ 1609 sk_queue = &sk->sk_receive_queue; 1610 if (!rx_queue_lock_held) 1611 spin_lock(&sk_queue->lock); 1612 1613 amt = (size + sk->sk_forward_alloc - partial) & ~(PAGE_SIZE - 1); 1614 sk_forward_alloc_add(sk, size - amt); 1615 1616 if (amt) 1617 __sk_mem_reduce_allocated(sk, amt >> PAGE_SHIFT); 1618 1619 atomic_sub(size, &sk->sk_rmem_alloc); 1620 1621 /* this can save us from acquiring the rx queue lock on next receive */ 1622 skb_queue_splice_tail_init(sk_queue, &up->reader_queue); 1623 1624 if (!rx_queue_lock_held) 1625 spin_unlock(&sk_queue->lock); 1626 } 1627 1628 /* Note: called with reader_queue.lock held. 1629 * Instead of using skb->truesize here, find a copy of it in skb->dev_scratch 1630 * This avoids a cache line miss while receive_queue lock is held. 1631 * Look at __udp_enqueue_schedule_skb() to find where this copy is done. 1632 */ 1633 void udp_skb_destructor(struct sock *sk, struct sk_buff *skb) 1634 { 1635 prefetch(&skb->data); 1636 udp_rmem_release(sk, udp_skb_truesize(skb), 1, false); 1637 } 1638 1639 /* as above, but the caller held the rx queue lock, too */ 1640 static void udp_skb_dtor_locked(struct sock *sk, struct sk_buff *skb) 1641 { 1642 prefetch(&skb->data); 1643 udp_rmem_release(sk, udp_skb_truesize(skb), 1, true); 1644 } 1645 1646 static int udp_rmem_schedule(struct sock *sk, int size) 1647 { 1648 int delta; 1649 1650 delta = size - sk->sk_forward_alloc; 1651 if (delta > 0 && !__sk_mem_schedule(sk, delta, SK_MEM_RECV)) 1652 return -ENOBUFS; 1653 1654 return 0; 1655 } 1656 1657 int __udp_enqueue_schedule_skb(struct sock *sk, struct sk_buff *skb) 1658 { 1659 struct sk_buff_head *list = &sk->sk_receive_queue; 1660 struct udp_prod_queue *udp_prod_queue; 1661 struct sk_buff *next, *to_drop = NULL; 1662 struct llist_node *ll_list; 1663 unsigned int rmem, rcvbuf; 1664 int size, err = -ENOMEM; 1665 int total_size = 0; 1666 int q_size = 0; 1667 int dropcount; 1668 int nb = 0; 1669 1670 rmem = atomic_read(&sk->sk_rmem_alloc); 1671 rcvbuf = READ_ONCE(sk->sk_rcvbuf); 1672 size = skb->truesize; 1673 1674 udp_prod_queue = &udp_sk(sk)->udp_prod_queue[numa_node_id()]; 1675 1676 rmem += atomic_read(&udp_prod_queue->rmem_alloc); 1677 1678 /* Immediately drop when the receive queue is full. 1679 * Cast to unsigned int performs the boundary check for INT_MAX. 1680 */ 1681 if (rmem + size > rcvbuf) { 1682 if (rcvbuf > INT_MAX >> 1) 1683 goto drop; 1684 1685 /* Accept the packet if queue is empty. */ 1686 if (rmem) 1687 goto drop; 1688 } 1689 1690 /* Under mem pressure, it might be helpful to help udp_recvmsg() 1691 * having linear skbs : 1692 * - Reduce memory overhead and thus increase receive queue capacity 1693 * - Less cache line misses at copyout() time 1694 * - Less work at consume_skb() (less alien page frag freeing) 1695 */ 1696 if (rmem > (rcvbuf >> 1)) { 1697 skb_condense(skb); 1698 size = skb->truesize; 1699 } 1700 1701 udp_set_dev_scratch(skb); 1702 1703 atomic_add(size, &udp_prod_queue->rmem_alloc); 1704 1705 if (!llist_add(&skb->ll_node, &udp_prod_queue->ll_root)) 1706 return 0; 1707 1708 dropcount = sock_flag(sk, SOCK_RXQ_OVFL) ? sk_drops_read(sk) : 0; 1709 1710 spin_lock(&list->lock); 1711 1712 ll_list = llist_del_all(&udp_prod_queue->ll_root); 1713 1714 ll_list = llist_reverse_order(ll_list); 1715 1716 llist_for_each_entry_safe(skb, next, ll_list, ll_node) { 1717 size = udp_skb_truesize(skb); 1718 total_size += size; 1719 err = udp_rmem_schedule(sk, size); 1720 if (unlikely(err)) { 1721 /* Free the skbs outside of locked section. */ 1722 skb->next = to_drop; 1723 to_drop = skb; 1724 continue; 1725 } 1726 1727 q_size += size; 1728 sk_forward_alloc_add(sk, -size); 1729 1730 /* no need to setup a destructor, we will explicitly release the 1731 * forward allocated memory on dequeue 1732 */ 1733 SOCK_SKB_CB(skb)->dropcount = dropcount; 1734 nb++; 1735 __skb_queue_tail(list, skb); 1736 } 1737 1738 atomic_add(q_size, &sk->sk_rmem_alloc); 1739 1740 spin_unlock(&list->lock); 1741 1742 if (!sock_flag(sk, SOCK_DEAD)) { 1743 /* Multiple threads might be blocked in recvmsg(), 1744 * using prepare_to_wait_exclusive(). 1745 */ 1746 while (nb) { 1747 INDIRECT_CALL_1(READ_ONCE(sk->sk_data_ready), 1748 sock_def_readable, sk); 1749 nb--; 1750 } 1751 } 1752 1753 if (unlikely(to_drop)) { 1754 int err_ipv4 = 0; 1755 int err_ipv6 = 0; 1756 1757 for (nb = 0; to_drop != NULL; nb++) { 1758 skb = to_drop; 1759 if (skb->protocol == htons(ETH_P_IP)) 1760 err_ipv4++; 1761 else 1762 err_ipv6++; 1763 to_drop = skb->next; 1764 skb_mark_not_on_list(skb); 1765 sk_skb_reason_drop(sk, skb, SKB_DROP_REASON_PROTO_MEM); 1766 } 1767 numa_drop_add(&udp_sk(sk)->drop_counters, nb); 1768 if (err_ipv4 > 0) { 1769 SNMP_ADD_STATS(__UDPX_MIB(sk, true), UDP_MIB_MEMERRORS, 1770 err_ipv4); 1771 SNMP_ADD_STATS(__UDPX_MIB(sk, true), UDP_MIB_INERRORS, 1772 err_ipv4); 1773 } 1774 if (err_ipv6 > 0) { 1775 SNMP_ADD_STATS(__UDPX_MIB(sk, false), UDP_MIB_MEMERRORS, 1776 err_ipv6); 1777 SNMP_ADD_STATS(__UDPX_MIB(sk, false), UDP_MIB_INERRORS, 1778 err_ipv6); 1779 } 1780 } 1781 1782 atomic_sub(total_size, &udp_prod_queue->rmem_alloc); 1783 1784 return 0; 1785 1786 drop: 1787 udp_drops_inc(sk); 1788 return err; 1789 } 1790 1791 void udp_destruct_common(struct sock *sk) 1792 { 1793 /* reclaim completely the forward allocated memory */ 1794 struct udp_sock *up = udp_sk(sk); 1795 unsigned int total = 0; 1796 struct sk_buff *skb; 1797 1798 skb_queue_splice_tail_init(&sk->sk_receive_queue, &up->reader_queue); 1799 while ((skb = __skb_dequeue(&up->reader_queue)) != NULL) { 1800 total += skb->truesize; 1801 kfree_skb(skb); 1802 } 1803 udp_rmem_release(sk, total, 0, true); 1804 kfree(up->udp_prod_queue); 1805 } 1806 1807 static void udp_destruct_sock(struct sock *sk) 1808 { 1809 udp_destruct_common(sk); 1810 inet_sock_destruct(sk); 1811 } 1812 1813 static int udp_init_sock(struct sock *sk) 1814 { 1815 int res = udp_lib_init_sock(sk); 1816 1817 sk->sk_destruct = udp_destruct_sock; 1818 set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags); 1819 return res; 1820 } 1821 1822 void skb_consume_udp(struct sock *sk, struct sk_buff *skb, int len) 1823 { 1824 if (unlikely(READ_ONCE(udp_sk(sk)->peeking_with_offset))) 1825 sk_peek_offset_bwd(sk, len); 1826 1827 if (!skb_shared(skb)) { 1828 skb_orphan(skb); 1829 skb_attempt_defer_free(skb); 1830 return; 1831 } 1832 1833 if (!skb_unref(skb)) 1834 return; 1835 1836 /* In the more common cases we cleared the head states previously, 1837 * see __udp_queue_rcv_skb(). 1838 */ 1839 if (unlikely(udp_skb_has_head_state(skb))) 1840 skb_release_head_state(skb); 1841 __consume_stateless_skb(skb); 1842 } 1843 1844 static struct sk_buff *__first_packet_length(struct sock *sk, 1845 struct sk_buff_head *rcvq, 1846 unsigned int *total) 1847 { 1848 struct sk_buff *skb; 1849 1850 while ((skb = skb_peek(rcvq)) != NULL) { 1851 if (udp_lib_checksum_complete(skb)) { 1852 struct net *net = sock_net(sk); 1853 1854 __UDP_INC_STATS(net, UDP_MIB_CSUMERRORS); 1855 __UDP_INC_STATS(net, UDP_MIB_INERRORS); 1856 udp_drops_inc(sk); 1857 __skb_unlink(skb, rcvq); 1858 *total += skb->truesize; 1859 kfree_skb_reason(skb, SKB_DROP_REASON_UDP_CSUM); 1860 } else { 1861 udp_skb_csum_unnecessary_set(skb); 1862 break; 1863 } 1864 } 1865 return skb; 1866 } 1867 1868 /** 1869 * first_packet_length - return length of first packet in receive queue 1870 * @sk: socket 1871 * 1872 * Drops all bad checksum frames, until a valid one is found. 1873 * Returns the length of found skb, or -1 if none is found. 1874 */ 1875 static int first_packet_length(struct sock *sk) 1876 { 1877 struct sk_buff_head *rcvq = &udp_sk(sk)->reader_queue; 1878 struct sk_buff_head *sk_queue = &sk->sk_receive_queue; 1879 unsigned int total = 0; 1880 struct sk_buff *skb; 1881 int res; 1882 1883 spin_lock_bh(&rcvq->lock); 1884 skb = __first_packet_length(sk, rcvq, &total); 1885 if (!skb && !skb_queue_empty_lockless(sk_queue)) { 1886 spin_lock(&sk_queue->lock); 1887 skb_queue_splice_tail_init(sk_queue, rcvq); 1888 spin_unlock(&sk_queue->lock); 1889 1890 skb = __first_packet_length(sk, rcvq, &total); 1891 } 1892 res = skb ? skb->len : -1; 1893 if (total) 1894 udp_rmem_release(sk, total, 1, false); 1895 spin_unlock_bh(&rcvq->lock); 1896 return res; 1897 } 1898 1899 /* 1900 * IOCTL requests applicable to the UDP protocol 1901 */ 1902 1903 int udp_ioctl(struct sock *sk, int cmd, int *karg) 1904 { 1905 switch (cmd) { 1906 case SIOCOUTQ: 1907 { 1908 *karg = sk_wmem_alloc_get(sk); 1909 return 0; 1910 } 1911 1912 case SIOCINQ: 1913 { 1914 *karg = max_t(int, 0, first_packet_length(sk)); 1915 return 0; 1916 } 1917 1918 default: 1919 return -ENOIOCTLCMD; 1920 } 1921 1922 return 0; 1923 } 1924 1925 struct sk_buff *__skb_recv_udp(struct sock *sk, unsigned int flags, 1926 int *off, int *err) 1927 { 1928 struct sk_buff_head *sk_queue = &sk->sk_receive_queue; 1929 struct sk_buff_head *queue; 1930 struct sk_buff *last; 1931 long timeo; 1932 int error; 1933 1934 queue = &udp_sk(sk)->reader_queue; 1935 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 1936 do { 1937 struct sk_buff *skb; 1938 1939 error = sock_error(sk); 1940 if (error) 1941 break; 1942 1943 error = -EAGAIN; 1944 do { 1945 spin_lock_bh(&queue->lock); 1946 skb = __skb_try_recv_from_queue(queue, flags, off, err, 1947 &last); 1948 if (skb) { 1949 if (!(flags & MSG_PEEK)) 1950 udp_skb_destructor(sk, skb); 1951 spin_unlock_bh(&queue->lock); 1952 return skb; 1953 } 1954 1955 if (skb_queue_empty_lockless(sk_queue)) { 1956 spin_unlock_bh(&queue->lock); 1957 goto busy_check; 1958 } 1959 1960 /* refill the reader queue and walk it again 1961 * keep both queues locked to avoid re-acquiring 1962 * the sk_receive_queue lock if fwd memory scheduling 1963 * is needed. 1964 */ 1965 spin_lock(&sk_queue->lock); 1966 skb_queue_splice_tail_init(sk_queue, queue); 1967 1968 skb = __skb_try_recv_from_queue(queue, flags, off, err, 1969 &last); 1970 if (skb && !(flags & MSG_PEEK)) 1971 udp_skb_dtor_locked(sk, skb); 1972 spin_unlock(&sk_queue->lock); 1973 spin_unlock_bh(&queue->lock); 1974 if (skb) 1975 return skb; 1976 1977 busy_check: 1978 if (!sk_can_busy_loop(sk)) 1979 break; 1980 1981 sk_busy_loop(sk, flags & MSG_DONTWAIT); 1982 } while (!skb_queue_empty_lockless(sk_queue)); 1983 1984 /* sk_queue is empty, reader_queue may contain peeked packets */ 1985 } while (timeo && 1986 !__skb_wait_for_more_packets(sk, &sk->sk_receive_queue, 1987 &error, &timeo, 1988 (struct sk_buff *)sk_queue)); 1989 1990 *err = error; 1991 return NULL; 1992 } 1993 EXPORT_SYMBOL(__skb_recv_udp); 1994 1995 int udp_read_skb(struct sock *sk, skb_read_actor_t recv_actor) 1996 { 1997 struct sk_buff *skb; 1998 int err; 1999 2000 try_again: 2001 skb = skb_recv_udp(sk, MSG_DONTWAIT, &err); 2002 if (!skb) 2003 return err; 2004 2005 if (udp_lib_checksum_complete(skb)) { 2006 struct net *net = sock_net(sk); 2007 2008 __UDP_INC_STATS(net, UDP_MIB_CSUMERRORS); 2009 __UDP_INC_STATS(net, UDP_MIB_INERRORS); 2010 udp_drops_inc(sk); 2011 kfree_skb_reason(skb, SKB_DROP_REASON_UDP_CSUM); 2012 goto try_again; 2013 } 2014 2015 WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk)); 2016 2017 /* 2018 * skb->dev still aliases the UDP rx dev_scratch (its charge was freed 2019 * on dequeue above); a sockmap verdict program may deref it via 2020 * bpf_sk_lookup_*(), so clear it -> bpf_skc_lookup() uses skb->sk 2021 */ 2022 skb->dev = NULL; 2023 2024 return recv_actor(sk, skb); 2025 } 2026 2027 /* 2028 * This should be easy, if there is something there we 2029 * return it, otherwise we block. 2030 */ 2031 2032 INDIRECT_CALLABLE_SCOPE 2033 int udp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags) 2034 { 2035 DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name); 2036 int off, err, peeking = flags & MSG_PEEK; 2037 struct inet_sock *inet = inet_sk(sk); 2038 struct net *net = sock_net(sk); 2039 bool checksum_valid = false; 2040 unsigned int ulen, copied; 2041 struct sk_buff *skb; 2042 2043 if (flags & MSG_ERRQUEUE) 2044 return ip_recv_error(sk, msg, len); 2045 2046 try_again: 2047 off = sk_peek_offset(sk, flags); 2048 skb = __skb_recv_udp(sk, flags, &off, &err); 2049 if (!skb) 2050 return err; 2051 2052 ulen = udp_skb_len(skb); 2053 copied = len; 2054 if (copied > ulen - off) 2055 copied = ulen - off; 2056 else if (copied < ulen) 2057 msg->msg_flags |= MSG_TRUNC; 2058 2059 /* If checksum is needed at all, try to do it while copying the 2060 * data. If the data is truncated, do it before the copy. 2061 */ 2062 if (copied < ulen || peeking) { 2063 checksum_valid = udp_skb_csum_unnecessary(skb) || 2064 !__udp_lib_checksum_complete(skb); 2065 if (!checksum_valid) 2066 goto csum_copy_err; 2067 } 2068 2069 if (checksum_valid || udp_skb_csum_unnecessary(skb)) { 2070 if (udp_skb_is_linear(skb)) 2071 err = copy_linear_skb(skb, copied, off, &msg->msg_iter); 2072 else 2073 err = skb_copy_datagram_msg(skb, off, msg, copied); 2074 } else { 2075 err = skb_copy_and_csum_datagram_msg(skb, off, msg); 2076 2077 if (err == -EINVAL) 2078 goto csum_copy_err; 2079 } 2080 2081 if (unlikely(err)) { 2082 if (!peeking) { 2083 udp_drops_inc(sk); 2084 UDP_INC_STATS(net, UDP_MIB_INERRORS); 2085 } 2086 kfree_skb(skb); 2087 return err; 2088 } 2089 2090 if (!peeking) 2091 UDP_INC_STATS(net, UDP_MIB_INDATAGRAMS); 2092 2093 sock_recv_cmsgs(msg, sk, skb); 2094 2095 /* Copy the address. */ 2096 if (sin) { 2097 sin->sin_family = AF_INET; 2098 sin->sin_port = udp_hdr(skb)->source; 2099 sin->sin_addr.s_addr = ip_hdr(skb)->saddr; 2100 memset(sin->sin_zero, 0, sizeof(sin->sin_zero)); 2101 msg->msg_namelen = sizeof(*sin); 2102 2103 BPF_CGROUP_RUN_PROG_UDP4_RECVMSG_LOCK(sk, 2104 (struct sockaddr *)sin, 2105 &msg->msg_namelen); 2106 } 2107 2108 if (udp_test_bit(GRO_ENABLED, sk)) 2109 udp_cmsg_recv(msg, sk, skb); 2110 2111 if (inet_cmsg_flags(inet)) 2112 ip_cmsg_recv_offset(msg, sk, skb, sizeof(struct udphdr), off); 2113 2114 err = copied; 2115 if (flags & MSG_TRUNC) 2116 err = ulen; 2117 2118 skb_consume_udp(sk, skb, peeking ? -err : err); 2119 return err; 2120 2121 csum_copy_err: 2122 if (!__sk_queue_drop_skb(sk, &udp_sk(sk)->reader_queue, skb, flags, 2123 udp_skb_destructor)) { 2124 UDP_INC_STATS(net, UDP_MIB_CSUMERRORS); 2125 UDP_INC_STATS(net, UDP_MIB_INERRORS); 2126 } 2127 kfree_skb_reason(skb, SKB_DROP_REASON_UDP_CSUM); 2128 2129 /* starting over for a new packet, but check if we need to yield */ 2130 cond_resched(); 2131 msg->msg_flags &= ~MSG_TRUNC; 2132 goto try_again; 2133 } 2134 2135 int udp_pre_connect(struct sock *sk, struct sockaddr_unsized *uaddr, 2136 int addr_len) 2137 { 2138 /* This check is replicated from __ip4_datagram_connect() and 2139 * intended to prevent BPF program called below from accessing bytes 2140 * that are out of the bound specified by user in addr_len. 2141 */ 2142 if (addr_len < sizeof(struct sockaddr_in)) 2143 return -EINVAL; 2144 2145 return BPF_CGROUP_RUN_PROG_INET4_CONNECT_LOCK(sk, uaddr, &addr_len); 2146 } 2147 2148 static int udp_connect(struct sock *sk, struct sockaddr_unsized *uaddr, 2149 int addr_len) 2150 { 2151 int res; 2152 2153 lock_sock(sk); 2154 res = __ip4_datagram_connect(sk, uaddr, addr_len); 2155 if (!res) 2156 udp4_hash4(sk); 2157 release_sock(sk); 2158 return res; 2159 } 2160 2161 int __udp_disconnect(struct sock *sk, int flags) 2162 { 2163 struct inet_sock *inet = inet_sk(sk); 2164 /* 2165 * 1003.1g - break association. 2166 */ 2167 2168 sk->sk_state = TCP_CLOSE; 2169 inet->inet_daddr = 0; 2170 inet->inet_dport = 0; 2171 sock_rps_reset_rxhash(sk); 2172 sk->sk_bound_dev_if = 0; 2173 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) { 2174 inet_reset_saddr(sk); 2175 if (sk->sk_prot->rehash && 2176 (sk->sk_userlocks & SOCK_BINDPORT_LOCK)) 2177 sk->sk_prot->rehash(sk); 2178 } 2179 2180 if (!(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) { 2181 sk->sk_prot->unhash(sk); 2182 inet->inet_sport = 0; 2183 } 2184 sk_dst_reset(sk); 2185 return 0; 2186 } 2187 EXPORT_SYMBOL(__udp_disconnect); 2188 2189 int udp_disconnect(struct sock *sk, int flags) 2190 { 2191 lock_sock(sk); 2192 __udp_disconnect(sk, flags); 2193 release_sock(sk); 2194 return 0; 2195 } 2196 2197 void udp_lib_unhash(struct sock *sk) 2198 { 2199 if (sk_hashed(sk)) { 2200 struct udp_hslot *hslot, *hslot2; 2201 struct net *net = sock_net(sk); 2202 struct udp_table *udptable; 2203 2204 sock_rps_delete_flow(sk); 2205 udptable = net->ipv4.udp_table; 2206 hslot = udp_hashslot(udptable, net, udp_sk(sk)->udp_port_hash); 2207 hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash); 2208 2209 spin_lock_bh(&hslot->lock); 2210 if (rcu_access_pointer(sk->sk_reuseport_cb)) 2211 reuseport_detach_sock(sk); 2212 if (sk_del_node_init_rcu(sk)) { 2213 hslot->count--; 2214 inet_sk(sk)->inet_num = 0; 2215 sock_prot_inuse_add(net, sk->sk_prot, -1); 2216 2217 spin_lock(&hslot2->lock); 2218 hlist_del_init_rcu(&udp_sk(sk)->udp_portaddr_node); 2219 hslot2->count--; 2220 spin_unlock(&hslot2->lock); 2221 2222 udp_unhash4(udptable, sk); 2223 } 2224 spin_unlock_bh(&hslot->lock); 2225 } 2226 } 2227 2228 /* 2229 * inet_rcv_saddr was changed, we must rehash secondary hash 2230 */ 2231 void udp_lib_rehash(struct sock *sk, u16 newhash, u16 newhash4) 2232 { 2233 if (sk_hashed(sk)) { 2234 struct udp_hslot *hslot, *hslot2, *nhslot2; 2235 struct net *net = sock_net(sk); 2236 struct udp_table *udptable; 2237 2238 udptable = net->ipv4.udp_table; 2239 hslot = udp_hashslot(udptable, net, udp_sk(sk)->udp_port_hash); 2240 hslot2 = udp_hashslot2(udptable, udp_sk(sk)->udp_portaddr_hash); 2241 nhslot2 = udp_hashslot2(udptable, newhash); 2242 2243 if (hslot2 != nhslot2 || 2244 rcu_access_pointer(sk->sk_reuseport_cb)) { 2245 /* we must lock primary chain too */ 2246 spin_lock_bh(&hslot->lock); 2247 if (rcu_access_pointer(sk->sk_reuseport_cb)) 2248 reuseport_detach_sock(sk); 2249 2250 if (hslot2 != nhslot2) { 2251 spin_lock(&hslot2->lock); 2252 hlist_del_init_rcu(&udp_sk(sk)->udp_portaddr_node); 2253 hslot2->count--; 2254 spin_unlock(&hslot2->lock); 2255 2256 spin_lock(&nhslot2->lock); 2257 hlist_add_head_rcu(&udp_sk(sk)->udp_portaddr_node, 2258 &nhslot2->head); 2259 nhslot2->count++; 2260 spin_unlock(&nhslot2->lock); 2261 } 2262 2263 spin_unlock_bh(&hslot->lock); 2264 } 2265 2266 /* Now process hash4 if necessary: 2267 * (1) update hslot4; 2268 * (2) update hslot2->hash4_cnt. 2269 * Note that hslot2/hslot4 should be checked separately, as 2270 * either of them may change with the other unchanged. 2271 */ 2272 if (udp_hashed4(sk)) { 2273 spin_lock_bh(&hslot->lock); 2274 2275 if (inet_rcv_saddr_any(sk)) { 2276 udp_unhash4(udptable, sk); 2277 } else { 2278 udp_rehash4(udptable, sk, newhash4); 2279 if (hslot2 != nhslot2) { 2280 spin_lock(&hslot2->lock); 2281 udp_hash4_dec(hslot2); 2282 spin_unlock(&hslot2->lock); 2283 2284 spin_lock(&nhslot2->lock); 2285 udp_hash4_inc(nhslot2); 2286 spin_unlock(&nhslot2->lock); 2287 } 2288 } 2289 2290 spin_unlock_bh(&hslot->lock); 2291 } 2292 2293 udp_sk(sk)->udp_portaddr_hash = newhash; 2294 } 2295 } 2296 2297 static void udp_v4_rehash(struct sock *sk) 2298 { 2299 u16 new_hash = ipv4_portaddr_hash(sock_net(sk), 2300 inet_sk(sk)->inet_rcv_saddr, 2301 inet_sk(sk)->inet_num); 2302 u16 new_hash4 = udp_ehashfn(sock_net(sk), 2303 sk->sk_rcv_saddr, sk->sk_num, 2304 sk->sk_daddr, sk->sk_dport); 2305 2306 udp_lib_rehash(sk, new_hash, new_hash4); 2307 } 2308 2309 static int __udp_queue_rcv_skb(struct sock *sk, struct sk_buff *skb) 2310 { 2311 int rc; 2312 2313 if (inet_sk(sk)->inet_daddr) { 2314 sock_rps_save_rxhash(sk, skb); 2315 sk_mark_napi_id(sk, skb); 2316 sk_incoming_cpu_update(sk); 2317 } else { 2318 sk_mark_napi_id_once(sk, skb); 2319 } 2320 2321 rc = __udp_enqueue_schedule_skb(sk, skb); 2322 if (rc < 0) { 2323 struct net *net = sock_net(sk); 2324 int drop_reason; 2325 2326 /* Note that an ENOMEM error is charged twice */ 2327 if (rc == -ENOMEM) { 2328 UDP_INC_STATS(net, UDP_MIB_RCVBUFERRORS); 2329 drop_reason = SKB_DROP_REASON_SOCKET_RCVBUFF; 2330 } else { 2331 UDP_INC_STATS(net, UDP_MIB_MEMERRORS); 2332 drop_reason = SKB_DROP_REASON_PROTO_MEM; 2333 } 2334 UDP_INC_STATS(net, UDP_MIB_INERRORS); 2335 trace_udp_fail_queue_rcv_skb(rc, sk, skb); 2336 sk_skb_reason_drop(sk, skb, drop_reason); 2337 return -1; 2338 } 2339 2340 return 0; 2341 } 2342 2343 /* returns: 2344 * -1: error 2345 * 0: success 2346 * >0: "udp encap" protocol resubmission 2347 * 2348 * Note that in the success and error cases, the skb is assumed to 2349 * have either been requeued or freed. 2350 */ 2351 static int udp_queue_rcv_one_skb(struct sock *sk, struct sk_buff *skb) 2352 { 2353 enum skb_drop_reason drop_reason = SKB_DROP_REASON_NOT_SPECIFIED; 2354 struct udp_sock *up = udp_sk(sk); 2355 struct net *net = sock_net(sk); 2356 2357 /* 2358 * Charge it to the socket, dropping if the queue is full. 2359 */ 2360 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb)) { 2361 drop_reason = SKB_DROP_REASON_XFRM_POLICY; 2362 goto drop; 2363 } 2364 nf_reset_ct(skb); 2365 2366 if (static_branch_unlikely(&udp_encap_needed_key) && 2367 READ_ONCE(up->encap_type)) { 2368 int (*encap_rcv)(struct sock *sk, struct sk_buff *skb); 2369 2370 /* 2371 * This is an encapsulation socket so pass the skb to 2372 * the socket's udp_encap_rcv() hook. Otherwise, just 2373 * fall through and pass this up the UDP socket. 2374 * up->encap_rcv() returns the following value: 2375 * =0 if skb was successfully passed to the encap 2376 * handler or was discarded by it. 2377 * >0 if skb should be passed on to UDP. 2378 * <0 if skb should be resubmitted as proto -N 2379 */ 2380 2381 /* if we're overly short, let UDP handle it */ 2382 encap_rcv = READ_ONCE(up->encap_rcv); 2383 if (encap_rcv) { 2384 int ret; 2385 2386 /* Verify checksum before giving to encap */ 2387 if (udp_lib_checksum_complete(skb)) 2388 goto csum_error; 2389 2390 ret = encap_rcv(sk, skb); 2391 if (ret <= 0) { 2392 __UDP_INC_STATS(net, UDP_MIB_INDATAGRAMS); 2393 return -ret; 2394 } 2395 } 2396 2397 /* FALLTHROUGH -- it's a UDP Packet */ 2398 } 2399 2400 prefetch(&sk->sk_rmem_alloc); 2401 if (rcu_access_pointer(sk->sk_filter) && 2402 udp_lib_checksum_complete(skb)) 2403 goto csum_error; 2404 2405 drop_reason = sk_filter_trim_cap(sk, skb, sizeof(struct udphdr)); 2406 if (drop_reason) 2407 goto drop; 2408 2409 udp_csum_pull_header(skb); 2410 2411 ipv4_pktinfo_prepare(sk, skb, true); 2412 return __udp_queue_rcv_skb(sk, skb); 2413 2414 csum_error: 2415 drop_reason = SKB_DROP_REASON_UDP_CSUM; 2416 __UDP_INC_STATS(net, UDP_MIB_CSUMERRORS); 2417 drop: 2418 __UDP_INC_STATS(net, UDP_MIB_INERRORS); 2419 udp_drops_inc(sk); 2420 sk_skb_reason_drop(sk, skb, drop_reason); 2421 return -1; 2422 } 2423 2424 static int udp_queue_rcv_skb(struct sock *sk, struct sk_buff *skb) 2425 { 2426 struct sk_buff *next, *segs; 2427 int ret; 2428 2429 if (likely(!udp_unexpected_gso(sk, skb))) 2430 return udp_queue_rcv_one_skb(sk, skb); 2431 2432 BUILD_BUG_ON(sizeof(struct udp_skb_cb) > SKB_GSO_CB_OFFSET); 2433 __skb_push(skb, -skb_mac_offset(skb)); 2434 segs = udp_rcv_segment(sk, skb, true); 2435 skb_list_walk_safe(segs, skb, next) { 2436 __skb_pull(skb, skb_transport_offset(skb)); 2437 2438 udp_post_segment_fix_csum(skb); 2439 ret = udp_queue_rcv_one_skb(sk, skb); 2440 if (ret > 0) 2441 ip_protocol_deliver_rcu(dev_net(skb->dev), skb, ret); 2442 } 2443 return 0; 2444 } 2445 2446 /* For TCP sockets, sk_rx_dst is protected by socket lock 2447 * For UDP, we use xchg() to guard against concurrent changes. 2448 */ 2449 bool udp_sk_rx_dst_set(struct sock *sk, struct dst_entry *dst) 2450 { 2451 struct dst_entry *old; 2452 2453 if (dst_hold_safe(dst)) { 2454 old = unrcu_pointer(xchg(&sk->sk_rx_dst, RCU_INITIALIZER(dst))); 2455 dst_release(old); 2456 return old != dst; 2457 } 2458 return false; 2459 } 2460 2461 /* 2462 * Multicasts and broadcasts go to each listener. 2463 * 2464 * Note: called only from the BH handler context. 2465 */ 2466 static int __udp4_lib_mcast_deliver(struct net *net, struct sk_buff *skb, 2467 struct udphdr *uh, 2468 __be32 saddr, __be32 daddr) 2469 { 2470 struct udp_table *udptable = net->ipv4.udp_table; 2471 unsigned int hash2, hash2_any, offset; 2472 unsigned short hnum = ntohs(uh->dest); 2473 struct sock *sk, *first = NULL; 2474 int dif = skb->dev->ifindex; 2475 int sdif = inet_sdif(skb); 2476 struct hlist_node *node; 2477 struct udp_hslot *hslot; 2478 struct sk_buff *nskb; 2479 bool use_hash2; 2480 int ret; 2481 2482 hash2_any = 0; 2483 hash2 = 0; 2484 hslot = udp_hashslot(udptable, net, hnum); 2485 use_hash2 = hslot->count > 10; 2486 offset = offsetof(typeof(*sk), sk_node); 2487 2488 if (use_hash2) { 2489 hash2_any = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum) & 2490 udptable->mask; 2491 hash2 = ipv4_portaddr_hash(net, daddr, hnum) & udptable->mask; 2492 start_lookup: 2493 hslot = &udptable->hash2[hash2].hslot; 2494 offset = offsetof(typeof(*sk), __sk_common.skc_portaddr_node); 2495 } 2496 2497 sk_for_each_entry_offset_rcu(sk, node, &hslot->head, offset) { 2498 if (!__udp_is_mcast_sock(net, sk, uh->dest, daddr, 2499 uh->source, saddr, dif, sdif, hnum)) 2500 continue; 2501 2502 if (!first) { 2503 first = sk; 2504 continue; 2505 } 2506 nskb = skb_clone(skb, GFP_ATOMIC); 2507 2508 if (unlikely(!nskb)) { 2509 udp_drops_inc(sk); 2510 __UDP_INC_STATS(net, UDP_MIB_RCVBUFERRORS); 2511 __UDP_INC_STATS(net, UDP_MIB_INERRORS); 2512 continue; 2513 } 2514 if (udp_queue_rcv_skb(sk, nskb) > 0) 2515 consume_skb(nskb); 2516 } 2517 2518 /* Also lookup *:port if we are using hash2 and haven't done so yet. */ 2519 if (use_hash2 && hash2 != hash2_any) { 2520 hash2 = hash2_any; 2521 goto start_lookup; 2522 } 2523 2524 if (first) { 2525 ret = udp_queue_rcv_skb(first, skb); 2526 if (ret > 0) 2527 return -ret; 2528 } else { 2529 kfree_skb(skb); 2530 __UDP_INC_STATS(net, UDP_MIB_IGNOREDMULTI); 2531 } 2532 return 0; 2533 } 2534 2535 /* Initialize UDP checksum. If exited with zero value (success), 2536 * CHECKSUM_UNNECESSARY means, that no more checks are required. 2537 * Otherwise, csum completion requires checksumming packet body, 2538 * including udp header and folding it to skb->csum. 2539 */ 2540 static inline int udp4_csum_init(struct sk_buff *skb, struct udphdr *uh) 2541 { 2542 int err; 2543 2544 /* Note, we are only interested in != 0 or == 0, thus the 2545 * force to int. 2546 */ 2547 err = (__force int)skb_checksum_init_zero_check(skb, IPPROTO_UDP, uh->check, 2548 inet_compute_pseudo); 2549 if (err) 2550 return err; 2551 2552 if (skb->ip_summed == CHECKSUM_COMPLETE && !skb->csum_valid) { 2553 /* If SW calculated the value, we know it's bad */ 2554 if (skb->csum_complete_sw) 2555 return 1; 2556 2557 /* HW says the value is bad. Let's validate that. 2558 * skb->csum is no longer the full packet checksum, 2559 * so don't treat it as such. 2560 */ 2561 skb_checksum_complete_unset(skb); 2562 } 2563 2564 return 0; 2565 } 2566 2567 /* wrapper for udp_queue_rcv_skb taking care of csum conversion and 2568 * return code conversion for ip layer consumption 2569 */ 2570 static int udp_unicast_rcv_skb(struct sock *sk, struct sk_buff *skb, 2571 struct udphdr *uh) 2572 { 2573 int ret; 2574 2575 if (inet_get_convert_csum(sk) && uh->check) 2576 skb_checksum_try_convert(skb, IPPROTO_UDP, inet_compute_pseudo); 2577 2578 ret = udp_queue_rcv_skb(sk, skb); 2579 2580 /* a return value > 0 means to resubmit the input, but 2581 * it wants the return to be -protocol, or 0 2582 */ 2583 if (ret > 0) 2584 return -ret; 2585 return 0; 2586 } 2587 2588 /* 2589 * All we need to do is get the socket, and then do a checksum. 2590 */ 2591 2592 int udp_rcv(struct sk_buff *skb) 2593 { 2594 struct rtable *rt = skb_rtable(skb); 2595 struct net *net = dev_net(skb->dev); 2596 struct sock *sk = NULL; 2597 __be32 saddr, daddr; 2598 unsigned int ulen; 2599 struct udphdr *uh; 2600 bool refcounted; 2601 int drop_reason; 2602 2603 drop_reason = SKB_DROP_REASON_NOT_SPECIFIED; 2604 2605 /* 2606 * Validate the packet. 2607 */ 2608 if (!pskb_may_pull(skb, sizeof(struct udphdr))) 2609 goto drop; /* No space for header. */ 2610 2611 uh = udp_hdr(skb); 2612 ulen = udp_get_len(skb, uh, 0); 2613 saddr = ip_hdr(skb)->saddr; 2614 daddr = ip_hdr(skb)->daddr; 2615 2616 if (ulen > skb->len) 2617 goto short_packet; 2618 2619 if (ulen < sizeof(*uh)) 2620 goto short_packet; 2621 2622 if (ulen < skb->len) { 2623 if (pskb_trim_rcsum(skb, ulen)) 2624 goto short_packet; 2625 2626 uh = udp_hdr(skb); 2627 } 2628 2629 if (udp4_csum_init(skb, uh)) 2630 goto csum_error; 2631 2632 sk = inet_steal_sock(net, skb, sizeof(struct udphdr), saddr, uh->source, daddr, uh->dest, 2633 &refcounted, udp_ehashfn); 2634 if (IS_ERR(sk)) 2635 goto no_sk; 2636 2637 if (sk) { 2638 struct dst_entry *dst = skb_dst(skb); 2639 int ret; 2640 2641 if (unlikely(rcu_dereference(sk->sk_rx_dst) != dst)) 2642 udp_sk_rx_dst_set(sk, dst); 2643 2644 ret = udp_unicast_rcv_skb(sk, skb, uh); 2645 if (refcounted) 2646 sock_put(sk); 2647 return ret; 2648 } 2649 2650 if (rt->rt_flags & (RTCF_BROADCAST|RTCF_MULTICAST)) 2651 return __udp4_lib_mcast_deliver(net, skb, uh, saddr, daddr); 2652 2653 sk = __udp4_lib_lookup_skb(skb, uh->source, uh->dest); 2654 if (sk) 2655 return udp_unicast_rcv_skb(sk, skb, uh); 2656 no_sk: 2657 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) 2658 goto drop; 2659 nf_reset_ct(skb); 2660 2661 /* No socket. Drop packet silently, if checksum is wrong */ 2662 if (udp_lib_checksum_complete(skb)) 2663 goto csum_error; 2664 2665 drop_reason = SKB_DROP_REASON_NO_SOCKET; 2666 __UDP_INC_STATS(net, UDP_MIB_NOPORTS); 2667 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_PORT_UNREACH, 0); 2668 2669 /* 2670 * Hmm. We got an UDP packet to a port to which we 2671 * don't wanna listen. Ignore it. 2672 */ 2673 sk_skb_reason_drop(sk, skb, drop_reason); 2674 return 0; 2675 2676 short_packet: 2677 drop_reason = SKB_DROP_REASON_PKT_TOO_SMALL; 2678 net_dbg_ratelimited("UDP: short packet: From %pI4:%u %d/%d to %pI4:%u\n", 2679 &saddr, ntohs(uh->source), 2680 ulen, skb->len, 2681 &daddr, ntohs(uh->dest)); 2682 goto drop; 2683 2684 csum_error: 2685 /* 2686 * RFC1122: OK. Discards the bad packet silently (as far as 2687 * the network is concerned, anyway) as per 4.1.3.4 (MUST). 2688 */ 2689 drop_reason = SKB_DROP_REASON_UDP_CSUM; 2690 net_dbg_ratelimited("UDP: bad checksum. From %pI4:%u to %pI4:%u ulen %d\n", 2691 &saddr, ntohs(uh->source), &daddr, ntohs(uh->dest), 2692 ulen); 2693 __UDP_INC_STATS(net, UDP_MIB_CSUMERRORS); 2694 drop: 2695 __UDP_INC_STATS(net, UDP_MIB_INERRORS); 2696 sk_skb_reason_drop(sk, skb, drop_reason); 2697 return 0; 2698 } 2699 2700 /* We can only early demux multicast if there is a single matching socket. 2701 * If more than one socket found returns NULL 2702 */ 2703 static struct sock *__udp4_lib_mcast_demux_lookup(struct net *net, 2704 __be16 loc_port, __be32 loc_addr, 2705 __be16 rmt_port, __be32 rmt_addr, 2706 int dif, int sdif) 2707 { 2708 struct udp_table *udptable = net->ipv4.udp_table; 2709 unsigned short hnum = ntohs(loc_port); 2710 struct sock *sk, *result; 2711 struct udp_hslot *hslot; 2712 unsigned int slot; 2713 2714 slot = udp_hashfn(net, hnum, udptable->mask); 2715 hslot = &udptable->hash[slot]; 2716 2717 /* Do not bother scanning a too big list */ 2718 if (hslot->count > 10) 2719 return NULL; 2720 2721 result = NULL; 2722 sk_for_each_rcu(sk, &hslot->head) { 2723 if (__udp_is_mcast_sock(net, sk, loc_port, loc_addr, 2724 rmt_port, rmt_addr, dif, sdif, hnum)) { 2725 if (result) 2726 return NULL; 2727 result = sk; 2728 } 2729 } 2730 2731 return result; 2732 } 2733 2734 /* For unicast we should only early demux connected sockets or we can 2735 * break forwarding setups. The chains here can be long so only check 2736 * if the first socket is an exact match and if not move on. 2737 */ 2738 static struct sock *__udp4_lib_demux_lookup(struct net *net, 2739 __be16 loc_port, __be32 loc_addr, 2740 __be16 rmt_port, __be32 rmt_addr, 2741 int dif, int sdif) 2742 { 2743 struct udp_table *udptable = net->ipv4.udp_table; 2744 INET_ADDR_COOKIE(acookie, rmt_addr, loc_addr); 2745 unsigned short hnum = ntohs(loc_port); 2746 struct udp_hslot *hslot2; 2747 unsigned int hash2; 2748 __portpair ports; 2749 struct sock *sk; 2750 2751 hash2 = ipv4_portaddr_hash(net, loc_addr, hnum); 2752 hslot2 = udp_hashslot2(udptable, hash2); 2753 ports = INET_COMBINED_PORTS(rmt_port, hnum); 2754 2755 udp_portaddr_for_each_entry_rcu(sk, &hslot2->head) { 2756 if (inet_match(net, sk, acookie, ports, dif, sdif)) 2757 return sk; 2758 /* Only check first socket in chain */ 2759 break; 2760 } 2761 return NULL; 2762 } 2763 2764 enum skb_drop_reason udp_v4_early_demux(struct sk_buff *skb) 2765 { 2766 struct net *net = dev_net(skb->dev); 2767 struct in_device *in_dev = NULL; 2768 const struct iphdr *iph; 2769 const struct udphdr *uh; 2770 struct sock *sk = NULL; 2771 struct dst_entry *dst; 2772 int dif = skb->dev->ifindex; 2773 int sdif = inet_sdif(skb); 2774 int ours; 2775 2776 /* validate the packet */ 2777 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct udphdr))) 2778 return SKB_NOT_DROPPED_YET; 2779 2780 iph = ip_hdr(skb); 2781 uh = udp_hdr(skb); 2782 2783 if (skb->pkt_type == PACKET_MULTICAST) { 2784 in_dev = __in_dev_get_rcu(skb->dev); 2785 2786 if (!in_dev) 2787 return SKB_NOT_DROPPED_YET; 2788 2789 ours = ip_check_mc_rcu(in_dev, iph->daddr, iph->saddr, 2790 iph->protocol); 2791 if (!ours) 2792 return SKB_NOT_DROPPED_YET; 2793 2794 sk = __udp4_lib_mcast_demux_lookup(net, uh->dest, iph->daddr, 2795 uh->source, iph->saddr, 2796 dif, sdif); 2797 } else if (skb->pkt_type == PACKET_HOST) { 2798 sk = __udp4_lib_demux_lookup(net, uh->dest, iph->daddr, 2799 uh->source, iph->saddr, dif, sdif); 2800 } 2801 2802 if (!sk) 2803 return SKB_NOT_DROPPED_YET; 2804 2805 skb->sk = sk; 2806 DEBUG_NET_WARN_ON_ONCE(sk_is_refcounted(sk)); 2807 skb->destructor = sock_pfree; 2808 dst = rcu_dereference(sk->sk_rx_dst); 2809 2810 if (dst) 2811 dst = dst_check(dst, 0); 2812 if (dst) { 2813 u32 itag = 0; 2814 2815 /* set noref for now. 2816 * any place which wants to hold dst has to call 2817 * dst_hold_safe() 2818 */ 2819 skb_dst_set_noref(skb, dst); 2820 2821 /* for unconnected multicast sockets we need to validate 2822 * the source on each packet 2823 */ 2824 if (!inet_sk(sk)->inet_daddr && in_dev) 2825 return ip_mc_validate_source(skb, iph->daddr, 2826 iph->saddr, 2827 ip4h_dscp(iph), 2828 skb->dev, in_dev, &itag); 2829 } 2830 return SKB_NOT_DROPPED_YET; 2831 } 2832 2833 static void udp_destroy_sock(struct sock *sk) 2834 { 2835 struct udp_sock *up = udp_sk(sk); 2836 bool slow = lock_sock_fast(sk); 2837 2838 /* protects from races with udp_abort() */ 2839 sock_set_flag(sk, SOCK_DEAD); 2840 udp_flush_pending_frames(sk); 2841 unlock_sock_fast(sk, slow); 2842 if (static_branch_unlikely(&udp_encap_needed_key)) { 2843 if (up->encap_type) { 2844 void (*encap_destroy)(struct sock *sk); 2845 encap_destroy = READ_ONCE(up->encap_destroy); 2846 if (encap_destroy) 2847 encap_destroy(sk); 2848 } 2849 if (udp_test_bit(ENCAP_ENABLED, sk)) { 2850 static_branch_dec(&udp_encap_needed_key); 2851 udp_tunnel_cleanup_gro(sk); 2852 } 2853 } 2854 } 2855 2856 typedef struct sk_buff *(*udp_gro_receive_t)(struct sock *sk, 2857 struct list_head *head, 2858 struct sk_buff *skb); 2859 2860 static void set_xfrm_gro_udp_encap_rcv(__u16 encap_type, unsigned short family, 2861 struct sock *sk) 2862 { 2863 #ifdef CONFIG_XFRM 2864 udp_gro_receive_t new_gro_receive; 2865 2866 if (udp_test_bit(GRO_ENABLED, sk) && encap_type == UDP_ENCAP_ESPINUDP) { 2867 if (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6) 2868 new_gro_receive = xfrm6_gro_udp_encap_rcv; 2869 else 2870 new_gro_receive = xfrm4_gro_udp_encap_rcv; 2871 2872 if (udp_sk(sk)->gro_receive != new_gro_receive) { 2873 /* 2874 * With IPV6_ADDRFORM the gro callback could change 2875 * after being set, unregister the old one, if valid. 2876 */ 2877 if (udp_sk(sk)->gro_receive) 2878 udp_tunnel_update_gro_rcv(sk, false); 2879 2880 WRITE_ONCE(udp_sk(sk)->gro_receive, new_gro_receive); 2881 udp_tunnel_update_gro_rcv(sk, true); 2882 } 2883 } 2884 #endif 2885 } 2886 2887 /* 2888 * Socket option code for UDP 2889 */ 2890 int udp_lib_setsockopt(struct sock *sk, int level, int optname, 2891 sockptr_t optval, unsigned int optlen, 2892 int (*push_pending_frames)(struct sock *)) 2893 { 2894 struct udp_sock *up = udp_sk(sk); 2895 int val, valbool; 2896 int err = 0; 2897 2898 if (level == SOL_SOCKET) { 2899 err = sk_setsockopt(sk, level, optname, optval, optlen); 2900 2901 if (optname == SO_RCVBUF || optname == SO_RCVBUFFORCE) { 2902 sockopt_lock_sock(sk); 2903 /* paired with READ_ONCE in udp_rmem_release() */ 2904 WRITE_ONCE(up->forward_threshold, sk->sk_rcvbuf >> 2); 2905 sockopt_release_sock(sk); 2906 } 2907 return err; 2908 } 2909 2910 if (optlen < sizeof(int)) 2911 return -EINVAL; 2912 2913 if (copy_from_sockptr(&val, optval, sizeof(val))) 2914 return -EFAULT; 2915 2916 valbool = val ? 1 : 0; 2917 2918 switch (optname) { 2919 case UDP_CORK: 2920 if (val != 0) { 2921 udp_set_bit(CORK, sk); 2922 } else { 2923 udp_clear_bit(CORK, sk); 2924 lock_sock(sk); 2925 push_pending_frames(sk); 2926 release_sock(sk); 2927 } 2928 break; 2929 2930 case UDP_ENCAP: 2931 sockopt_lock_sock(sk); 2932 switch (val) { 2933 case 0: 2934 #ifdef CONFIG_XFRM 2935 case UDP_ENCAP_ESPINUDP: 2936 set_xfrm_gro_udp_encap_rcv(val, sk->sk_family, sk); 2937 #if IS_ENABLED(CONFIG_IPV6) 2938 if (sk->sk_family == AF_INET6) 2939 WRITE_ONCE(up->encap_rcv, 2940 xfrm6_udp_encap_rcv); 2941 else 2942 #endif 2943 WRITE_ONCE(up->encap_rcv, 2944 xfrm4_udp_encap_rcv); 2945 #endif 2946 fallthrough; 2947 case UDP_ENCAP_L2TPINUDP: 2948 WRITE_ONCE(up->encap_type, val); 2949 udp_tunnel_encap_enable(sk); 2950 break; 2951 default: 2952 err = -ENOPROTOOPT; 2953 break; 2954 } 2955 sockopt_release_sock(sk); 2956 break; 2957 2958 case UDP_NO_CHECK6_TX: 2959 udp_set_no_check6_tx(sk, valbool); 2960 break; 2961 2962 case UDP_NO_CHECK6_RX: 2963 udp_set_no_check6_rx(sk, valbool); 2964 break; 2965 2966 case UDP_SEGMENT: 2967 if (val < 0 || val > USHRT_MAX) 2968 return -EINVAL; 2969 WRITE_ONCE(up->gso_size, val); 2970 break; 2971 2972 case UDP_GRO: 2973 sockopt_lock_sock(sk); 2974 /* when enabling GRO, accept the related GSO packet type */ 2975 if (valbool) 2976 udp_tunnel_encap_enable(sk); 2977 udp_assign_bit(GRO_ENABLED, sk, valbool); 2978 udp_assign_bit(ACCEPT_L4, sk, valbool); 2979 set_xfrm_gro_udp_encap_rcv(up->encap_type, sk->sk_family, sk); 2980 sockopt_release_sock(sk); 2981 break; 2982 2983 default: 2984 err = -ENOPROTOOPT; 2985 break; 2986 } 2987 2988 return err; 2989 } 2990 2991 static int udp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval, 2992 unsigned int optlen) 2993 { 2994 if (level == SOL_UDP || level == SOL_SOCKET) 2995 return udp_lib_setsockopt(sk, level, optname, 2996 optval, optlen, 2997 udp_push_pending_frames); 2998 return ip_setsockopt(sk, level, optname, optval, optlen); 2999 } 3000 3001 int udp_lib_getsockopt(struct sock *sk, int level, int optname, 3002 sockopt_t *opt) 3003 { 3004 struct udp_sock *up = udp_sk(sk); 3005 int val, len; 3006 3007 len = opt->optlen; 3008 /* keep the check so direct sockopt_t callers stay covered. */ 3009 if (len < 0) 3010 return -EINVAL; 3011 3012 len = min_t(unsigned int, len, sizeof(int)); 3013 3014 switch (optname) { 3015 case UDP_CORK: 3016 val = udp_test_bit(CORK, sk); 3017 break; 3018 3019 case UDP_ENCAP: 3020 val = READ_ONCE(up->encap_type); 3021 break; 3022 3023 case UDP_NO_CHECK6_TX: 3024 val = udp_get_no_check6_tx(sk); 3025 break; 3026 3027 case UDP_NO_CHECK6_RX: 3028 val = udp_get_no_check6_rx(sk); 3029 break; 3030 3031 case UDP_SEGMENT: 3032 val = READ_ONCE(up->gso_size); 3033 break; 3034 3035 case UDP_GRO: 3036 val = udp_test_bit(GRO_ENABLED, sk); 3037 break; 3038 3039 default: 3040 return -ENOPROTOOPT; 3041 } 3042 3043 opt->optlen = len; 3044 if (copy_to_iter(&val, len, &opt->iter_out) != len) 3045 return -EFAULT; 3046 return 0; 3047 } 3048 3049 static int udp_getsockopt(struct sock *sk, int level, int optname, 3050 char __user *optval, int __user *optlen) 3051 { 3052 sockopt_t opt; 3053 int err; 3054 3055 /* 3056 * keep the old __user pointers, until ip_getsockopt() moves 3057 * to sockopt_t 3058 */ 3059 if (level != SOL_UDP) 3060 return ip_getsockopt(sk, level, optname, optval, optlen); 3061 3062 err = sockopt_init_user(&opt, optval, optlen); 3063 if (err) 3064 return err; 3065 3066 err = udp_lib_getsockopt(sk, level, optname, &opt); 3067 if (err) 3068 return err; 3069 3070 /* optval was written by copy_to_iter() in udp_lib_getsockopt() */ 3071 if (put_user(opt.optlen, optlen)) 3072 return -EFAULT; 3073 3074 return 0; 3075 } 3076 3077 /** 3078 * udp_poll - wait for a UDP event. 3079 * @file: - file struct 3080 * @sock: - socket 3081 * @wait: - poll table 3082 * 3083 * This is same as datagram poll, except for the special case of 3084 * blocking sockets. If application is using a blocking fd 3085 * and a packet with checksum error is in the queue; 3086 * then it could get return from select indicating data available 3087 * but then block when reading it. Add special case code 3088 * to work around these arguably broken applications. 3089 */ 3090 __poll_t udp_poll(struct file *file, struct socket *sock, poll_table *wait) 3091 { 3092 __poll_t mask = datagram_poll(file, sock, wait); 3093 struct sock *sk = sock->sk; 3094 3095 if (!skb_queue_empty_lockless(&udp_sk(sk)->reader_queue)) 3096 mask |= EPOLLIN | EPOLLRDNORM; 3097 3098 /* Check for false positives due to checksum errors */ 3099 if ((mask & EPOLLRDNORM) && !(file->f_flags & O_NONBLOCK) && 3100 !(sk->sk_shutdown & RCV_SHUTDOWN) && first_packet_length(sk) == -1) 3101 mask &= ~(EPOLLIN | EPOLLRDNORM); 3102 3103 /* psock ingress_msg queue should not contain any bad checksum frames */ 3104 if (sk_is_readable(sk)) 3105 mask |= EPOLLIN | EPOLLRDNORM; 3106 return mask; 3107 3108 } 3109 3110 int udp_abort(struct sock *sk, int err) 3111 { 3112 if (!has_current_bpf_ctx()) 3113 lock_sock(sk); 3114 3115 /* udp{v6}_destroy_sock() sets it under the sk lock, avoid racing 3116 * with close() 3117 */ 3118 if (sock_flag(sk, SOCK_DEAD)) 3119 goto out; 3120 3121 sk->sk_err = err; 3122 sk_error_report(sk); 3123 __udp_disconnect(sk, 0); 3124 3125 out: 3126 if (!has_current_bpf_ctx()) 3127 release_sock(sk); 3128 3129 return 0; 3130 } 3131 3132 struct proto udp_prot = { 3133 .name = "UDP", 3134 .owner = THIS_MODULE, 3135 .close = udp_lib_close, 3136 .pre_connect = udp_pre_connect, 3137 .connect = udp_connect, 3138 .disconnect = udp_disconnect, 3139 .ioctl = udp_ioctl, 3140 .init = udp_init_sock, 3141 .destroy = udp_destroy_sock, 3142 .setsockopt = udp_setsockopt, 3143 .getsockopt = udp_getsockopt, 3144 .sendmsg = udp_sendmsg, 3145 .recvmsg = udp_recvmsg, 3146 .splice_eof = udp_splice_eof, 3147 .release_cb = ip4_datagram_release_cb, 3148 .hash = udp_lib_hash, 3149 .unhash = udp_lib_unhash, 3150 .rehash = udp_v4_rehash, 3151 .get_port = udp_v4_get_port, 3152 .put_port = udp_lib_unhash, 3153 #ifdef CONFIG_BPF_SYSCALL 3154 .psock_update_sk_prot = udp_bpf_update_proto, 3155 #endif 3156 .memory_allocated = &net_aligned_data.udp_memory_allocated, 3157 .per_cpu_fw_alloc = &udp_memory_per_cpu_fw_alloc, 3158 3159 .sysctl_mem = sysctl_udp_mem, 3160 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_udp_wmem_min), 3161 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_udp_rmem_min), 3162 .obj_size = sizeof(struct udp_sock), 3163 .diag_destroy = udp_abort, 3164 }; 3165 EXPORT_SYMBOL(udp_prot); 3166 3167 /* ------------------------------------------------------------------------ */ 3168 #ifdef CONFIG_PROC_FS 3169 3170 static unsigned short seq_file_family(const struct seq_file *seq); 3171 static bool seq_sk_match(struct seq_file *seq, const struct sock *sk) 3172 { 3173 unsigned short family = seq_file_family(seq); 3174 3175 /* AF_UNSPEC is used as a match all */ 3176 return ((family == AF_UNSPEC || family == sk->sk_family) && 3177 net_eq(sock_net(sk), seq_file_net(seq))); 3178 } 3179 3180 #ifdef CONFIG_BPF_SYSCALL 3181 static const struct seq_operations bpf_iter_udp_seq_ops; 3182 #endif 3183 3184 static struct sock *udp_get_first(struct seq_file *seq, int start) 3185 { 3186 struct udp_iter_state *state = seq->private; 3187 struct net *net = seq_file_net(seq); 3188 struct udp_table *udptable; 3189 struct sock *sk; 3190 3191 udptable = net->ipv4.udp_table; 3192 3193 for (state->bucket = start; state->bucket <= udptable->mask; 3194 ++state->bucket) { 3195 struct udp_hslot *hslot = &udptable->hash[state->bucket]; 3196 3197 if (hlist_empty(&hslot->head)) 3198 continue; 3199 3200 spin_lock_bh(&hslot->lock); 3201 sk_for_each(sk, &hslot->head) { 3202 if (seq_sk_match(seq, sk)) 3203 goto found; 3204 } 3205 spin_unlock_bh(&hslot->lock); 3206 } 3207 sk = NULL; 3208 found: 3209 return sk; 3210 } 3211 3212 static struct sock *udp_get_next(struct seq_file *seq, struct sock *sk) 3213 { 3214 struct udp_iter_state *state = seq->private; 3215 struct net *net = seq_file_net(seq); 3216 struct udp_table *udptable; 3217 3218 do { 3219 sk = sk_next(sk); 3220 } while (sk && !seq_sk_match(seq, sk)); 3221 3222 if (!sk) { 3223 udptable = net->ipv4.udp_table; 3224 3225 if (state->bucket <= udptable->mask) 3226 spin_unlock_bh(&udptable->hash[state->bucket].lock); 3227 3228 return udp_get_first(seq, state->bucket + 1); 3229 } 3230 return sk; 3231 } 3232 3233 static struct sock *udp_get_idx(struct seq_file *seq, loff_t pos) 3234 { 3235 struct sock *sk = udp_get_first(seq, 0); 3236 3237 if (sk) 3238 while (pos && (sk = udp_get_next(seq, sk)) != NULL) 3239 --pos; 3240 return pos ? NULL : sk; 3241 } 3242 3243 void *udp_seq_start(struct seq_file *seq, loff_t *pos) 3244 { 3245 struct udp_iter_state *state = seq->private; 3246 state->bucket = MAX_UDP_PORTS; 3247 3248 return *pos ? udp_get_idx(seq, *pos-1) : SEQ_START_TOKEN; 3249 } 3250 3251 void *udp_seq_next(struct seq_file *seq, void *v, loff_t *pos) 3252 { 3253 struct sock *sk; 3254 3255 if (v == SEQ_START_TOKEN) 3256 sk = udp_get_idx(seq, 0); 3257 else 3258 sk = udp_get_next(seq, v); 3259 3260 ++*pos; 3261 return sk; 3262 } 3263 3264 void udp_seq_stop(struct seq_file *seq, void *v) 3265 { 3266 struct udp_iter_state *state = seq->private; 3267 struct udp_table *udptable; 3268 3269 udptable = seq_file_net(seq)->ipv4.udp_table; 3270 3271 if (state->bucket <= udptable->mask) 3272 spin_unlock_bh(&udptable->hash[state->bucket].lock); 3273 } 3274 3275 /* ------------------------------------------------------------------------ */ 3276 static void udp4_format_sock(struct sock *sp, struct seq_file *f, 3277 int bucket) 3278 { 3279 struct inet_sock *inet = inet_sk(sp); 3280 __be32 dest = inet->inet_daddr; 3281 __be32 src = inet->inet_rcv_saddr; 3282 __u16 destp = ntohs(inet->inet_dport); 3283 __u16 srcp = ntohs(inet->inet_sport); 3284 3285 seq_printf(f, "%5d: %08X:%04X %08X:%04X" 3286 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %llu %d %pK %u", 3287 bucket, src, srcp, dest, destp, sp->sk_state, 3288 sk_wmem_alloc_get(sp), 3289 udp_rqueue_get(sp), 3290 0, 0L, 0, 3291 from_kuid_munged(seq_user_ns(f), sk_uid(sp)), 3292 0, sock_i_ino(sp), 3293 refcount_read(&sp->sk_refcnt), sp, 3294 sk_drops_read(sp)); 3295 } 3296 3297 static int udp4_seq_show(struct seq_file *seq, void *v) 3298 { 3299 seq_setwidth(seq, 127); 3300 if (v == SEQ_START_TOKEN) 3301 seq_puts(seq, " sl local_address rem_address st tx_queue " 3302 "rx_queue tr tm->when retrnsmt uid timeout " 3303 "inode ref pointer drops"); 3304 else { 3305 struct udp_iter_state *state = seq->private; 3306 3307 udp4_format_sock(v, seq, state->bucket); 3308 } 3309 seq_pad(seq, '\n'); 3310 return 0; 3311 } 3312 3313 #ifdef CONFIG_BPF_SYSCALL 3314 struct bpf_iter__udp { 3315 __bpf_md_ptr(struct bpf_iter_meta *, meta); 3316 __bpf_md_ptr(struct udp_sock *, udp_sk); 3317 uid_t uid __aligned(8); 3318 int bucket __aligned(8); 3319 }; 3320 3321 union bpf_udp_iter_batch_item { 3322 struct sock *sk; 3323 __u64 cookie; 3324 }; 3325 3326 struct bpf_udp_iter_state { 3327 struct udp_iter_state state; 3328 unsigned int cur_sk; 3329 unsigned int end_sk; 3330 unsigned int max_sk; 3331 union bpf_udp_iter_batch_item *batch; 3332 }; 3333 3334 static int bpf_iter_udp_realloc_batch(struct bpf_udp_iter_state *iter, 3335 unsigned int new_batch_sz, gfp_t flags); 3336 static struct sock *bpf_iter_udp_resume(struct sock *first_sk, 3337 union bpf_udp_iter_batch_item *cookies, 3338 int n_cookies) 3339 { 3340 struct sock *sk = NULL; 3341 int i; 3342 3343 for (i = 0; i < n_cookies; i++) { 3344 sk = first_sk; 3345 udp_portaddr_for_each_entry_from(sk) 3346 if (cookies[i].cookie == atomic64_read(&sk->sk_cookie)) 3347 goto done; 3348 } 3349 done: 3350 return sk; 3351 } 3352 3353 static struct sock *bpf_iter_udp_batch(struct seq_file *seq) 3354 { 3355 struct bpf_udp_iter_state *iter = seq->private; 3356 struct udp_iter_state *state = &iter->state; 3357 unsigned int find_cookie, end_cookie; 3358 struct net *net = seq_file_net(seq); 3359 struct udp_table *udptable; 3360 unsigned int batch_sks = 0; 3361 int resume_bucket; 3362 int resizes = 0; 3363 struct sock *sk; 3364 int err = 0; 3365 3366 resume_bucket = state->bucket; 3367 3368 /* The current batch is done, so advance the bucket. */ 3369 if (iter->cur_sk == iter->end_sk) 3370 state->bucket++; 3371 3372 udptable = net->ipv4.udp_table; 3373 3374 again: 3375 /* New batch for the next bucket. 3376 * Iterate over the hash table to find a bucket with sockets matching 3377 * the iterator attributes, and return the first matching socket from 3378 * the bucket. The remaining matched sockets from the bucket are batched 3379 * before releasing the bucket lock. This allows BPF programs that are 3380 * called in seq_show to acquire the bucket lock if needed. 3381 */ 3382 find_cookie = iter->cur_sk; 3383 end_cookie = iter->end_sk; 3384 iter->cur_sk = 0; 3385 iter->end_sk = 0; 3386 batch_sks = 0; 3387 3388 for (; state->bucket <= udptable->mask; state->bucket++) { 3389 struct udp_hslot *hslot2 = &udptable->hash2[state->bucket].hslot; 3390 3391 if (hlist_empty(&hslot2->head)) 3392 goto next_bucket; 3393 3394 spin_lock_bh(&hslot2->lock); 3395 sk = hlist_entry_safe(hslot2->head.first, struct sock, 3396 __sk_common.skc_portaddr_node); 3397 /* Resume from the first (in iteration order) unseen socket from 3398 * the last batch that still exists in resume_bucket. Most of 3399 * the time this will just be where the last iteration left off 3400 * in resume_bucket unless that socket disappeared between 3401 * reads. 3402 */ 3403 if (state->bucket == resume_bucket) 3404 sk = bpf_iter_udp_resume(sk, &iter->batch[find_cookie], 3405 end_cookie - find_cookie); 3406 fill_batch: 3407 udp_portaddr_for_each_entry_from(sk) { 3408 if (seq_sk_match(seq, sk)) { 3409 if (iter->end_sk < iter->max_sk) { 3410 sock_hold(sk); 3411 iter->batch[iter->end_sk++].sk = sk; 3412 } 3413 batch_sks++; 3414 } 3415 } 3416 3417 /* Allocate a larger batch and try again. */ 3418 if (unlikely(resizes <= 1 && iter->end_sk && 3419 iter->end_sk != batch_sks)) { 3420 resizes++; 3421 3422 /* First, try with GFP_USER to maximize the chances of 3423 * grabbing more memory. 3424 */ 3425 if (resizes == 1) { 3426 spin_unlock_bh(&hslot2->lock); 3427 err = bpf_iter_udp_realloc_batch(iter, 3428 batch_sks * 3 / 2, 3429 GFP_USER); 3430 if (err) 3431 return ERR_PTR(err); 3432 /* Start over. */ 3433 goto again; 3434 } 3435 3436 /* Next, hold onto the lock, so the bucket doesn't 3437 * change while we get the rest of the sockets. 3438 */ 3439 err = bpf_iter_udp_realloc_batch(iter, batch_sks, 3440 GFP_NOWAIT); 3441 if (err) { 3442 spin_unlock_bh(&hslot2->lock); 3443 return ERR_PTR(err); 3444 } 3445 3446 /* Pick up where we left off. */ 3447 sk = iter->batch[iter->end_sk - 1].sk; 3448 sk = hlist_entry_safe(sk->__sk_common.skc_portaddr_node.next, 3449 struct sock, 3450 __sk_common.skc_portaddr_node); 3451 batch_sks = iter->end_sk; 3452 goto fill_batch; 3453 } 3454 3455 spin_unlock_bh(&hslot2->lock); 3456 3457 if (iter->end_sk) 3458 break; 3459 next_bucket: 3460 resizes = 0; 3461 } 3462 3463 WARN_ON_ONCE(iter->end_sk != batch_sks); 3464 return iter->end_sk ? iter->batch[0].sk : NULL; 3465 } 3466 3467 static void *bpf_iter_udp_seq_next(struct seq_file *seq, void *v, loff_t *pos) 3468 { 3469 struct bpf_udp_iter_state *iter = seq->private; 3470 struct sock *sk; 3471 3472 /* Whenever seq_next() is called, the iter->cur_sk is 3473 * done with seq_show(), so unref the iter->cur_sk. 3474 */ 3475 if (iter->cur_sk < iter->end_sk) 3476 sock_put(iter->batch[iter->cur_sk++].sk); 3477 3478 /* After updating iter->cur_sk, check if there are more sockets 3479 * available in the current bucket batch. 3480 */ 3481 if (iter->cur_sk < iter->end_sk) 3482 sk = iter->batch[iter->cur_sk].sk; 3483 else 3484 /* Prepare a new batch. */ 3485 sk = bpf_iter_udp_batch(seq); 3486 3487 ++*pos; 3488 return sk; 3489 } 3490 3491 static void *bpf_iter_udp_seq_start(struct seq_file *seq, loff_t *pos) 3492 { 3493 /* bpf iter does not support lseek, so it always 3494 * continue from where it was stop()-ped. 3495 */ 3496 if (*pos) 3497 return bpf_iter_udp_batch(seq); 3498 3499 return SEQ_START_TOKEN; 3500 } 3501 3502 static int udp_prog_seq_show(struct bpf_prog *prog, struct bpf_iter_meta *meta, 3503 struct udp_sock *udp_sk, uid_t uid, int bucket) 3504 { 3505 struct bpf_iter__udp ctx; 3506 3507 meta->seq_num--; /* skip SEQ_START_TOKEN */ 3508 ctx.meta = meta; 3509 ctx.udp_sk = udp_sk; 3510 ctx.uid = uid; 3511 ctx.bucket = bucket; 3512 return bpf_iter_run_prog(prog, &ctx); 3513 } 3514 3515 static int bpf_iter_udp_seq_show(struct seq_file *seq, void *v) 3516 { 3517 struct udp_iter_state *state = seq->private; 3518 struct bpf_iter_meta meta; 3519 struct bpf_prog *prog; 3520 struct sock *sk = v; 3521 uid_t uid; 3522 int ret; 3523 3524 if (v == SEQ_START_TOKEN) 3525 return 0; 3526 3527 lock_sock(sk); 3528 3529 if (unlikely(sk_unhashed(sk))) { 3530 ret = SEQ_SKIP; 3531 goto unlock; 3532 } 3533 3534 uid = from_kuid_munged(seq_user_ns(seq), sk_uid(sk)); 3535 meta.seq = seq; 3536 prog = bpf_iter_get_info(&meta, false); 3537 ret = udp_prog_seq_show(prog, &meta, v, uid, state->bucket); 3538 3539 unlock: 3540 release_sock(sk); 3541 return ret; 3542 } 3543 3544 static void bpf_iter_udp_put_batch(struct bpf_udp_iter_state *iter) 3545 { 3546 union bpf_udp_iter_batch_item *item; 3547 unsigned int cur_sk = iter->cur_sk; 3548 __u64 cookie; 3549 3550 /* Remember the cookies of the sockets we haven't seen yet, so we can 3551 * pick up where we left off next time around. 3552 */ 3553 while (cur_sk < iter->end_sk) { 3554 item = &iter->batch[cur_sk++]; 3555 cookie = sock_gen_cookie(item->sk); 3556 sock_put(item->sk); 3557 item->cookie = cookie; 3558 } 3559 } 3560 3561 static void bpf_iter_udp_seq_stop(struct seq_file *seq, void *v) 3562 { 3563 struct bpf_udp_iter_state *iter = seq->private; 3564 struct bpf_iter_meta meta; 3565 struct bpf_prog *prog; 3566 3567 if (!v) { 3568 meta.seq = seq; 3569 prog = bpf_iter_get_info(&meta, true); 3570 if (prog) 3571 (void)udp_prog_seq_show(prog, &meta, v, 0, 0); 3572 } 3573 3574 if (iter->cur_sk < iter->end_sk) 3575 bpf_iter_udp_put_batch(iter); 3576 } 3577 3578 static const struct seq_operations bpf_iter_udp_seq_ops = { 3579 .start = bpf_iter_udp_seq_start, 3580 .next = bpf_iter_udp_seq_next, 3581 .stop = bpf_iter_udp_seq_stop, 3582 .show = bpf_iter_udp_seq_show, 3583 }; 3584 #endif 3585 3586 static unsigned short seq_file_family(const struct seq_file *seq) 3587 { 3588 const struct udp_seq_afinfo *afinfo; 3589 3590 #ifdef CONFIG_BPF_SYSCALL 3591 /* BPF iterator: bpf programs to filter sockets. */ 3592 if (seq->op == &bpf_iter_udp_seq_ops) 3593 return AF_UNSPEC; 3594 #endif 3595 3596 /* Proc fs iterator */ 3597 afinfo = pde_data(file_inode(seq->file)); 3598 return afinfo->family; 3599 } 3600 3601 static const struct seq_operations udp_seq_ops = { 3602 .start = udp_seq_start, 3603 .next = udp_seq_next, 3604 .stop = udp_seq_stop, 3605 .show = udp4_seq_show, 3606 }; 3607 3608 static struct udp_seq_afinfo udp4_seq_afinfo = { 3609 .family = AF_INET, 3610 }; 3611 3612 static int __net_init udp4_proc_init_net(struct net *net) 3613 { 3614 if (!proc_create_net_data("udp", 0444, net->proc_net, &udp_seq_ops, 3615 sizeof(struct udp_iter_state), &udp4_seq_afinfo)) 3616 return -ENOMEM; 3617 return 0; 3618 } 3619 3620 static void __net_exit udp4_proc_exit_net(struct net *net) 3621 { 3622 remove_proc_entry("udp", net->proc_net); 3623 } 3624 3625 static struct pernet_operations udp4_net_ops = { 3626 .init = udp4_proc_init_net, 3627 .exit = udp4_proc_exit_net, 3628 }; 3629 3630 int __init udp4_proc_init(void) 3631 { 3632 return register_pernet_subsys(&udp4_net_ops); 3633 } 3634 3635 void udp4_proc_exit(void) 3636 { 3637 unregister_pernet_subsys(&udp4_net_ops); 3638 } 3639 #endif /* CONFIG_PROC_FS */ 3640 3641 static __initdata unsigned long uhash_entries; 3642 static int __init set_uhash_entries(char *str) 3643 { 3644 ssize_t ret; 3645 3646 if (!str) 3647 return 0; 3648 3649 ret = kstrtoul(str, 0, &uhash_entries); 3650 if (ret) 3651 return 0; 3652 3653 if (uhash_entries && uhash_entries < UDP_HTABLE_SIZE_MIN) 3654 uhash_entries = UDP_HTABLE_SIZE_MIN; 3655 return 1; 3656 } 3657 __setup("uhash_entries=", set_uhash_entries); 3658 3659 static void __init udp_table_init(struct udp_table *table, const char *name) 3660 { 3661 unsigned int i, slot_size; 3662 3663 slot_size = sizeof(struct udp_hslot) + sizeof(struct udp_hslot_main) + 3664 udp_hash4_slot_size(); 3665 table->hash = alloc_large_system_hash(name, 3666 slot_size, 3667 uhash_entries, 3668 21, /* one slot per 2 MB */ 3669 0, 3670 &table->log, 3671 &table->mask, 3672 UDP_HTABLE_SIZE_MIN, 3673 UDP_HTABLE_SIZE_MAX); 3674 3675 table->hash2 = (void *)(table->hash + (table->mask + 1)); 3676 for (i = 0; i <= table->mask; i++) { 3677 INIT_HLIST_HEAD(&table->hash[i].head); 3678 table->hash[i].count = 0; 3679 spin_lock_init(&table->hash[i].lock); 3680 } 3681 for (i = 0; i <= table->mask; i++) { 3682 INIT_HLIST_HEAD(&table->hash2[i].hslot.head); 3683 table->hash2[i].hslot.count = 0; 3684 spin_lock_init(&table->hash2[i].hslot.lock); 3685 } 3686 udp_table_hash4_init(table); 3687 } 3688 3689 u32 udp_flow_hashrnd(void) 3690 { 3691 static u32 hashrnd __read_mostly; 3692 3693 net_get_random_once(&hashrnd, sizeof(hashrnd)); 3694 3695 return hashrnd; 3696 } 3697 EXPORT_SYMBOL(udp_flow_hashrnd); 3698 3699 static void __net_init udp_sysctl_init(struct net *net) 3700 { 3701 net->ipv4.sysctl_udp_rmem_min = PAGE_SIZE; 3702 net->ipv4.sysctl_udp_wmem_min = PAGE_SIZE; 3703 3704 #ifdef CONFIG_NET_L3_MASTER_DEV 3705 net->ipv4.sysctl_udp_l3mdev_accept = 0; 3706 #endif 3707 } 3708 3709 static struct udp_table __net_init *udp_pernet_table_alloc(unsigned int hash_entries) 3710 { 3711 struct udp_table *udptable; 3712 unsigned int slot_size; 3713 int i; 3714 3715 udptable = kmalloc_obj(*udptable); 3716 if (!udptable) 3717 goto out; 3718 3719 slot_size = sizeof(struct udp_hslot) + sizeof(struct udp_hslot_main) + 3720 udp_hash4_slot_size(); 3721 udptable->hash = vmalloc_huge(hash_entries * slot_size, 3722 GFP_KERNEL_ACCOUNT); 3723 if (!udptable->hash) 3724 goto free_table; 3725 3726 udptable->hash2 = (void *)(udptable->hash + hash_entries); 3727 udptable->mask = hash_entries - 1; 3728 udptable->log = ilog2(hash_entries); 3729 3730 for (i = 0; i < hash_entries; i++) { 3731 INIT_HLIST_HEAD(&udptable->hash[i].head); 3732 udptable->hash[i].count = 0; 3733 spin_lock_init(&udptable->hash[i].lock); 3734 3735 INIT_HLIST_HEAD(&udptable->hash2[i].hslot.head); 3736 udptable->hash2[i].hslot.count = 0; 3737 spin_lock_init(&udptable->hash2[i].hslot.lock); 3738 } 3739 udp_table_hash4_init(udptable); 3740 3741 return udptable; 3742 3743 free_table: 3744 kfree(udptable); 3745 out: 3746 return NULL; 3747 } 3748 3749 static void __net_exit udp_pernet_table_free(struct net *net) 3750 { 3751 struct udp_table *udptable = net->ipv4.udp_table; 3752 3753 if (udptable == &udp_table) 3754 return; 3755 3756 kvfree(udptable->hash); 3757 kfree(udptable); 3758 } 3759 3760 static void __net_init udp_set_table(struct net *net) 3761 { 3762 struct udp_table *udptable; 3763 unsigned int hash_entries; 3764 struct net *old_net; 3765 3766 if (net_eq(net, &init_net)) 3767 goto fallback; 3768 3769 old_net = current->nsproxy->net_ns; 3770 hash_entries = READ_ONCE(old_net->ipv4.sysctl_udp_child_hash_entries); 3771 if (!hash_entries) 3772 goto fallback; 3773 3774 /* Set min to keep the bitmap on stack in udp_lib_get_port() */ 3775 if (hash_entries < UDP_HTABLE_SIZE_MIN_PERNET) 3776 hash_entries = UDP_HTABLE_SIZE_MIN_PERNET; 3777 else 3778 hash_entries = roundup_pow_of_two(hash_entries); 3779 3780 udptable = udp_pernet_table_alloc(hash_entries); 3781 if (udptable) { 3782 net->ipv4.udp_table = udptable; 3783 } else { 3784 pr_warn("Failed to allocate UDP hash table (entries: %u) " 3785 "for a netns, fallback to the global one\n", 3786 hash_entries); 3787 fallback: 3788 net->ipv4.udp_table = &udp_table; 3789 } 3790 } 3791 3792 static int __net_init udp_pernet_init(struct net *net) 3793 { 3794 #if IS_ENABLED(CONFIG_NET_UDP_TUNNEL) 3795 int i; 3796 3797 /* No tunnel is configured */ 3798 for (i = 0; i < ARRAY_SIZE(net->ipv4.udp_tunnel_gro); ++i) { 3799 INIT_HLIST_HEAD(&net->ipv4.udp_tunnel_gro[i].list); 3800 RCU_INIT_POINTER(net->ipv4.udp_tunnel_gro[i].sk, NULL); 3801 } 3802 #endif 3803 udp_sysctl_init(net); 3804 udp_set_table(net); 3805 3806 return 0; 3807 } 3808 3809 static void __net_exit udp_pernet_exit(struct net *net) 3810 { 3811 udp_pernet_table_free(net); 3812 } 3813 3814 static struct pernet_operations __net_initdata udp_sysctl_ops = { 3815 .init = udp_pernet_init, 3816 .exit = udp_pernet_exit, 3817 }; 3818 3819 #if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS) 3820 DEFINE_BPF_ITER_FUNC(udp, struct bpf_iter_meta *meta, 3821 struct udp_sock *udp_sk, uid_t uid, int bucket) 3822 3823 static int bpf_iter_udp_realloc_batch(struct bpf_udp_iter_state *iter, 3824 unsigned int new_batch_sz, gfp_t flags) 3825 { 3826 union bpf_udp_iter_batch_item *new_batch; 3827 3828 new_batch = kvmalloc_objs(*new_batch, new_batch_sz, 3829 flags | __GFP_NOWARN); 3830 if (!new_batch) 3831 return -ENOMEM; 3832 3833 if (flags != GFP_NOWAIT) 3834 bpf_iter_udp_put_batch(iter); 3835 3836 memcpy(new_batch, iter->batch, sizeof(*iter->batch) * iter->end_sk); 3837 kvfree(iter->batch); 3838 iter->batch = new_batch; 3839 iter->max_sk = new_batch_sz; 3840 3841 return 0; 3842 } 3843 3844 #define INIT_BATCH_SZ 16 3845 3846 static int bpf_iter_init_udp(void *priv_data, struct bpf_iter_aux_info *aux) 3847 { 3848 struct bpf_udp_iter_state *iter = priv_data; 3849 int ret; 3850 3851 ret = bpf_iter_init_seq_net(priv_data, aux); 3852 if (ret) 3853 return ret; 3854 3855 ret = bpf_iter_udp_realloc_batch(iter, INIT_BATCH_SZ, GFP_USER); 3856 if (ret) 3857 bpf_iter_fini_seq_net(priv_data); 3858 3859 iter->state.bucket = -1; 3860 3861 return ret; 3862 } 3863 3864 static void bpf_iter_fini_udp(void *priv_data) 3865 { 3866 struct bpf_udp_iter_state *iter = priv_data; 3867 3868 bpf_iter_fini_seq_net(priv_data); 3869 kvfree(iter->batch); 3870 } 3871 3872 static const struct bpf_iter_seq_info udp_seq_info = { 3873 .seq_ops = &bpf_iter_udp_seq_ops, 3874 .init_seq_private = bpf_iter_init_udp, 3875 .fini_seq_private = bpf_iter_fini_udp, 3876 .seq_priv_size = sizeof(struct bpf_udp_iter_state), 3877 }; 3878 3879 static struct bpf_iter_reg udp_reg_info = { 3880 .target = "udp", 3881 .ctx_arg_info_size = 1, 3882 .ctx_arg_info = { 3883 { offsetof(struct bpf_iter__udp, udp_sk), 3884 PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED }, 3885 }, 3886 .seq_info = &udp_seq_info, 3887 }; 3888 3889 static void __init bpf_iter_register(void) 3890 { 3891 udp_reg_info.ctx_arg_info[0].btf_id = btf_sock_ids[BTF_SOCK_TYPE_UDP]; 3892 if (bpf_iter_reg_target(&udp_reg_info)) 3893 pr_warn("Warning: could not register bpf iterator udp\n"); 3894 } 3895 #endif 3896 3897 void __init udp_init(void) 3898 { 3899 unsigned long limit; 3900 3901 udp_table_init(&udp_table, "UDP"); 3902 limit = nr_free_buffer_pages() / 8; 3903 limit = max(limit, 128UL); 3904 sysctl_udp_mem[0] = limit / 4 * 3; 3905 sysctl_udp_mem[1] = limit; 3906 sysctl_udp_mem[2] = sysctl_udp_mem[0] * 2; 3907 3908 if (register_pernet_subsys(&udp_sysctl_ops)) 3909 panic("UDP: failed to init sysctl parameters.\n"); 3910 3911 #if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS) 3912 bpf_iter_register(); 3913 #endif 3914 } 3915