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 * Support for INET connection oriented protocols. 8 * 9 * Authors: See the TCP sources 10 */ 11 12 #include <linux/module.h> 13 #include <linux/jhash.h> 14 15 #include <net/inet_connection_sock.h> 16 #include <net/inet_hashtables.h> 17 #include <net/inet_timewait_sock.h> 18 #include <net/ip.h> 19 #include <net/route.h> 20 #include <net/tcp_states.h> 21 #include <net/xfrm.h> 22 #include <net/tcp.h> 23 #include <net/tcp_ecn.h> 24 #include <net/sock_reuseport.h> 25 #include <net/addrconf.h> 26 27 #if IS_ENABLED(CONFIG_IPV6) 28 /* match_sk*_wildcard == true: IPV6_ADDR_ANY equals to any IPv6 addresses 29 * if IPv6 only, and any IPv4 addresses 30 * if not IPv6 only 31 * match_sk*_wildcard == false: addresses must be exactly the same, i.e. 32 * IPV6_ADDR_ANY only equals to IPV6_ADDR_ANY, 33 * and 0.0.0.0 equals to 0.0.0.0 only 34 */ 35 static bool ipv6_rcv_saddr_equal(const struct in6_addr *sk1_rcv_saddr6, 36 const struct in6_addr *sk2_rcv_saddr6, 37 __be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr, 38 bool sk1_ipv6only, bool sk2_ipv6only, 39 bool match_sk1_wildcard, 40 bool match_sk2_wildcard) 41 { 42 int addr_type = ipv6_addr_type(sk1_rcv_saddr6); 43 int addr_type2 = sk2_rcv_saddr6 ? ipv6_addr_type(sk2_rcv_saddr6) : IPV6_ADDR_MAPPED; 44 45 /* if both are mapped, treat as IPv4 */ 46 if (addr_type == IPV6_ADDR_MAPPED && addr_type2 == IPV6_ADDR_MAPPED) { 47 if (!sk2_ipv6only) { 48 if (sk1_rcv_saddr == sk2_rcv_saddr) 49 return true; 50 return (match_sk1_wildcard && !sk1_rcv_saddr) || 51 (match_sk2_wildcard && !sk2_rcv_saddr); 52 } 53 return false; 54 } 55 56 if (addr_type == IPV6_ADDR_ANY && addr_type2 == IPV6_ADDR_ANY) 57 return true; 58 59 if (addr_type2 == IPV6_ADDR_ANY && match_sk2_wildcard && 60 !(sk2_ipv6only && addr_type == IPV6_ADDR_MAPPED)) 61 return true; 62 63 if (addr_type == IPV6_ADDR_ANY && match_sk1_wildcard && 64 !(sk1_ipv6only && addr_type2 == IPV6_ADDR_MAPPED)) 65 return true; 66 67 if (sk2_rcv_saddr6 && 68 ipv6_addr_equal(sk1_rcv_saddr6, sk2_rcv_saddr6)) 69 return true; 70 71 return false; 72 } 73 #endif 74 75 /* match_sk*_wildcard == true: 0.0.0.0 equals to any IPv4 addresses 76 * match_sk*_wildcard == false: addresses must be exactly the same, i.e. 77 * 0.0.0.0 only equals to 0.0.0.0 78 */ 79 static bool ipv4_rcv_saddr_equal(__be32 sk1_rcv_saddr, __be32 sk2_rcv_saddr, 80 bool sk2_ipv6only, bool match_sk1_wildcard, 81 bool match_sk2_wildcard) 82 { 83 if (!sk2_ipv6only) { 84 if (sk1_rcv_saddr == sk2_rcv_saddr) 85 return true; 86 return (match_sk1_wildcard && !sk1_rcv_saddr) || 87 (match_sk2_wildcard && !sk2_rcv_saddr); 88 } 89 return false; 90 } 91 92 bool inet_rcv_saddr_equal(const struct sock *sk, const struct sock *sk2, 93 bool match_wildcard) 94 { 95 #if IS_ENABLED(CONFIG_IPV6) 96 if (sk->sk_family == AF_INET6) 97 return ipv6_rcv_saddr_equal(&sk->sk_v6_rcv_saddr, 98 inet6_rcv_saddr(sk2), 99 sk->sk_rcv_saddr, 100 sk2->sk_rcv_saddr, 101 ipv6_only_sock(sk), 102 ipv6_only_sock(sk2), 103 match_wildcard, 104 match_wildcard); 105 #endif 106 return ipv4_rcv_saddr_equal(sk->sk_rcv_saddr, sk2->sk_rcv_saddr, 107 ipv6_only_sock(sk2), match_wildcard, 108 match_wildcard); 109 } 110 EXPORT_SYMBOL(inet_rcv_saddr_equal); 111 112 bool inet_rcv_saddr_any(const struct sock *sk) 113 { 114 #if IS_ENABLED(CONFIG_IPV6) 115 if (sk->sk_family == AF_INET6) 116 return ipv6_addr_any(&sk->sk_v6_rcv_saddr); 117 #endif 118 return !sk->sk_rcv_saddr; 119 } 120 121 /** 122 * inet_sk_get_local_port_range - fetch ephemeral ports range 123 * @sk: socket 124 * @low: pointer to low port 125 * @high: pointer to high port 126 * 127 * Fetch netns port range (/proc/sys/net/ipv4/ip_local_port_range) 128 * Range can be overridden if socket got IP_LOCAL_PORT_RANGE option. 129 * Returns true if IP_LOCAL_PORT_RANGE was set on this socket. 130 */ 131 bool inet_sk_get_local_port_range(const struct sock *sk, int *low, int *high) 132 { 133 int lo, hi, sk_lo, sk_hi; 134 bool local_range = false; 135 u32 sk_range; 136 137 inet_get_local_port_range(sock_net(sk), &lo, &hi); 138 139 sk_range = READ_ONCE(inet_sk(sk)->local_port_range); 140 if (unlikely(sk_range)) { 141 sk_lo = sk_range & 0xffff; 142 sk_hi = sk_range >> 16; 143 144 if (lo <= sk_lo && sk_lo <= hi) 145 lo = sk_lo; 146 if (lo <= sk_hi && sk_hi <= hi) 147 hi = sk_hi; 148 local_range = true; 149 } 150 151 *low = lo; 152 *high = hi; 153 return local_range; 154 } 155 EXPORT_SYMBOL(inet_sk_get_local_port_range); 156 157 static bool inet_use_bhash2_on_bind(const struct sock *sk) 158 { 159 #if IS_ENABLED(CONFIG_IPV6) 160 if (sk->sk_family == AF_INET6) { 161 if (ipv6_addr_any(&sk->sk_v6_rcv_saddr)) 162 return false; 163 164 if (!ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr)) 165 return true; 166 } 167 #endif 168 return sk->sk_rcv_saddr != htonl(INADDR_ANY); 169 } 170 171 static bool inet_bind_conflict(const struct sock *sk, struct sock *sk2, 172 kuid_t uid, bool relax, 173 bool reuseport_cb_ok, bool reuseport_ok) 174 { 175 int bound_dev_if2; 176 177 if (sk == sk2) 178 return false; 179 180 bound_dev_if2 = READ_ONCE(sk2->sk_bound_dev_if); 181 182 if (!sk->sk_bound_dev_if || !bound_dev_if2 || 183 sk->sk_bound_dev_if == bound_dev_if2) { 184 if (sk->sk_reuse && sk2->sk_reuse && 185 sk2->sk_state != TCP_LISTEN) { 186 if (!relax || (!reuseport_ok && sk->sk_reuseport && 187 sk2->sk_reuseport && reuseport_cb_ok && 188 (sk2->sk_state == TCP_TIME_WAIT || 189 uid_eq(uid, sk_uid(sk2))))) 190 return true; 191 } else if (!reuseport_ok || !sk->sk_reuseport || 192 !sk2->sk_reuseport || !reuseport_cb_ok || 193 (sk2->sk_state != TCP_TIME_WAIT && 194 !uid_eq(uid, sk_uid(sk2)))) { 195 return true; 196 } 197 } 198 return false; 199 } 200 201 static bool __inet_bhash2_conflict(const struct sock *sk, struct sock *sk2, 202 kuid_t uid, bool relax, 203 bool reuseport_cb_ok, bool reuseport_ok) 204 { 205 if (ipv6_only_sock(sk2)) { 206 if (sk->sk_family == AF_INET) 207 return false; 208 209 #if IS_ENABLED(CONFIG_IPV6) 210 if (ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr)) 211 return false; 212 #endif 213 } 214 215 return inet_bind_conflict(sk, sk2, uid, relax, 216 reuseport_cb_ok, reuseport_ok); 217 } 218 219 static bool inet_bhash2_conflict(const struct sock *sk, 220 const struct inet_bind2_bucket *tb2, 221 kuid_t uid, 222 bool relax, bool reuseport_cb_ok, 223 bool reuseport_ok) 224 { 225 struct sock *sk2; 226 227 sk_for_each_bound(sk2, &tb2->owners) { 228 if (__inet_bhash2_conflict(sk, sk2, uid, relax, 229 reuseport_cb_ok, reuseport_ok)) 230 return true; 231 } 232 233 return false; 234 } 235 236 #define sk_for_each_bound_bhash(__sk, __tb2, __tb) \ 237 hlist_for_each_entry(__tb2, &(__tb)->bhash2, bhash_node) \ 238 sk_for_each_bound((__sk), &(__tb2)->owners) 239 240 /* This should be called only when the tb and tb2 hashbuckets' locks are held */ 241 static int inet_csk_bind_conflict(const struct sock *sk, 242 const struct inet_bind_bucket *tb, 243 const struct inet_bind2_bucket *tb2, /* may be null */ 244 bool relax, bool reuseport_ok) 245 { 246 struct sock_reuseport *reuseport_cb; 247 kuid_t uid = sk_uid(sk); 248 bool reuseport_cb_ok; 249 struct sock *sk2; 250 251 rcu_read_lock(); 252 reuseport_cb = rcu_dereference(sk->sk_reuseport_cb); 253 /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */ 254 reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks); 255 rcu_read_unlock(); 256 257 /* Conflicts with an existing IPV6_ADDR_ANY (if ipv6) or INADDR_ANY (if 258 * ipv4) should have been checked already. We need to do these two 259 * checks separately because their spinlocks have to be acquired/released 260 * independently of each other, to prevent possible deadlocks 261 */ 262 if (inet_use_bhash2_on_bind(sk)) 263 return tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, 264 reuseport_cb_ok, reuseport_ok); 265 266 /* Unlike other sk lookup places we do not check 267 * for sk_net here, since _all_ the socks listed 268 * in tb->owners and tb2->owners list belong 269 * to the same net - the one this bucket belongs to. 270 */ 271 sk_for_each_bound_bhash(sk2, tb2, tb) { 272 if (!inet_bind_conflict(sk, sk2, uid, relax, reuseport_cb_ok, reuseport_ok)) 273 continue; 274 275 if (inet_rcv_saddr_equal(sk, sk2, true)) 276 return true; 277 } 278 279 return false; 280 } 281 282 /* Determine if there is a bind conflict with an existing IPV6_ADDR_ANY (if ipv6) or 283 * INADDR_ANY (if ipv4) socket. 284 * 285 * Caller must hold bhash hashbucket lock with local bh disabled, to protect 286 * against concurrent binds on the port for addr any 287 */ 288 static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l3mdev, 289 bool relax, bool reuseport_ok) 290 { 291 const struct net *net = sock_net(sk); 292 struct sock_reuseport *reuseport_cb; 293 struct inet_bind_hashbucket *head2; 294 struct inet_bind2_bucket *tb2; 295 kuid_t uid = sk_uid(sk); 296 bool conflict = false; 297 bool reuseport_cb_ok; 298 299 rcu_read_lock(); 300 reuseport_cb = rcu_dereference(sk->sk_reuseport_cb); 301 /* paired with WRITE_ONCE() in __reuseport_(add|detach)_closed_sock */ 302 reuseport_cb_ok = !reuseport_cb || READ_ONCE(reuseport_cb->num_closed_socks); 303 rcu_read_unlock(); 304 305 head2 = inet_bhash2_addr_any_hashbucket(sk, net, port); 306 307 spin_lock(&head2->lock); 308 309 inet_bind_bucket_for_each(tb2, &head2->chain) { 310 if (!inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk)) 311 continue; 312 313 if (!inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok, reuseport_ok)) 314 continue; 315 316 conflict = true; 317 break; 318 } 319 320 spin_unlock(&head2->lock); 321 322 return conflict; 323 } 324 325 /* 326 * Find an open port number for the socket. Returns with the 327 * inet_bind_hashbucket locks held if successful. 328 */ 329 static struct inet_bind_hashbucket * 330 inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret, 331 struct inet_bind2_bucket **tb2_ret, 332 struct inet_bind_hashbucket **head2_ret, int *port_ret) 333 { 334 struct inet_hashinfo *hinfo = tcp_get_hashinfo(sk); 335 int i, low, high, attempt_half, port, l3mdev; 336 struct inet_bind_hashbucket *head, *head2; 337 struct net *net = sock_net(sk); 338 struct inet_bind2_bucket *tb2; 339 struct inet_bind_bucket *tb; 340 u32 remaining, offset; 341 bool relax = false; 342 343 l3mdev = inet_sk_bound_l3mdev(sk); 344 ports_exhausted: 345 attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0; 346 other_half_scan: 347 inet_sk_get_local_port_range(sk, &low, &high); 348 high++; /* [32768, 60999] -> [32768, 61000[ */ 349 if (high - low < 4) 350 attempt_half = 0; 351 if (attempt_half) { 352 int half = low + (((high - low) >> 2) << 1); 353 354 if (attempt_half == 1) 355 high = half; 356 else 357 low = half; 358 } 359 remaining = high - low; 360 if (likely(remaining > 1)) 361 remaining &= ~1U; 362 363 offset = get_random_u32_below(remaining); 364 /* __inet_hash_connect() favors ports having @low parity 365 * We do the opposite to not pollute connect() users. 366 */ 367 offset |= 1U; 368 369 other_parity_scan: 370 port = low + offset; 371 for (i = 0; i < remaining; i += 2, port += 2) { 372 if (unlikely(port >= high)) 373 port -= remaining; 374 if (inet_is_local_reserved_port(net, port)) 375 continue; 376 head = &hinfo->bhash[inet_bhashfn(net, port, 377 hinfo->bhash_size)]; 378 spin_lock_bh(&head->lock); 379 if (inet_use_bhash2_on_bind(sk)) { 380 if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, relax, false)) 381 goto next_port; 382 } 383 384 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port); 385 spin_lock(&head2->lock); 386 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk); 387 inet_bind_bucket_for_each(tb, &head->chain) 388 if (inet_bind_bucket_match(tb, net, port, l3mdev)) { 389 if (!inet_csk_bind_conflict(sk, tb, tb2, 390 relax, false)) 391 goto success; 392 spin_unlock(&head2->lock); 393 goto next_port; 394 } 395 tb = NULL; 396 goto success; 397 next_port: 398 spin_unlock_bh(&head->lock); 399 cond_resched(); 400 } 401 402 offset--; 403 if (!(offset & 1)) 404 goto other_parity_scan; 405 406 if (attempt_half == 1) { 407 /* OK we now try the upper half of the range */ 408 attempt_half = 2; 409 goto other_half_scan; 410 } 411 412 if (READ_ONCE(net->ipv4.sysctl_ip_autobind_reuse) && !relax) { 413 /* We still have a chance to connect to different destinations */ 414 relax = true; 415 goto ports_exhausted; 416 } 417 return NULL; 418 success: 419 *port_ret = port; 420 *tb_ret = tb; 421 *tb2_ret = tb2; 422 *head2_ret = head2; 423 return head; 424 } 425 426 static inline int sk_reuseport_match(struct inet_bind_bucket *tb, 427 const struct sock *sk) 428 { 429 if (tb->fastreuseport <= 0) 430 return 0; 431 if (!sk->sk_reuseport) 432 return 0; 433 if (rcu_access_pointer(sk->sk_reuseport_cb)) 434 return 0; 435 if (!uid_eq(tb->fastuid, sk_uid(sk))) 436 return 0; 437 /* We only need to check the rcv_saddr if this tb was once marked 438 * without fastreuseport and then was reset, as we can only know that 439 * the fast_*rcv_saddr doesn't have any conflicts with the socks on the 440 * owners list. 441 */ 442 if (tb->fastreuseport == FASTREUSEPORT_ANY) 443 return 1; 444 #if IS_ENABLED(CONFIG_IPV6) 445 if (tb->fast_sk_family == AF_INET6) 446 return ipv6_rcv_saddr_equal(&tb->fast_v6_rcv_saddr, 447 inet6_rcv_saddr(sk), 448 tb->fast_rcv_saddr, 449 sk->sk_rcv_saddr, 450 tb->fast_ipv6_only, 451 ipv6_only_sock(sk), true, false); 452 #endif 453 return ipv4_rcv_saddr_equal(tb->fast_rcv_saddr, sk->sk_rcv_saddr, 454 ipv6_only_sock(sk), true, false); 455 } 456 457 void inet_csk_update_fastreuse(const struct sock *sk, 458 struct inet_bind_bucket *tb, 459 struct inet_bind2_bucket *tb2) 460 { 461 bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN; 462 463 if (hlist_empty(&tb->bhash2)) { 464 tb->fastreuse = reuse; 465 if (sk->sk_reuseport) { 466 tb->fastreuseport = FASTREUSEPORT_ANY; 467 tb->fastuid = sk_uid(sk); 468 tb->fast_rcv_saddr = sk->sk_rcv_saddr; 469 tb->fast_ipv6_only = ipv6_only_sock(sk); 470 tb->fast_sk_family = sk->sk_family; 471 #if IS_ENABLED(CONFIG_IPV6) 472 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 473 #endif 474 } else { 475 tb->fastreuseport = 0; 476 } 477 } else { 478 if (!reuse) 479 tb->fastreuse = 0; 480 if (sk->sk_reuseport) { 481 /* We didn't match or we don't have fastreuseport set on 482 * the tb, but we have sk_reuseport set on this socket 483 * and we know that there are no bind conflicts with 484 * this socket in this tb, so reset our tb's reuseport 485 * settings so that any subsequent sockets that match 486 * our current socket will be put on the fast path. 487 * 488 * If we reset we need to set FASTREUSEPORT_STRICT so we 489 * do extra checking for all subsequent sk_reuseport 490 * socks. 491 */ 492 if (!sk_reuseport_match(tb, sk)) { 493 tb->fastreuseport = FASTREUSEPORT_STRICT; 494 tb->fastuid = sk_uid(sk); 495 tb->fast_rcv_saddr = sk->sk_rcv_saddr; 496 tb->fast_ipv6_only = ipv6_only_sock(sk); 497 tb->fast_sk_family = sk->sk_family; 498 #if IS_ENABLED(CONFIG_IPV6) 499 tb->fast_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 500 #endif 501 } 502 } else { 503 tb->fastreuseport = 0; 504 } 505 } 506 507 tb2->fastreuse = tb->fastreuse; 508 tb2->fastreuseport = tb->fastreuseport; 509 } 510 511 /* Obtain a reference to a local port for the given sock, 512 * if snum is zero it means select any available local port. 513 * We try to allocate an odd port (and leave even ports for connect()) 514 */ 515 int inet_csk_get_port(struct sock *sk, unsigned short snum) 516 { 517 bool reuse = sk->sk_reuse && sk->sk_state != TCP_LISTEN; 518 bool found_port = false, check_bind_conflict = true; 519 bool bhash_created = false, bhash2_created = false; 520 struct inet_hashinfo *hinfo = tcp_get_hashinfo(sk); 521 int ret = -EADDRINUSE, port = snum, l3mdev; 522 struct inet_bind_hashbucket *head, *head2; 523 struct inet_bind2_bucket *tb2 = NULL; 524 struct inet_bind_bucket *tb = NULL; 525 bool head2_lock_acquired = false; 526 struct net *net = sock_net(sk); 527 528 l3mdev = inet_sk_bound_l3mdev(sk); 529 530 if (!port) { 531 head = inet_csk_find_open_port(sk, &tb, &tb2, &head2, &port); 532 if (!head) 533 return ret; 534 535 head2_lock_acquired = true; 536 537 if (tb && tb2) 538 goto success; 539 found_port = true; 540 } else { 541 head = &hinfo->bhash[inet_bhashfn(net, port, 542 hinfo->bhash_size)]; 543 spin_lock_bh(&head->lock); 544 inet_bind_bucket_for_each(tb, &head->chain) 545 if (inet_bind_bucket_match(tb, net, port, l3mdev)) 546 break; 547 } 548 549 if (!tb) { 550 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep, net, 551 head, port, l3mdev); 552 if (!tb) 553 goto fail_unlock; 554 bhash_created = true; 555 } 556 557 if (!found_port) { 558 if (!hlist_empty(&tb->bhash2)) { 559 if (sk->sk_reuse == SK_FORCE_REUSE || 560 (tb->fastreuse > 0 && reuse) || 561 sk_reuseport_match(tb, sk)) 562 check_bind_conflict = false; 563 } 564 565 if (check_bind_conflict && inet_use_bhash2_on_bind(sk)) { 566 if (inet_bhash2_addr_any_conflict(sk, port, l3mdev, true, true)) 567 goto fail_unlock; 568 } 569 570 head2 = inet_bhashfn_portaddr(hinfo, sk, net, port); 571 spin_lock(&head2->lock); 572 head2_lock_acquired = true; 573 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk); 574 } 575 576 if (!tb2) { 577 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, 578 net, head2, tb, sk); 579 if (!tb2) 580 goto fail_unlock; 581 bhash2_created = true; 582 } 583 584 if (!found_port && check_bind_conflict) { 585 if (inet_csk_bind_conflict(sk, tb, tb2, true, true)) 586 goto fail_unlock; 587 } 588 589 success: 590 inet_csk_update_fastreuse(sk, tb, tb2); 591 592 if (!inet_csk(sk)->icsk_bind_hash) 593 inet_bind_hash(sk, tb, tb2, port); 594 WARN_ON(inet_csk(sk)->icsk_bind_hash != tb); 595 WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2); 596 ret = 0; 597 598 fail_unlock: 599 if (ret) { 600 if (bhash2_created) 601 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep, tb2); 602 if (bhash_created) 603 inet_bind_bucket_destroy(tb); 604 } 605 if (head2_lock_acquired) 606 spin_unlock(&head2->lock); 607 spin_unlock_bh(&head->lock); 608 return ret; 609 } 610 EXPORT_SYMBOL_GPL(inet_csk_get_port); 611 612 /* 613 * Wait for an incoming connection, avoid race conditions. This must be called 614 * with the socket locked. 615 */ 616 static int inet_csk_wait_for_connect(struct sock *sk, long timeo) 617 { 618 struct inet_connection_sock *icsk = inet_csk(sk); 619 DEFINE_WAIT(wait); 620 int err; 621 622 /* 623 * True wake-one mechanism for incoming connections: only 624 * one process gets woken up, not the 'whole herd'. 625 * Since we do not 'race & poll' for established sockets 626 * anymore, the common case will execute the loop only once. 627 * 628 * Subtle issue: "add_wait_queue_exclusive()" will be added 629 * after any current non-exclusive waiters, and we know that 630 * it will always _stay_ after any new non-exclusive waiters 631 * because all non-exclusive waiters are added at the 632 * beginning of the wait-queue. As such, it's ok to "drop" 633 * our exclusiveness temporarily when we get woken up without 634 * having to remove and re-insert us on the wait queue. 635 */ 636 for (;;) { 637 prepare_to_wait_exclusive(sk_sleep(sk), &wait, 638 TASK_INTERRUPTIBLE); 639 release_sock(sk); 640 if (reqsk_queue_empty(&icsk->icsk_accept_queue)) 641 timeo = schedule_timeout(timeo); 642 sched_annotate_sleep(); 643 lock_sock(sk); 644 err = 0; 645 if (!reqsk_queue_empty(&icsk->icsk_accept_queue)) 646 break; 647 err = -EINVAL; 648 if (sk->sk_state != TCP_LISTEN) 649 break; 650 err = sock_intr_errno(timeo); 651 if (signal_pending(current)) 652 break; 653 err = -EAGAIN; 654 if (!timeo) 655 break; 656 } 657 finish_wait(sk_sleep(sk), &wait); 658 return err; 659 } 660 661 /* 662 * This will accept the next outstanding connection. 663 */ 664 struct sock *inet_csk_accept(struct sock *sk, struct proto_accept_arg *arg) 665 { 666 struct inet_connection_sock *icsk = inet_csk(sk); 667 struct request_sock_queue *queue = &icsk->icsk_accept_queue; 668 struct request_sock *req; 669 struct sock *newsk; 670 int error; 671 672 lock_sock(sk); 673 674 /* We need to make sure that this socket is listening, 675 * and that it has something pending. 676 */ 677 error = -EINVAL; 678 if (sk->sk_state != TCP_LISTEN) 679 goto out_err; 680 681 /* Find already established connection */ 682 if (reqsk_queue_empty(queue)) { 683 long timeo = sock_rcvtimeo(sk, arg->flags & O_NONBLOCK); 684 685 /* If this is a non blocking socket don't sleep */ 686 error = -EAGAIN; 687 if (!timeo) 688 goto out_err; 689 690 error = inet_csk_wait_for_connect(sk, timeo); 691 if (error) 692 goto out_err; 693 } 694 req = reqsk_queue_remove(queue, sk); 695 arg->is_empty = reqsk_queue_empty(queue); 696 newsk = req->sk; 697 698 if (sk->sk_protocol == IPPROTO_TCP && 699 tcp_rsk(req)->tfo_listener) { 700 spin_lock_bh(&queue->fastopenq.lock); 701 if (tcp_rsk(req)->tfo_listener) { 702 /* We are still waiting for the final ACK from 3WHS 703 * so can't free req now. Instead, we set req->sk to 704 * NULL to signify that the child socket is taken 705 * so reqsk_fastopen_remove() will free the req 706 * when 3WHS finishes (or is aborted). 707 */ 708 req->sk = NULL; 709 req = NULL; 710 } 711 spin_unlock_bh(&queue->fastopenq.lock); 712 } 713 714 release_sock(sk); 715 716 if (req) 717 reqsk_put(req); 718 719 inet_init_csk_locks(newsk); 720 return newsk; 721 722 out_err: 723 release_sock(sk); 724 arg->err = error; 725 return NULL; 726 } 727 EXPORT_SYMBOL(inet_csk_accept); 728 729 /* 730 * Using different timers for retransmit, delayed acks and probes 731 * We may wish use just one timer maintaining a list of expire jiffies 732 * to optimize. 733 */ 734 void inet_csk_init_xmit_timers(struct sock *sk, 735 void (*retransmit_handler)(struct timer_list *t), 736 void (*delack_handler)(struct timer_list *t), 737 void (*keepalive_handler)(struct timer_list *t)) 738 { 739 struct inet_connection_sock *icsk = inet_csk(sk); 740 741 timer_setup(&sk->tcp_retransmit_timer, retransmit_handler, 0); 742 timer_setup(&icsk->icsk_delack_timer, delack_handler, 0); 743 timer_setup(&icsk->icsk_keepalive_timer, keepalive_handler, 0); 744 icsk->icsk_pending = icsk->icsk_ack.pending = 0; 745 } 746 747 void inet_csk_clear_xmit_timers(struct sock *sk) 748 { 749 struct inet_connection_sock *icsk = inet_csk(sk); 750 751 smp_store_release(&icsk->icsk_pending, 0); 752 smp_store_release(&icsk->icsk_ack.pending, 0); 753 754 sk_stop_timer(sk, &sk->tcp_retransmit_timer); 755 sk_stop_timer(sk, &icsk->icsk_delack_timer); 756 sk_stop_timer(sk, &icsk->icsk_keepalive_timer); 757 } 758 759 void inet_csk_clear_xmit_timers_sync(struct sock *sk) 760 { 761 struct inet_connection_sock *icsk = inet_csk(sk); 762 763 /* ongoing timer handlers need to acquire socket lock. */ 764 sock_not_owned_by_me(sk); 765 766 smp_store_release(&icsk->icsk_pending, 0); 767 smp_store_release(&icsk->icsk_ack.pending, 0); 768 769 sk_stop_timer_sync(sk, &sk->tcp_retransmit_timer); 770 sk_stop_timer_sync(sk, &icsk->icsk_delack_timer); 771 sk_stop_timer_sync(sk, &icsk->icsk_keepalive_timer); 772 } 773 774 struct dst_entry *inet_csk_route_req(const struct sock *sk, 775 struct flowi4 *fl4, 776 const struct request_sock *req) 777 { 778 const struct inet_request_sock *ireq = inet_rsk(req); 779 struct net *net = read_pnet(&ireq->ireq_net); 780 struct ip_options_rcu *opt; 781 struct rtable *rt; 782 783 rcu_read_lock(); 784 opt = rcu_dereference(ireq->ireq_opt); 785 786 flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark, 787 ip_sock_rt_tos(sk), ip_sock_rt_scope(sk), 788 sk->sk_protocol, inet_sk_flowi_flags(sk), 789 (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr, 790 ireq->ir_loc_addr, ireq->ir_rmt_port, 791 htons(ireq->ir_num), sk_uid(sk)); 792 security_req_classify_flow(req, flowi4_to_flowi_common(fl4)); 793 rt = ip_route_output_flow(net, fl4, sk); 794 if (IS_ERR(rt)) 795 goto no_route; 796 if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway) 797 goto route_err; 798 rcu_read_unlock(); 799 return &rt->dst; 800 801 route_err: 802 ip_rt_put(rt); 803 no_route: 804 rcu_read_unlock(); 805 __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES); 806 return NULL; 807 } 808 809 struct dst_entry *inet_csk_route_child_sock(const struct sock *sk, 810 struct sock *newsk, 811 const struct request_sock *req) 812 { 813 const struct inet_request_sock *ireq = inet_rsk(req); 814 struct net *net = read_pnet(&ireq->ireq_net); 815 struct inet_sock *newinet = inet_sk(newsk); 816 struct ip_options_rcu *opt; 817 struct flowi4 *fl4; 818 struct rtable *rt; 819 820 opt = rcu_dereference(ireq->ireq_opt); 821 fl4 = &newinet->cork.fl.u.ip4; 822 823 flowi4_init_output(fl4, ireq->ir_iif, ireq->ir_mark, 824 ip_sock_rt_tos(sk), ip_sock_rt_scope(sk), 825 sk->sk_protocol, inet_sk_flowi_flags(sk), 826 (opt && opt->opt.srr) ? opt->opt.faddr : ireq->ir_rmt_addr, 827 ireq->ir_loc_addr, ireq->ir_rmt_port, 828 htons(ireq->ir_num), sk_uid(sk)); 829 security_req_classify_flow(req, flowi4_to_flowi_common(fl4)); 830 rt = ip_route_output_flow(net, fl4, sk); 831 if (IS_ERR(rt)) 832 goto no_route; 833 if (opt && opt->opt.is_strictroute && rt->rt_uses_gateway) 834 goto route_err; 835 return &rt->dst; 836 837 route_err: 838 ip_rt_put(rt); 839 no_route: 840 __IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES); 841 return NULL; 842 } 843 EXPORT_SYMBOL_GPL(inet_csk_route_child_sock); 844 845 /* Decide when to expire the request and when to resend SYN-ACK */ 846 static void syn_ack_recalc(struct request_sock *req, 847 const int max_syn_ack_retries, 848 const u8 rskq_defer_accept, 849 int *expire, int *resend) 850 { 851 if (!rskq_defer_accept) { 852 *expire = req->num_timeout >= max_syn_ack_retries; 853 *resend = 1; 854 return; 855 } 856 *expire = req->num_timeout >= max_syn_ack_retries && 857 (!inet_rsk(req)->acked || req->num_timeout >= rskq_defer_accept); 858 /* Do not resend while waiting for data after ACK, 859 * start to resend on end of deferring period to give 860 * last chance for data or ACK to create established socket. 861 */ 862 *resend = !inet_rsk(req)->acked || 863 req->num_timeout >= rskq_defer_accept - 1; 864 } 865 866 static struct request_sock * 867 reqsk_alloc_noprof(const struct request_sock_ops *ops, struct sock *sk_listener, 868 bool attach_listener) 869 { 870 struct request_sock *req; 871 872 req = kmem_cache_alloc_noprof(ops->slab, GFP_ATOMIC | __GFP_NOWARN); 873 if (!req) 874 return NULL; 875 req->rsk_listener = NULL; 876 if (attach_listener) { 877 if (unlikely(!refcount_inc_not_zero(&sk_listener->sk_refcnt))) { 878 kmem_cache_free(ops->slab, req); 879 return NULL; 880 } 881 req->rsk_listener = sk_listener; 882 } 883 req->rsk_ops = ops; 884 req_to_sk(req)->sk_prot = sk_listener->sk_prot; 885 sk_node_init(&req_to_sk(req)->sk_node); 886 sk_tx_queue_clear(req_to_sk(req)); 887 req->saved_syn = NULL; 888 req->syncookie = 0; 889 req->num_timeout = 0; 890 req->num_retrans = 0; 891 req->sk = NULL; 892 refcount_set(&req->rsk_refcnt, 0); 893 894 return req; 895 } 896 #define reqsk_alloc(...) alloc_hooks(reqsk_alloc_noprof(__VA_ARGS__)) 897 898 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 899 struct sock *sk_listener, 900 bool attach_listener) 901 { 902 struct request_sock *req = reqsk_alloc(ops, sk_listener, 903 attach_listener); 904 905 if (req) { 906 struct inet_request_sock *ireq = inet_rsk(req); 907 908 ireq->ireq_opt = NULL; 909 #if IS_ENABLED(CONFIG_IPV6) 910 ireq->pktopts = NULL; 911 #endif 912 atomic64_set(&ireq->ir_cookie, 0); 913 ireq->ireq_state = TCP_NEW_SYN_RECV; 914 write_pnet(&ireq->ireq_net, sock_net(sk_listener)); 915 ireq->ireq_family = sk_listener->sk_family; 916 } 917 918 return req; 919 } 920 EXPORT_SYMBOL(inet_reqsk_alloc); 921 922 static struct request_sock *inet_reqsk_clone(struct request_sock *req, 923 struct sock *sk) 924 { 925 struct sock *req_sk, *nreq_sk; 926 struct request_sock *nreq; 927 928 nreq = kmem_cache_alloc(req->rsk_ops->slab, GFP_ATOMIC | __GFP_NOWARN); 929 if (!nreq) { 930 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 931 932 /* paired with refcount_inc_not_zero() in reuseport_migrate_sock() */ 933 sock_put(sk); 934 return NULL; 935 } 936 937 req_sk = req_to_sk(req); 938 nreq_sk = req_to_sk(nreq); 939 940 memcpy(nreq_sk, req_sk, 941 offsetof(struct sock, sk_dontcopy_begin)); 942 unsafe_memcpy(&nreq_sk->sk_dontcopy_end, &req_sk->sk_dontcopy_end, 943 req->rsk_ops->obj_size - offsetof(struct sock, sk_dontcopy_end), 944 /* alloc is larger than struct, see above */); 945 946 sk_node_init(&nreq_sk->sk_node); 947 nreq_sk->sk_tx_queue_mapping = req_sk->sk_tx_queue_mapping; 948 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING 949 nreq_sk->sk_rx_queue_mapping = req_sk->sk_rx_queue_mapping; 950 #endif 951 nreq_sk->sk_incoming_cpu = req_sk->sk_incoming_cpu; 952 953 nreq->rsk_listener = sk; 954 955 /* We need not acquire fastopenq->lock 956 * because the child socket is locked in inet_csk_listen_stop(). 957 */ 958 if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(nreq)->tfo_listener) 959 rcu_assign_pointer(tcp_sk(nreq->sk)->fastopen_rsk, nreq); 960 961 return nreq; 962 } 963 964 static void reqsk_queue_migrated(struct request_sock_queue *queue, 965 const struct request_sock *req) 966 { 967 if (req->num_timeout == 0) 968 atomic_inc(&queue->young); 969 atomic_inc(&queue->qlen); 970 } 971 972 static void reqsk_migrate_reset(struct request_sock *req) 973 { 974 req->saved_syn = NULL; 975 #if IS_ENABLED(CONFIG_IPV6) 976 inet_rsk(req)->ipv6_opt = NULL; 977 inet_rsk(req)->pktopts = NULL; 978 #else 979 inet_rsk(req)->ireq_opt = NULL; 980 #endif 981 } 982 983 /* return true if req was found in the ehash table */ 984 static bool reqsk_queue_unlink(struct request_sock *req) 985 { 986 struct sock *sk = req_to_sk(req); 987 bool found = false; 988 989 if (sk_hashed(sk)) { 990 struct inet_hashinfo *hashinfo = tcp_get_hashinfo(sk); 991 spinlock_t *lock; 992 993 lock = inet_ehash_lockp(hashinfo, req->rsk_hash); 994 spin_lock(lock); 995 found = __sk_nulls_del_node_init_rcu(sk); 996 spin_unlock(lock); 997 } 998 999 return found; 1000 } 1001 1002 static bool __inet_csk_reqsk_queue_drop(struct sock *sk, 1003 struct request_sock *req, 1004 bool from_timer) 1005 { 1006 bool unlinked = reqsk_queue_unlink(req); 1007 1008 if (!from_timer && timer_delete_sync(&req->rsk_timer)) 1009 reqsk_put(req); 1010 1011 if (unlinked) { 1012 reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req); 1013 reqsk_put(req); 1014 } 1015 1016 return unlinked; 1017 } 1018 1019 bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req) 1020 { 1021 return __inet_csk_reqsk_queue_drop(sk, req, false); 1022 } 1023 1024 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req) 1025 { 1026 inet_csk_reqsk_queue_drop(sk, req); 1027 reqsk_put(req); 1028 } 1029 EXPORT_IPV6_MOD(inet_csk_reqsk_queue_drop_and_put); 1030 1031 static void reqsk_timer_handler(struct timer_list *t) 1032 { 1033 struct request_sock *req = timer_container_of(req, t, rsk_timer); 1034 struct request_sock *nreq = NULL, *oreq = req; 1035 struct sock *sk_listener = req->rsk_listener; 1036 struct inet_connection_sock *icsk; 1037 struct request_sock_queue *queue; 1038 struct net *net; 1039 int max_syn_ack_retries, qlen, expire = 0, resend = 0; 1040 1041 if (inet_sk_state_load(sk_listener) != TCP_LISTEN) { 1042 struct sock *nsk; 1043 1044 nsk = reuseport_migrate_sock(sk_listener, req_to_sk(req), NULL); 1045 if (!nsk) 1046 goto drop; 1047 1048 nreq = inet_reqsk_clone(req, nsk); 1049 if (!nreq) 1050 goto drop; 1051 1052 /* The new timer for the cloned req can decrease the 2 1053 * by calling inet_csk_reqsk_queue_drop_and_put(), so 1054 * hold another count to prevent use-after-free and 1055 * call reqsk_put() just before return. 1056 */ 1057 refcount_set(&nreq->rsk_refcnt, 2 + 1); 1058 timer_setup(&nreq->rsk_timer, reqsk_timer_handler, TIMER_PINNED); 1059 reqsk_queue_migrated(&inet_csk(nsk)->icsk_accept_queue, req); 1060 1061 req = nreq; 1062 sk_listener = nsk; 1063 } 1064 1065 icsk = inet_csk(sk_listener); 1066 net = sock_net(sk_listener); 1067 max_syn_ack_retries = READ_ONCE(icsk->icsk_syn_retries) ? : 1068 READ_ONCE(net->ipv4.sysctl_tcp_synack_retries); 1069 /* Normally all the openreqs are young and become mature 1070 * (i.e. converted to established socket) for first timeout. 1071 * If synack was not acknowledged for 1 second, it means 1072 * one of the following things: synack was lost, ack was lost, 1073 * rtt is high or nobody planned to ack (i.e. synflood). 1074 * When server is a bit loaded, queue is populated with old 1075 * open requests, reducing effective size of queue. 1076 * When server is well loaded, queue size reduces to zero 1077 * after several minutes of work. It is not synflood, 1078 * it is normal operation. The solution is pruning 1079 * too old entries overriding normal timeout, when 1080 * situation becomes dangerous. 1081 * 1082 * Essentially, we reserve half of room for young 1083 * embrions; and abort old ones without pity, if old 1084 * ones are about to clog our table. 1085 */ 1086 queue = &icsk->icsk_accept_queue; 1087 qlen = reqsk_queue_len(queue); 1088 if ((qlen << 1) > max(8U, READ_ONCE(sk_listener->sk_max_ack_backlog))) { 1089 int young = reqsk_queue_len_young(queue) << 1; 1090 1091 while (max_syn_ack_retries > 2) { 1092 if (qlen < young) 1093 break; 1094 max_syn_ack_retries--; 1095 young <<= 1; 1096 } 1097 } 1098 1099 syn_ack_recalc(req, max_syn_ack_retries, READ_ONCE(queue->rskq_defer_accept), 1100 &expire, &resend); 1101 tcp_syn_ack_timeout(req); 1102 1103 if (!expire && 1104 (!resend || 1105 !tcp_rtx_synack(sk_listener, req) || 1106 inet_rsk(req)->acked)) { 1107 if (req->num_retrans > 1 && tcp_rsk(req)->accecn_ok) 1108 tcp_rsk(req)->accecn_fail_mode |= TCP_ACCECN_ACE_FAIL_SEND; 1109 if (req->num_timeout++ == 0) 1110 atomic_dec(&queue->young); 1111 mod_timer(&req->rsk_timer, jiffies + tcp_reqsk_timeout(req)); 1112 1113 if (!nreq) 1114 return; 1115 1116 if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) { 1117 /* delete timer */ 1118 __inet_csk_reqsk_queue_drop(sk_listener, nreq, true); 1119 goto no_ownership; 1120 } 1121 1122 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQSUCCESS); 1123 reqsk_migrate_reset(oreq); 1124 reqsk_queue_removed(&inet_csk(oreq->rsk_listener)->icsk_accept_queue, oreq); 1125 reqsk_put(oreq); 1126 1127 reqsk_put(nreq); 1128 return; 1129 } 1130 1131 /* Even if we can clone the req, we may need not retransmit any more 1132 * SYN+ACKs (nreq->num_timeout > max_syn_ack_retries, etc), or another 1133 * CPU may win the "own_req" race so that inet_ehash_insert() fails. 1134 */ 1135 if (nreq) { 1136 __NET_INC_STATS(net, LINUX_MIB_TCPMIGRATEREQFAILURE); 1137 no_ownership: 1138 reqsk_migrate_reset(nreq); 1139 reqsk_queue_removed(queue, nreq); 1140 __reqsk_free(nreq); 1141 } 1142 1143 drop: 1144 __inet_csk_reqsk_queue_drop(sk_listener, oreq, true); 1145 reqsk_put(oreq); 1146 } 1147 1148 static bool reqsk_queue_hash_req(struct request_sock *req) 1149 { 1150 bool found_dup_sk = false; 1151 1152 if (!inet_ehash_insert(req_to_sk(req), NULL, &found_dup_sk)) 1153 return false; 1154 1155 /* The timer needs to be setup after a successful insertion. */ 1156 req->timeout = tcp_timeout_init((struct sock *)req); 1157 timer_setup(&req->rsk_timer, reqsk_timer_handler, TIMER_PINNED); 1158 mod_timer(&req->rsk_timer, jiffies + req->timeout); 1159 1160 /* before letting lookups find us, make sure all req fields 1161 * are committed to memory and refcnt initialized. 1162 */ 1163 smp_wmb(); 1164 refcount_set(&req->rsk_refcnt, 2 + 1); 1165 return true; 1166 } 1167 1168 bool inet_csk_reqsk_queue_hash_add(struct sock *sk, struct request_sock *req) 1169 { 1170 if (!reqsk_queue_hash_req(req)) 1171 return false; 1172 1173 inet_csk_reqsk_queue_added(sk); 1174 return true; 1175 } 1176 1177 static void inet_clone_ulp(const struct request_sock *req, struct sock *newsk, 1178 const gfp_t priority) 1179 { 1180 struct inet_connection_sock *icsk = inet_csk(newsk); 1181 1182 if (!icsk->icsk_ulp_ops) 1183 return; 1184 1185 icsk->icsk_ulp_ops->clone(req, newsk, priority); 1186 } 1187 1188 /** 1189 * inet_csk_clone_lock - clone an inet socket, and lock its clone 1190 * @sk: the socket to clone 1191 * @req: request_sock 1192 * @priority: for allocation (%GFP_KERNEL, %GFP_ATOMIC, etc) 1193 * 1194 * Caller must unlock socket even in error path (bh_unlock_sock(newsk)) 1195 */ 1196 struct sock *inet_csk_clone_lock(const struct sock *sk, 1197 const struct request_sock *req, 1198 const gfp_t priority) 1199 { 1200 struct sock *newsk = sk_clone_lock(sk, priority); 1201 struct inet_connection_sock *newicsk; 1202 const struct inet_request_sock *ireq; 1203 struct inet_sock *newinet; 1204 1205 if (!newsk) 1206 return NULL; 1207 1208 newicsk = inet_csk(newsk); 1209 newinet = inet_sk(newsk); 1210 ireq = inet_rsk(req); 1211 1212 newicsk->icsk_bind_hash = NULL; 1213 newicsk->icsk_bind2_hash = NULL; 1214 1215 newinet->inet_dport = ireq->ir_rmt_port; 1216 newinet->inet_num = ireq->ir_num; 1217 newinet->inet_sport = htons(ireq->ir_num); 1218 1219 newsk->sk_bound_dev_if = ireq->ir_iif; 1220 1221 newsk->sk_daddr = ireq->ir_rmt_addr; 1222 newsk->sk_rcv_saddr = ireq->ir_loc_addr; 1223 newinet->inet_saddr = ireq->ir_loc_addr; 1224 1225 #if IS_ENABLED(CONFIG_IPV6) 1226 newsk->sk_v6_daddr = ireq->ir_v6_rmt_addr; 1227 newsk->sk_v6_rcv_saddr = ireq->ir_v6_loc_addr; 1228 #endif 1229 1230 /* listeners have SOCK_RCU_FREE, not the children */ 1231 sock_reset_flag(newsk, SOCK_RCU_FREE); 1232 1233 inet_sk(newsk)->mc_list = NULL; 1234 1235 newsk->sk_mark = inet_rsk(req)->ir_mark; 1236 atomic64_set(&newsk->sk_cookie, 1237 atomic64_read(&inet_rsk(req)->ir_cookie)); 1238 1239 newicsk->icsk_retransmits = 0; 1240 newicsk->icsk_backoff = 0; 1241 newicsk->icsk_probes_out = 0; 1242 newicsk->icsk_probes_tstamp = 0; 1243 1244 /* Deinitialize accept_queue to trap illegal accesses. */ 1245 memset(&newicsk->icsk_accept_queue, 0, 1246 sizeof(newicsk->icsk_accept_queue)); 1247 1248 inet_sk_set_state(newsk, TCP_SYN_RECV); 1249 1250 inet_clone_ulp(req, newsk, priority); 1251 1252 security_inet_csk_clone(newsk, req); 1253 1254 return newsk; 1255 } 1256 1257 /* 1258 * At this point, there should be no process reference to this 1259 * socket, and thus no user references at all. Therefore we 1260 * can assume the socket waitqueue is inactive and nobody will 1261 * try to jump onto it. 1262 */ 1263 void inet_csk_destroy_sock(struct sock *sk) 1264 { 1265 WARN_ON(sk->sk_state != TCP_CLOSE); 1266 WARN_ON(!sock_flag(sk, SOCK_DEAD)); 1267 1268 /* It cannot be in hash table! */ 1269 WARN_ON(!sk_unhashed(sk)); 1270 1271 /* If it has not 0 inet_sk(sk)->inet_num, it must be bound */ 1272 WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash); 1273 1274 sk->sk_prot->destroy(sk); 1275 1276 sk_stream_kill_queues(sk); 1277 1278 xfrm_sk_free_policy(sk); 1279 1280 tcp_orphan_count_dec(); 1281 1282 sock_put(sk); 1283 } 1284 EXPORT_SYMBOL(inet_csk_destroy_sock); 1285 1286 void inet_csk_prepare_for_destroy_sock(struct sock *sk) 1287 { 1288 /* The below has to be done to allow calling inet_csk_destroy_sock */ 1289 sock_set_flag(sk, SOCK_DEAD); 1290 tcp_orphan_count_inc(); 1291 } 1292 1293 /* This function allows to force a closure of a socket after the call to 1294 * tcp_create_openreq_child(). 1295 */ 1296 void inet_csk_prepare_forced_close(struct sock *sk) 1297 __releases(&sk->sk_lock.slock) 1298 { 1299 /* sk_clone_lock locked the socket and set refcnt to 2 */ 1300 bh_unlock_sock(sk); 1301 sock_put(sk); 1302 inet_csk_prepare_for_destroy_sock(sk); 1303 inet_sk(sk)->inet_num = 0; 1304 } 1305 EXPORT_SYMBOL(inet_csk_prepare_forced_close); 1306 1307 static int inet_ulp_can_listen(const struct sock *sk) 1308 { 1309 const struct inet_connection_sock *icsk = inet_csk(sk); 1310 1311 if (icsk->icsk_ulp_ops && !icsk->icsk_ulp_ops->clone) 1312 return -EINVAL; 1313 1314 return 0; 1315 } 1316 1317 int inet_csk_listen_start(struct sock *sk) 1318 { 1319 struct inet_connection_sock *icsk = inet_csk(sk); 1320 struct inet_sock *inet = inet_sk(sk); 1321 int err; 1322 1323 err = inet_ulp_can_listen(sk); 1324 if (unlikely(err)) 1325 return err; 1326 1327 reqsk_queue_alloc(&icsk->icsk_accept_queue); 1328 1329 sk->sk_ack_backlog = 0; 1330 inet_csk_delack_init(sk); 1331 1332 /* There is race window here: we announce ourselves listening, 1333 * but this transition is still not validated by get_port(). 1334 * It is OK, because this socket enters to hash table only 1335 * after validation is complete. 1336 */ 1337 inet_sk_state_store(sk, TCP_LISTEN); 1338 err = sk->sk_prot->get_port(sk, inet->inet_num); 1339 if (!err) { 1340 inet->inet_sport = htons(inet->inet_num); 1341 1342 sk_dst_reset(sk); 1343 err = sk->sk_prot->hash(sk); 1344 1345 if (likely(!err)) 1346 return 0; 1347 } 1348 1349 inet_sk_set_state(sk, TCP_CLOSE); 1350 return err; 1351 } 1352 1353 static void inet_child_forget(struct sock *sk, struct request_sock *req, 1354 struct sock *child) 1355 { 1356 sk->sk_prot->disconnect(child, O_NONBLOCK); 1357 1358 sock_orphan(child); 1359 1360 tcp_orphan_count_inc(); 1361 1362 if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) { 1363 BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req); 1364 BUG_ON(sk != req->rsk_listener); 1365 1366 /* Paranoid, to prevent race condition if 1367 * an inbound pkt destined for child is 1368 * blocked by sock lock in tcp_v4_rcv(). 1369 * Also to satisfy an assertion in 1370 * tcp_v4_destroy_sock(). 1371 */ 1372 RCU_INIT_POINTER(tcp_sk(child)->fastopen_rsk, NULL); 1373 } 1374 inet_csk_destroy_sock(child); 1375 } 1376 1377 struct sock *inet_csk_reqsk_queue_add(struct sock *sk, 1378 struct request_sock *req, 1379 struct sock *child) 1380 { 1381 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 1382 1383 spin_lock(&queue->rskq_lock); 1384 if (unlikely(sk->sk_state != TCP_LISTEN)) { 1385 inet_child_forget(sk, req, child); 1386 child = NULL; 1387 } else { 1388 req->sk = child; 1389 req->dl_next = NULL; 1390 if (queue->rskq_accept_head == NULL) 1391 WRITE_ONCE(queue->rskq_accept_head, req); 1392 else 1393 queue->rskq_accept_tail->dl_next = req; 1394 queue->rskq_accept_tail = req; 1395 sk_acceptq_added(sk); 1396 } 1397 spin_unlock(&queue->rskq_lock); 1398 return child; 1399 } 1400 EXPORT_SYMBOL(inet_csk_reqsk_queue_add); 1401 1402 struct sock *inet_csk_complete_hashdance(struct sock *sk, struct sock *child, 1403 struct request_sock *req, bool own_req) 1404 { 1405 if (own_req) { 1406 inet_csk_reqsk_queue_drop(req->rsk_listener, req); 1407 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req); 1408 1409 if (sk != req->rsk_listener) { 1410 /* another listening sk has been selected, 1411 * migrate the req to it. 1412 */ 1413 struct request_sock *nreq; 1414 1415 /* hold a refcnt for the nreq->rsk_listener 1416 * which is assigned in inet_reqsk_clone() 1417 */ 1418 sock_hold(sk); 1419 nreq = inet_reqsk_clone(req, sk); 1420 if (!nreq) { 1421 inet_child_forget(sk, req, child); 1422 goto child_put; 1423 } 1424 1425 refcount_set(&nreq->rsk_refcnt, 1); 1426 if (inet_csk_reqsk_queue_add(sk, nreq, child)) { 1427 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQSUCCESS); 1428 reqsk_migrate_reset(req); 1429 reqsk_put(req); 1430 return child; 1431 } 1432 1433 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 1434 reqsk_migrate_reset(nreq); 1435 __reqsk_free(nreq); 1436 } else if (inet_csk_reqsk_queue_add(sk, req, child)) { 1437 return child; 1438 } 1439 } 1440 /* Too bad, another child took ownership of the request, undo. */ 1441 child_put: 1442 bh_unlock_sock(child); 1443 sock_put(child); 1444 return NULL; 1445 } 1446 1447 /* 1448 * This routine closes sockets which have been at least partially 1449 * opened, but not yet accepted. 1450 */ 1451 void inet_csk_listen_stop(struct sock *sk) 1452 { 1453 struct inet_connection_sock *icsk = inet_csk(sk); 1454 struct request_sock_queue *queue = &icsk->icsk_accept_queue; 1455 struct request_sock *next, *req; 1456 1457 /* Following specs, it would be better either to send FIN 1458 * (and enter FIN-WAIT-1, it is normal close) 1459 * or to send active reset (abort). 1460 * Certainly, it is pretty dangerous while synflood, but it is 1461 * bad justification for our negligence 8) 1462 * To be honest, we are not able to make either 1463 * of the variants now. --ANK 1464 */ 1465 while ((req = reqsk_queue_remove(queue, sk)) != NULL) { 1466 struct sock *child = req->sk, *nsk; 1467 struct request_sock *nreq; 1468 1469 local_bh_disable(); 1470 bh_lock_sock(child); 1471 WARN_ON(sock_owned_by_user(child)); 1472 sock_hold(child); 1473 1474 nsk = reuseport_migrate_sock(sk, child, NULL); 1475 if (nsk) { 1476 nreq = inet_reqsk_clone(req, nsk); 1477 if (nreq) { 1478 refcount_set(&nreq->rsk_refcnt, 1); 1479 1480 if (inet_csk_reqsk_queue_add(nsk, nreq, child)) { 1481 __NET_INC_STATS(sock_net(nsk), 1482 LINUX_MIB_TCPMIGRATEREQSUCCESS); 1483 reqsk_migrate_reset(req); 1484 } else { 1485 __NET_INC_STATS(sock_net(nsk), 1486 LINUX_MIB_TCPMIGRATEREQFAILURE); 1487 reqsk_migrate_reset(nreq); 1488 __reqsk_free(nreq); 1489 } 1490 1491 /* inet_csk_reqsk_queue_add() has already 1492 * called inet_child_forget() on failure case. 1493 */ 1494 goto skip_child_forget; 1495 } 1496 } 1497 1498 inet_child_forget(sk, req, child); 1499 skip_child_forget: 1500 reqsk_put(req); 1501 bh_unlock_sock(child); 1502 local_bh_enable(); 1503 sock_put(child); 1504 1505 cond_resched(); 1506 } 1507 if (queue->fastopenq.rskq_rst_head) { 1508 /* Free all the reqs queued in rskq_rst_head. */ 1509 spin_lock_bh(&queue->fastopenq.lock); 1510 req = queue->fastopenq.rskq_rst_head; 1511 queue->fastopenq.rskq_rst_head = NULL; 1512 spin_unlock_bh(&queue->fastopenq.lock); 1513 while (req != NULL) { 1514 next = req->dl_next; 1515 reqsk_put(req); 1516 req = next; 1517 } 1518 } 1519 WARN_ON_ONCE(sk->sk_ack_backlog); 1520 } 1521 EXPORT_SYMBOL_GPL(inet_csk_listen_stop); 1522 1523 static struct dst_entry *inet_csk_rebuild_route(struct sock *sk, struct flowi *fl) 1524 { 1525 const struct inet_sock *inet = inet_sk(sk); 1526 struct flowi4 *fl4; 1527 struct rtable *rt; 1528 1529 rcu_read_lock(); 1530 fl4 = &fl->u.ip4; 1531 inet_sk_init_flowi4(inet, fl4); 1532 rt = ip_route_output_flow(sock_net(sk), fl4, sk); 1533 if (IS_ERR(rt)) 1534 rt = NULL; 1535 if (rt) 1536 sk_setup_caps(sk, &rt->dst); 1537 rcu_read_unlock(); 1538 1539 return &rt->dst; 1540 } 1541 1542 struct dst_entry *inet_csk_update_pmtu(struct sock *sk, u32 mtu) 1543 { 1544 struct dst_entry *dst = __sk_dst_check(sk, 0); 1545 struct inet_sock *inet = inet_sk(sk); 1546 1547 if (!dst) { 1548 dst = inet_csk_rebuild_route(sk, &inet->cork.fl); 1549 if (!dst) 1550 goto out; 1551 } 1552 dst->ops->update_pmtu(dst, sk, NULL, mtu, true); 1553 1554 dst = __sk_dst_check(sk, 0); 1555 if (!dst) 1556 dst = inet_csk_rebuild_route(sk, &inet->cork.fl); 1557 out: 1558 return dst; 1559 } 1560