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