1 // SPDX-License-Identifier: GPL-2.0 2 /* Multipath TCP 3 * 4 * Copyright (c) 2019, Intel Corporation. 5 */ 6 #define pr_fmt(fmt) "MPTCP: " fmt 7 8 #include <linux/rculist.h> 9 #include <linux/spinlock.h> 10 #include "protocol.h" 11 #include "mib.h" 12 13 #define ADD_ADDR_RETRANS_MAX 3 14 15 struct mptcp_pm_add_addr { 16 struct list_head list; 17 struct mptcp_addr_info addr; 18 u8 retrans_times; 19 bool timer_done; 20 struct timer_list timer; 21 struct mptcp_sock *sock; 22 struct rcu_head rcu; 23 }; 24 25 static DEFINE_SPINLOCK(mptcp_pm_list_lock); 26 static LIST_HEAD(mptcp_pm_list); 27 28 /* path manager helpers */ 29 30 /* if sk is ipv4 or ipv6_only allows only same-family local and remote addresses, 31 * otherwise allow any matching local/remote pair 32 */ 33 bool mptcp_pm_addr_families_match(const struct sock *sk, 34 const struct mptcp_addr_info *loc, 35 const struct mptcp_addr_info *rem) 36 { 37 bool mptcp_is_v4 = sk->sk_family == AF_INET; 38 39 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 40 bool loc_is_v4 = loc->family == AF_INET || ipv6_addr_v4mapped(&loc->addr6); 41 bool rem_is_v4 = rem->family == AF_INET || ipv6_addr_v4mapped(&rem->addr6); 42 43 if (mptcp_is_v4) 44 return loc_is_v4 && rem_is_v4; 45 46 if (ipv6_only_sock(sk)) 47 return !loc_is_v4 && !rem_is_v4; 48 49 return loc_is_v4 == rem_is_v4; 50 #else 51 return mptcp_is_v4 && loc->family == AF_INET && rem->family == AF_INET; 52 #endif 53 } 54 55 bool mptcp_addresses_equal(const struct mptcp_addr_info *a, 56 const struct mptcp_addr_info *b, bool use_port) 57 { 58 bool addr_equals = false; 59 60 if (a->family == b->family) { 61 if (a->family == AF_INET) 62 addr_equals = a->addr.s_addr == b->addr.s_addr; 63 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 64 else 65 addr_equals = ipv6_addr_equal(&a->addr6, &b->addr6); 66 } else if (a->family == AF_INET) { 67 if (ipv6_addr_v4mapped(&b->addr6)) 68 addr_equals = a->addr.s_addr == b->addr6.s6_addr32[3]; 69 } else if (b->family == AF_INET) { 70 if (ipv6_addr_v4mapped(&a->addr6)) 71 addr_equals = a->addr6.s6_addr32[3] == b->addr.s_addr; 72 #endif 73 } 74 75 if (!addr_equals) 76 return false; 77 if (!use_port) 78 return true; 79 80 return a->port == b->port; 81 } 82 83 void mptcp_local_address(const struct sock_common *skc, 84 struct mptcp_addr_info *addr) 85 { 86 addr->family = skc->skc_family; 87 addr->port = htons(skc->skc_num); 88 if (addr->family == AF_INET) 89 addr->addr.s_addr = skc->skc_rcv_saddr; 90 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 91 else if (addr->family == AF_INET6) 92 addr->addr6 = skc->skc_v6_rcv_saddr; 93 #endif 94 } 95 96 void mptcp_remote_address(const struct sock_common *skc, 97 struct mptcp_addr_info *addr) 98 { 99 addr->family = skc->skc_family; 100 addr->port = skc->skc_dport; 101 if (addr->family == AF_INET) 102 addr->addr.s_addr = skc->skc_daddr; 103 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 104 else if (addr->family == AF_INET6) 105 addr->addr6 = skc->skc_v6_daddr; 106 #endif 107 } 108 109 static bool mptcp_pm_is_init_remote_addr(struct mptcp_sock *msk, 110 const struct mptcp_addr_info *remote) 111 { 112 struct mptcp_addr_info mpc_remote; 113 114 mptcp_remote_address((struct sock_common *)msk, &mpc_remote); 115 return mptcp_addresses_equal(&mpc_remote, remote, remote->port); 116 } 117 118 bool mptcp_pm_has_subflow_saddr(const struct mptcp_sock *msk, 119 const struct mptcp_addr_info *saddr) 120 { 121 struct mptcp_subflow_context *subflow; 122 struct mptcp_addr_info cur; 123 struct sock_common *skc; 124 125 mptcp_for_each_subflow(msk, subflow) { 126 skc = (struct sock_common *)mptcp_subflow_tcp_sock(subflow); 127 128 mptcp_local_address(skc, &cur); 129 if (mptcp_addresses_equal(&cur, saddr, saddr->port)) 130 return true; 131 } 132 133 return false; 134 } 135 136 static struct mptcp_pm_add_addr * 137 mptcp_pm_announced_lookup(const struct mptcp_sock *msk, 138 const struct mptcp_addr_info *addr) 139 { 140 struct mptcp_pm_add_addr *entry; 141 142 lockdep_assert_held(&msk->pm.lock); 143 144 list_for_each_entry(entry, &msk->pm.anno_list, list) { 145 if (mptcp_addresses_equal(&entry->addr, addr, true)) 146 return entry; 147 } 148 149 return NULL; 150 } 151 152 static struct mptcp_pm_add_addr * 153 mptcp_pm_announced_del_timer(struct mptcp_sock *msk, 154 const struct mptcp_addr_info *addr, bool check_id) 155 { 156 struct sock *sk = (struct sock *)msk; 157 struct mptcp_pm_add_addr *entry; 158 bool stop_timer = false; 159 160 rcu_read_lock(); 161 162 spin_lock_bh(&msk->pm.lock); 163 entry = mptcp_pm_announced_lookup(msk, addr); 164 if (entry && (!check_id || entry->addr.id == addr->id)) { 165 entry->retrans_times = ADD_ADDR_RETRANS_MAX; 166 stop_timer = true; 167 } 168 if (!check_id && entry) 169 list_del(&entry->list); 170 spin_unlock_bh(&msk->pm.lock); 171 172 /* Note: entry might have been removed by another thread. 173 * We hold rcu_read_lock() to ensure it is not freed under us. 174 */ 175 if (stop_timer) { 176 if (check_id) 177 sk_stop_timer(sk, &entry->timer); 178 else 179 sk_stop_timer_sync(sk, &entry->timer); 180 } 181 182 rcu_read_unlock(); 183 return entry; 184 } 185 186 bool mptcp_pm_announced_remove(struct mptcp_sock *msk, 187 const struct mptcp_addr_info *addr) 188 { 189 struct mptcp_pm_add_addr *entry; 190 bool ret; 191 192 entry = mptcp_pm_announced_del_timer(msk, addr, false); 193 ret = entry; 194 kfree_rcu(entry, rcu); 195 196 return ret; 197 } 198 199 bool mptcp_pm_announced_has_ssk(struct mptcp_sock *msk, const struct sock *ssk) 200 { 201 struct mptcp_pm_add_addr *entry; 202 struct mptcp_addr_info saddr; 203 bool ret = false; 204 205 mptcp_local_address((struct sock_common *)ssk, &saddr); 206 207 spin_lock_bh(&msk->pm.lock); 208 list_for_each_entry(entry, &msk->pm.anno_list, list) { 209 if (mptcp_addresses_equal(&entry->addr, &saddr, true)) { 210 ret = true; 211 goto out; 212 } 213 } 214 215 out: 216 spin_unlock_bh(&msk->pm.lock); 217 return ret; 218 } 219 220 static void __mptcp_pm_send_ack(struct mptcp_sock *msk, 221 struct mptcp_subflow_context *subflow, 222 bool prio, bool backup) 223 { 224 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 225 bool slow; 226 227 pr_debug("send ack for %s\n", 228 prio ? "mp_prio" : 229 (mptcp_pm_should_add_signal(msk) ? "add_addr" : "rm_addr")); 230 231 slow = lock_sock_fast(ssk); 232 if (prio) { 233 subflow->send_mp_prio = 1; 234 subflow->request_bkup = backup; 235 } 236 237 __mptcp_subflow_send_ack(ssk); 238 unlock_sock_fast(ssk, slow); 239 } 240 241 void mptcp_pm_send_ack(struct mptcp_sock *msk, 242 struct mptcp_subflow_context *subflow, 243 bool prio, bool backup) 244 { 245 spin_unlock_bh(&msk->pm.lock); 246 __mptcp_pm_send_ack(msk, subflow, prio, backup); 247 spin_lock_bh(&msk->pm.lock); 248 } 249 250 static bool subflow_in_rm_list(const struct mptcp_subflow_context *subflow, 251 const struct mptcp_rm_list *rm_list) 252 { 253 u8 i, id = subflow_get_local_id(subflow); 254 255 for (i = 0; i < rm_list->nr; i++) { 256 if (rm_list->ids[i] == id) 257 return true; 258 } 259 260 return false; 261 } 262 263 static void mptcp_pm_add_addr_send_ack(struct mptcp_sock *msk); 264 static void 265 mptcp_pm_addr_send_ack_avoid_list(struct mptcp_sock *msk, 266 const struct mptcp_rm_list *rm_list) 267 { 268 struct mptcp_subflow_context *subflow, *stale = NULL, *same_id = NULL; 269 270 msk_owned_by_me(msk); 271 lockdep_assert_held(&msk->pm.lock); 272 273 if (!mptcp_pm_should_add_signal(msk) && 274 !mptcp_pm_should_rm_signal(msk)) 275 return; 276 277 mptcp_for_each_subflow(msk, subflow) { 278 if (!__mptcp_subflow_active(subflow)) 279 continue; 280 281 if (unlikely(subflow->stale)) { 282 if (!stale) 283 stale = subflow; 284 } else if (unlikely(rm_list && 285 subflow_in_rm_list(subflow, rm_list))) { 286 if (!same_id) 287 same_id = subflow; 288 } else { 289 goto send_ack; 290 } 291 } 292 293 if (same_id) 294 subflow = same_id; 295 else if (stale) 296 subflow = stale; 297 else 298 return; 299 300 send_ack: 301 mptcp_pm_send_ack(msk, subflow, false, false); 302 } 303 304 void mptcp_pm_addr_send_ack(struct mptcp_sock *msk) 305 { 306 mptcp_pm_addr_send_ack_avoid_list(msk, NULL); 307 } 308 309 int mptcp_pm_mp_prio_send_ack(struct mptcp_sock *msk, 310 struct mptcp_addr_info *addr, 311 struct mptcp_addr_info *rem, 312 u8 bkup) 313 { 314 struct mptcp_subflow_context *subflow; 315 316 pr_debug("bkup=%d\n", bkup); 317 318 mptcp_for_each_subflow(msk, subflow) { 319 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 320 struct mptcp_addr_info local, remote; 321 322 if (!__mptcp_subflow_active(subflow)) 323 continue; 324 325 mptcp_local_address((struct sock_common *)ssk, &local); 326 if (!mptcp_addresses_equal(&local, addr, addr->port)) 327 continue; 328 329 if (rem && rem->family != AF_UNSPEC) { 330 mptcp_remote_address((struct sock_common *)ssk, &remote); 331 if (!mptcp_addresses_equal(&remote, rem, rem->port)) 332 continue; 333 } 334 335 __mptcp_pm_send_ack(msk, subflow, true, bkup); 336 return 0; 337 } 338 339 return -EINVAL; 340 } 341 342 static unsigned int mptcp_adjust_add_addr_timeout(struct mptcp_sock *msk) 343 { 344 const struct net *net = sock_net((struct sock *)msk); 345 unsigned int rto = mptcp_get_add_addr_timeout(net); 346 struct mptcp_subflow_context *subflow; 347 unsigned int max = 0, max_stale = 0; 348 349 if (!rto) 350 return 0; 351 352 mptcp_for_each_subflow(msk, subflow) { 353 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 354 struct inet_connection_sock *icsk = inet_csk(ssk); 355 356 if (!__mptcp_subflow_active(subflow)) 357 continue; 358 359 if (unlikely(subflow->stale)) { 360 if (icsk->icsk_rto > max_stale) 361 max_stale = icsk->icsk_rto; 362 } else if (icsk->icsk_rto > max) { 363 max = icsk->icsk_rto; 364 } 365 } 366 367 if (max) 368 return min(max, rto); 369 370 if (max_stale) 371 return min(max_stale, rto); 372 373 return rto; 374 } 375 376 static void mptcp_pm_add_addr_timer(struct timer_list *timer) 377 { 378 struct mptcp_pm_add_addr *entry = timer_container_of(entry, timer, 379 timer); 380 struct mptcp_sock *msk = entry->sock; 381 struct sock *sk = (struct sock *)msk; 382 unsigned int timeout = 0; 383 bool retransmit; 384 385 pr_debug("msk=%p\n", msk); 386 387 bh_lock_sock(sk); 388 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 389 goto out; 390 391 if (sock_owned_by_user(sk)) { 392 /* Try again later. */ 393 timeout = HZ / 20; 394 goto out; 395 } 396 397 timeout = mptcp_adjust_add_addr_timeout(msk); 398 if (!timeout || mptcp_pm_should_add_signal_addr(msk)) 399 goto out; 400 401 spin_lock_bh(&msk->pm.lock); 402 403 /* The cancel path (mptcp_pm_announced_del_timer()) can race with this 404 * callback. Once cancel updates retrans_times to MAX, suppress further 405 * retransmissions here. If this callback acquires pm.lock first, one 406 * final transmit attempt is still possible. 407 */ 408 if (entry->retrans_times < ADD_ADDR_RETRANS_MAX && 409 !mptcp_pm_should_add_signal_addr(msk)) { 410 pr_debug("retransmit ADD_ADDR id=%d\n", entry->addr.id); 411 mptcp_pm_announce_addr(msk, &entry->addr, false); 412 mptcp_pm_add_addr_send_ack(msk); 413 entry->retrans_times++; 414 } 415 416 retransmit = entry->retrans_times < ADD_ADDR_RETRANS_MAX; 417 if (retransmit) 418 timeout <<= entry->retrans_times; 419 else 420 timeout = 0; 421 422 spin_unlock_bh(&msk->pm.lock); 423 424 if (!retransmit) 425 mptcp_pm_subflow_established(msk); 426 427 out: 428 if (timeout) 429 sk_reset_timer(sk, timer, jiffies + timeout); 430 else 431 /* if sock_put calls sk_free: avoid waiting for this timer */ 432 entry->timer_done = true; 433 bh_unlock_sock(sk); 434 sock_put(sk); 435 } 436 437 bool mptcp_pm_announced_alloc(struct mptcp_sock *msk, 438 const struct mptcp_addr_info *addr) 439 { 440 struct mptcp_pm_add_addr *add_entry = NULL; 441 struct sock *sk = (struct sock *)msk; 442 unsigned int timeout; 443 444 lockdep_assert_held(&msk->pm.lock); 445 446 if (msk->pm.status & BIT(MPTCP_PM_DESTROYING)) 447 return false; 448 449 add_entry = mptcp_pm_announced_lookup(msk, addr); 450 if (add_entry) { 451 if (WARN_ON_ONCE(mptcp_pm_is_kernel(msk))) 452 return false; 453 454 goto reset_timer; 455 } 456 457 add_entry = kmalloc_obj(*add_entry, GFP_ATOMIC); 458 if (!add_entry) 459 return false; 460 461 list_add(&add_entry->list, &msk->pm.anno_list); 462 463 add_entry->addr = *addr; 464 add_entry->sock = msk; 465 add_entry->retrans_times = 0; 466 467 timer_setup(&add_entry->timer, mptcp_pm_add_addr_timer, 0); 468 reset_timer: 469 add_entry->timer_done = false; 470 timeout = mptcp_adjust_add_addr_timeout(msk); 471 if (timeout) 472 sk_reset_timer(sk, &add_entry->timer, jiffies + timeout); 473 474 return true; 475 } 476 477 static void mptcp_pm_free_announced_list(struct mptcp_sock *msk) 478 { 479 struct mptcp_pm_add_addr *entry, *tmp; 480 struct sock *sk = (struct sock *)msk; 481 LIST_HEAD(free_list); 482 483 pr_debug("msk=%p\n", msk); 484 485 spin_lock_bh(&msk->pm.lock); 486 list_splice_init(&msk->pm.anno_list, &free_list); 487 spin_unlock_bh(&msk->pm.lock); 488 489 list_for_each_entry_safe(entry, tmp, &free_list, list) { 490 if (!entry->timer_done) 491 sk_stop_timer_sync(sk, &entry->timer); 492 kfree_rcu(entry, rcu); 493 } 494 } 495 496 /* path manager command handlers */ 497 498 int mptcp_pm_announce_addr(struct mptcp_sock *msk, 499 const struct mptcp_addr_info *addr, 500 bool echo) 501 { 502 u8 add_addr = READ_ONCE(msk->pm.addr_signal); 503 504 pr_debug("msk=%p, local_id=%d, echo=%d\n", msk, addr->id, echo); 505 506 lockdep_assert_held(&msk->pm.lock); 507 508 if (add_addr & 509 (echo ? BIT(MPTCP_ADD_ADDR_ECHO) : BIT(MPTCP_ADD_ADDR_SIGNAL))) { 510 MPTCP_INC_STATS(sock_net((struct sock *)msk), 511 echo ? MPTCP_MIB_ECHOADDTXDROP : MPTCP_MIB_ADDADDRTXDROP); 512 return -EINVAL; 513 } 514 515 if (echo) { 516 msk->pm.remote = *addr; 517 add_addr |= BIT(MPTCP_ADD_ADDR_ECHO); 518 } else { 519 msk->pm.local = *addr; 520 add_addr |= BIT(MPTCP_ADD_ADDR_SIGNAL); 521 } 522 WRITE_ONCE(msk->pm.addr_signal, add_addr); 523 return 0; 524 } 525 526 int mptcp_pm_remove_addr(struct mptcp_sock *msk, const struct mptcp_rm_list *rm_list) 527 { 528 u8 rm_addr = READ_ONCE(msk->pm.addr_signal); 529 530 pr_debug("msk=%p, rm_list_nr=%d\n", msk, rm_list->nr); 531 532 if (rm_addr) { 533 MPTCP_ADD_STATS(sock_net((struct sock *)msk), 534 MPTCP_MIB_RMADDRTXDROP, rm_list->nr); 535 return -EINVAL; 536 } 537 538 msk->pm.rm_list_tx = *rm_list; 539 rm_addr |= BIT(MPTCP_RM_ADDR_SIGNAL); 540 WRITE_ONCE(msk->pm.addr_signal, rm_addr); 541 mptcp_pm_addr_send_ack_avoid_list(msk, rm_list); 542 return 0; 543 } 544 545 /* path manager event handlers */ 546 547 void mptcp_pm_new_connection(struct mptcp_sock *msk, const struct sock *ssk, int server_side) 548 { 549 struct mptcp_pm_data *pm = &msk->pm; 550 551 pr_debug("msk=%p, token=%u side=%d\n", msk, READ_ONCE(msk->token), server_side); 552 553 WRITE_ONCE(pm->server_side, server_side); 554 mptcp_event(MPTCP_EVENT_CREATED, msk, ssk, GFP_ATOMIC); 555 } 556 557 bool mptcp_pm_allow_new_subflow(struct mptcp_sock *msk) 558 { 559 struct mptcp_pm_data *pm = &msk->pm; 560 unsigned int limit_extra_subflows; 561 int ret = 0; 562 563 if (mptcp_pm_is_userspace(msk)) { 564 if (mptcp_userspace_pm_active(msk)) { 565 spin_lock_bh(&pm->lock); 566 pm->extra_subflows++; 567 spin_unlock_bh(&pm->lock); 568 return true; 569 } 570 return false; 571 } 572 573 limit_extra_subflows = mptcp_pm_get_limit_extra_subflows(msk); 574 575 pr_debug("msk=%p subflows=%d max=%d allow=%d\n", msk, 576 pm->extra_subflows, limit_extra_subflows, 577 READ_ONCE(pm->accept_subflow)); 578 579 /* try to avoid acquiring the lock below */ 580 if (!READ_ONCE(pm->accept_subflow)) 581 return false; 582 583 spin_lock_bh(&pm->lock); 584 if (READ_ONCE(pm->accept_subflow)) { 585 ret = pm->extra_subflows < limit_extra_subflows; 586 if (ret && ++pm->extra_subflows == limit_extra_subflows) 587 WRITE_ONCE(pm->accept_subflow, false); 588 } 589 spin_unlock_bh(&pm->lock); 590 591 return ret; 592 } 593 594 /* return true if the new status bit is currently cleared, that is, this event 595 * can be server, eventually by an already scheduled work 596 */ 597 static bool mptcp_pm_schedule_work(struct mptcp_sock *msk, 598 enum mptcp_pm_status new_status) 599 { 600 pr_debug("msk=%p status=%x new=%lx\n", msk, msk->pm.status, 601 BIT(new_status)); 602 if (msk->pm.status & BIT(new_status)) 603 return false; 604 605 msk->pm.status |= BIT(new_status); 606 mptcp_schedule_work((struct sock *)msk); 607 return true; 608 } 609 610 void mptcp_pm_fully_established(struct mptcp_sock *msk, const struct sock *ssk) 611 { 612 struct mptcp_pm_data *pm = &msk->pm; 613 bool announce = false; 614 615 pr_debug("msk=%p\n", msk); 616 617 spin_lock_bh(&pm->lock); 618 619 /* mptcp_pm_fully_established() can be invoked by multiple 620 * racing paths - accept() and check_fully_established() 621 * be sure to serve this event only once. 622 */ 623 if (READ_ONCE(pm->work_pending) && 624 !(pm->status & BIT(MPTCP_PM_ALREADY_ESTABLISHED))) 625 mptcp_pm_schedule_work(msk, MPTCP_PM_ESTABLISHED); 626 627 if ((pm->status & BIT(MPTCP_PM_ALREADY_ESTABLISHED)) == 0) 628 announce = true; 629 630 pm->status |= BIT(MPTCP_PM_ALREADY_ESTABLISHED); 631 spin_unlock_bh(&pm->lock); 632 633 if (announce) 634 mptcp_event(MPTCP_EVENT_ESTABLISHED, msk, ssk, GFP_ATOMIC); 635 } 636 637 void mptcp_pm_connection_closed(struct mptcp_sock *msk) 638 { 639 pr_debug("msk=%p\n", msk); 640 641 if (msk->token) 642 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL); 643 } 644 645 void mptcp_pm_subflow_established(struct mptcp_sock *msk) 646 { 647 struct mptcp_pm_data *pm = &msk->pm; 648 649 pr_debug("msk=%p\n", msk); 650 651 if (!READ_ONCE(pm->work_pending)) 652 return; 653 654 spin_lock_bh(&pm->lock); 655 656 if (READ_ONCE(pm->work_pending)) 657 mptcp_pm_schedule_work(msk, MPTCP_PM_SUBFLOW_ESTABLISHED); 658 659 spin_unlock_bh(&pm->lock); 660 } 661 662 void mptcp_pm_subflow_check_next(struct mptcp_sock *msk, 663 const struct mptcp_subflow_context *subflow) 664 { 665 struct sock *sk = (struct sock *)msk; 666 struct mptcp_pm_data *pm = &msk->pm; 667 bool update_subflows; 668 669 update_subflows = subflow->request_join || subflow->mp_join; 670 if (mptcp_pm_is_userspace(msk)) { 671 if (update_subflows) { 672 spin_lock_bh(&pm->lock); 673 pm->extra_subflows--; 674 spin_unlock_bh(&pm->lock); 675 } 676 return; 677 } 678 679 if (!READ_ONCE(pm->work_pending) && !update_subflows) 680 return; 681 682 spin_lock_bh(&pm->lock); 683 if (update_subflows) 684 __mptcp_pm_close_subflow(msk); 685 686 /* Even if this subflow is not really established, tell the PM to try 687 * to pick the next ones, if possible. 688 */ 689 if (mptcp_is_fully_established(sk) && 690 mptcp_pm_nl_check_work_pending(msk)) 691 mptcp_pm_schedule_work(msk, MPTCP_PM_SUBFLOW_ESTABLISHED); 692 693 spin_unlock_bh(&pm->lock); 694 } 695 696 void mptcp_pm_add_addr_received(const struct sock *ssk, 697 const struct mptcp_addr_info *addr) 698 { 699 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 700 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 701 struct mptcp_pm_data *pm = &msk->pm; 702 703 pr_debug("msk=%p remote_id=%d accept=%d\n", msk, addr->id, 704 READ_ONCE(pm->accept_addr)); 705 706 mptcp_event_addr_announced(ssk, addr); 707 708 spin_lock_bh(&pm->lock); 709 710 if (mptcp_pm_is_userspace(msk)) { 711 if (mptcp_userspace_pm_active(msk)) { 712 mptcp_pm_announce_addr(msk, addr, true); 713 mptcp_pm_add_addr_send_ack(msk); 714 } else { 715 __MPTCP_INC_STATS(sock_net((struct sock *)msk), MPTCP_MIB_ADDADDRDROP); 716 } 717 /* - id0 should not have a different address 718 * - special case for C-flag: linked to fill_local_addresses_vec() 719 */ 720 } else if ((addr->id == 0 && !mptcp_pm_is_init_remote_addr(msk, addr)) || 721 (addr->id > 0 && !READ_ONCE(pm->accept_addr) && 722 !mptcp_pm_add_addr_c_flag_case(msk))) { 723 mptcp_pm_announce_addr(msk, addr, true); 724 mptcp_pm_add_addr_send_ack(msk); 725 } else if (mptcp_pm_schedule_work(msk, MPTCP_PM_ADD_ADDR_RECEIVED)) { 726 pm->remote = *addr; 727 } else { 728 __MPTCP_INC_STATS(sock_net((struct sock *)msk), MPTCP_MIB_ADDADDRDROP); 729 } 730 731 spin_unlock_bh(&pm->lock); 732 } 733 734 void mptcp_pm_add_addr_echoed(struct mptcp_sock *msk, 735 const struct mptcp_addr_info *addr) 736 { 737 struct mptcp_pm_data *pm = &msk->pm; 738 struct mptcp_pm_add_addr *entry; 739 740 pr_debug("msk=%p\n", msk); 741 742 entry = mptcp_pm_announced_del_timer(msk, addr, true); 743 744 if (!entry || !READ_ONCE(pm->work_pending)) 745 return; 746 747 spin_lock_bh(&pm->lock); 748 749 if (READ_ONCE(pm->work_pending)) 750 mptcp_pm_schedule_work(msk, MPTCP_PM_SUBFLOW_ESTABLISHED); 751 752 spin_unlock_bh(&pm->lock); 753 } 754 755 /* To be called while pm->lock is held */ 756 static void mptcp_pm_add_addr_send_ack(struct mptcp_sock *msk) 757 { 758 if (!mptcp_pm_should_add_signal(msk)) 759 return; 760 761 mptcp_pm_schedule_work(msk, MPTCP_PM_ADD_ADDR_SEND_ACK); 762 } 763 764 static void mptcp_pm_rm_addr_or_subflow(struct mptcp_sock *msk, 765 const struct mptcp_rm_list *rm_list, 766 enum linux_mptcp_mib_field rm_type) 767 { 768 struct mptcp_subflow_context *subflow, *tmp; 769 struct sock *sk = (struct sock *)msk; 770 u8 i; 771 772 pr_debug("%s rm_list_nr %d\n", 773 rm_type == MPTCP_MIB_RMADDR ? "address" : "subflow", rm_list->nr); 774 775 msk_owned_by_me(msk); 776 777 if (sk->sk_state == TCP_LISTEN) 778 return; 779 780 if (!rm_list->nr) 781 return; 782 783 if (list_empty(&msk->conn_list)) 784 return; 785 786 for (i = 0; i < rm_list->nr; i++) { 787 u8 rm_id = rm_list->ids[i]; 788 bool removed = false; 789 790 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 791 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 792 u8 remote_id = READ_ONCE(subflow->remote_id); 793 int how = RCV_SHUTDOWN | SEND_SHUTDOWN; 794 u8 id = subflow_get_local_id(subflow); 795 796 if ((1 << inet_sk_state_load(ssk)) & 797 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSING | TCPF_CLOSE)) 798 continue; 799 if (rm_type == MPTCP_MIB_RMADDR && remote_id != rm_id) 800 continue; 801 if (rm_type == MPTCP_MIB_RMSUBFLOW && id != rm_id) 802 continue; 803 804 pr_debug(" -> %s rm_list_ids[%d]=%u local_id=%u remote_id=%u mpc_id=%u\n", 805 rm_type == MPTCP_MIB_RMADDR ? "address" : "subflow", 806 i, rm_id, id, remote_id, msk->mpc_endpoint_id); 807 spin_unlock_bh(&msk->pm.lock); 808 mptcp_subflow_shutdown(sk, ssk, how); 809 removed |= subflow->request_join; 810 811 /* the following takes care of updating the subflows counter */ 812 mptcp_close_ssk(sk, ssk, subflow); 813 spin_lock_bh(&msk->pm.lock); 814 815 if (rm_type == MPTCP_MIB_RMSUBFLOW) 816 __MPTCP_INC_STATS(sock_net(sk), rm_type); 817 } 818 819 if (rm_type == MPTCP_MIB_RMADDR) { 820 __MPTCP_INC_STATS(sock_net(sk), rm_type); 821 if (removed && mptcp_pm_is_kernel(msk)) 822 mptcp_pm_nl_rm_addr(msk, rm_id); 823 } 824 } 825 } 826 827 static void mptcp_pm_rm_addr_recv(struct mptcp_sock *msk) 828 { 829 mptcp_pm_rm_addr_or_subflow(msk, &msk->pm.rm_list_rx, MPTCP_MIB_RMADDR); 830 } 831 832 void mptcp_pm_rm_subflow(struct mptcp_sock *msk, 833 const struct mptcp_rm_list *rm_list) 834 { 835 mptcp_pm_rm_addr_or_subflow(msk, rm_list, MPTCP_MIB_RMSUBFLOW); 836 } 837 838 void mptcp_pm_rm_addr_received(struct mptcp_sock *msk, 839 const struct mptcp_rm_list *rm_list) 840 { 841 struct mptcp_pm_data *pm = &msk->pm; 842 u8 i; 843 844 pr_debug("msk=%p remote_ids_nr=%d\n", msk, rm_list->nr); 845 846 for (i = 0; i < rm_list->nr; i++) 847 mptcp_event_addr_removed(msk, rm_list->ids[i]); 848 849 spin_lock_bh(&pm->lock); 850 if (mptcp_pm_schedule_work(msk, MPTCP_PM_RM_ADDR_RECEIVED)) 851 pm->rm_list_rx = *rm_list; 852 else 853 __MPTCP_INC_STATS(sock_net((struct sock *)msk), MPTCP_MIB_RMADDRDROP); 854 spin_unlock_bh(&pm->lock); 855 } 856 857 void mptcp_pm_mp_prio_received(struct sock *ssk, u8 bkup) 858 { 859 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 860 struct sock *sk = subflow->conn; 861 struct mptcp_sock *msk; 862 863 pr_debug("subflow->backup=%d, bkup=%d\n", subflow->backup, bkup); 864 msk = mptcp_sk(sk); 865 if (subflow->backup != bkup) 866 subflow->backup = bkup; 867 868 mptcp_event(MPTCP_EVENT_SUB_PRIORITY, msk, ssk, GFP_ATOMIC); 869 } 870 871 void mptcp_pm_mp_fail_received(struct sock *sk, u64 fail_seq) 872 { 873 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 874 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 875 876 pr_debug("fail_seq=%llu\n", fail_seq); 877 878 /* After accepting the fail, we can't create any other subflows */ 879 spin_lock_bh(&msk->fallback_lock); 880 if (!msk->allow_infinite_fallback) { 881 spin_unlock_bh(&msk->fallback_lock); 882 return; 883 } 884 msk->allow_subflows = false; 885 spin_unlock_bh(&msk->fallback_lock); 886 887 if (!subflow->fail_tout) { 888 pr_debug("send MP_FAIL response and infinite map\n"); 889 890 subflow->send_mp_fail = 1; 891 subflow->send_infinite_map = 1; 892 tcp_send_ack(sk); 893 } else { 894 pr_debug("MP_FAIL response received\n"); 895 WRITE_ONCE(subflow->fail_tout, 0); 896 } 897 } 898 899 static int mptcp_add_addr_len(int family, bool echo, bool port) 900 { 901 int len = TCPOLEN_MPTCP_ADD_ADDR_BASE; 902 903 if (family == AF_INET6) 904 len = TCPOLEN_MPTCP_ADD_ADDR6_BASE; 905 if (!echo) 906 len += MPTCPOPT_THMAC_LEN; 907 /* account for 2 trailing 'nop' options */ 908 if (port) 909 len += TCPOLEN_MPTCP_PORT_LEN + TCPOLEN_MPTCP_PORT_ALIGN; 910 911 return len; 912 } 913 914 bool mptcp_pm_add_addr_signal(struct mptcp_sock *msk, int *size, int remaining, 915 struct mptcp_addr_info *addr, bool *echo, 916 bool *drop_ts) 917 { 918 bool skip_add_addr = false; 919 bool ret = false; 920 u8 add_addr; 921 int len = 0; 922 u8 family; 923 bool port; 924 925 spin_lock_bh(&msk->pm.lock); 926 927 /* double check after the lock is acquired */ 928 if (!mptcp_pm_should_add_signal(msk)) 929 goto out_unlock; 930 931 /* always drop every other options for pure ack ADD_ADDR; this is a 932 * plain dup-ack from TCP perspective. The other MPTCP-relevant info, 933 * if any, will be carried by the 'original' TCP ack 934 */ 935 len -= *size; 936 937 *echo = mptcp_pm_should_add_signal_echo(msk); 938 if (*echo) { 939 *addr = msk->pm.remote; 940 add_addr = msk->pm.addr_signal & ~BIT(MPTCP_ADD_ADDR_ECHO); 941 port = !!msk->pm.remote.port; 942 family = msk->pm.remote.family; 943 } else { 944 *addr = msk->pm.local; 945 add_addr = msk->pm.addr_signal & ~BIT(MPTCP_ADD_ADDR_SIGNAL); 946 port = !!msk->pm.local.port; 947 family = msk->pm.local.family; 948 } 949 950 len += mptcp_add_addr_len(family, *echo, port); 951 if (len > remaining) { 952 struct net *net = sock_net((struct sock *)msk); 953 954 if (*drop_ts && mptcp_add_addr_v6_port_drop_ts(net)) { 955 /* OK without TCP Timestamps? */ 956 len -= TCPOLEN_TSTAMP_ALIGNED; 957 if (len <= remaining) 958 goto enough_space; 959 } 960 961 if (*echo) { 962 MPTCP_INC_STATS(net, MPTCP_MIB_ECHOADDTXDROP); 963 } else { 964 skip_add_addr = true; 965 MPTCP_INC_STATS(net, MPTCP_MIB_ADDADDRTXDROP); 966 } 967 goto drop_signal_mark; 968 } 969 970 *drop_ts = false; 971 972 enough_space: 973 ret = true; 974 *size = len; 975 976 drop_signal_mark: 977 WRITE_ONCE(msk->pm.addr_signal, add_addr); 978 979 out_unlock: 980 spin_unlock_bh(&msk->pm.lock); 981 982 /* On pure-ACK option-space exhaustion, stop retrying this ADD_ADDR: 983 * clear the signal bit, cancel the matching retransmission timer, and 984 * let the PM state machine progress. 985 */ 986 if (skip_add_addr) { 987 mptcp_pm_announced_del_timer(msk, addr, true); 988 mptcp_pm_subflow_established(msk); 989 } 990 return ret; 991 } 992 993 static int mptcp_rm_addr_len(const struct mptcp_rm_list *rm_list) 994 { 995 if (rm_list->nr == 0 || rm_list->nr > MPTCP_RM_IDS_MAX) 996 return -EINVAL; 997 998 return TCPOLEN_MPTCP_RM_ADDR_BASE + roundup(rm_list->nr - 1, 4) + 1; 999 } 1000 1001 bool mptcp_pm_rm_addr_signal(struct mptcp_sock *msk, unsigned int remaining, 1002 struct mptcp_rm_list *rm_list, int *size) 1003 { 1004 int ret = false, len; 1005 u8 rm_addr; 1006 1007 spin_lock_bh(&msk->pm.lock); 1008 1009 /* double check after the lock is acquired */ 1010 if (!mptcp_pm_should_rm_signal(msk)) 1011 goto out_unlock; 1012 1013 rm_addr = msk->pm.addr_signal & ~BIT(MPTCP_RM_ADDR_SIGNAL); 1014 len = mptcp_rm_addr_len(&msk->pm.rm_list_tx); 1015 if (len < 0) { 1016 WRITE_ONCE(msk->pm.addr_signal, rm_addr); 1017 goto out_unlock; 1018 } 1019 if (remaining < len) 1020 goto out_unlock; 1021 1022 *size = len; 1023 *rm_list = msk->pm.rm_list_tx; 1024 WRITE_ONCE(msk->pm.addr_signal, rm_addr); 1025 ret = true; 1026 1027 out_unlock: 1028 spin_unlock_bh(&msk->pm.lock); 1029 return ret; 1030 } 1031 1032 int mptcp_pm_get_local_id(struct mptcp_sock *msk, struct sock_common *skc) 1033 { 1034 struct mptcp_pm_addr_entry skc_local = { 0 }; 1035 struct mptcp_addr_info msk_local; 1036 1037 if (WARN_ON_ONCE(!msk)) 1038 return -1; 1039 1040 /* The 0 ID mapping is defined by the first subflow, copied into the msk 1041 * addr 1042 */ 1043 mptcp_local_address((struct sock_common *)msk, &msk_local); 1044 mptcp_local_address((struct sock_common *)skc, &skc_local.addr); 1045 if (mptcp_addresses_equal(&msk_local, &skc_local.addr, false)) 1046 return 0; 1047 1048 skc_local.addr.id = 0; 1049 skc_local.flags = MPTCP_PM_ADDR_FLAG_IMPLICIT; 1050 1051 if (mptcp_pm_is_userspace(msk)) 1052 return mptcp_userspace_pm_get_local_id(msk, &skc_local); 1053 return mptcp_pm_nl_get_local_id(msk, &skc_local); 1054 } 1055 1056 bool mptcp_pm_is_backup(struct mptcp_sock *msk, struct sock_common *skc) 1057 { 1058 struct mptcp_addr_info skc_local; 1059 1060 mptcp_local_address((struct sock_common *)skc, &skc_local); 1061 1062 if (mptcp_pm_is_userspace(msk)) 1063 return mptcp_userspace_pm_is_backup(msk, &skc_local); 1064 1065 return mptcp_pm_nl_is_backup(msk, &skc_local); 1066 } 1067 1068 static void mptcp_pm_subflows_chk_stale(const struct mptcp_sock *msk, struct sock *ssk) 1069 { 1070 struct mptcp_subflow_context *iter, *subflow = mptcp_subflow_ctx(ssk); 1071 struct sock *sk = (struct sock *)msk; 1072 unsigned int active_max_loss_cnt; 1073 struct net *net = sock_net(sk); 1074 unsigned int stale_loss_cnt; 1075 bool slow; 1076 1077 stale_loss_cnt = mptcp_stale_loss_cnt(net); 1078 if (subflow->stale || !stale_loss_cnt || subflow->stale_count <= stale_loss_cnt) 1079 return; 1080 1081 /* look for another available subflow not in loss state */ 1082 active_max_loss_cnt = max_t(int, stale_loss_cnt - 1, 1); 1083 mptcp_for_each_subflow(msk, iter) { 1084 if (iter != subflow && mptcp_subflow_active(iter) && 1085 iter->stale_count < active_max_loss_cnt) { 1086 /* we have some alternatives, try to mark this subflow as idle ...*/ 1087 slow = lock_sock_fast(ssk); 1088 if (!tcp_rtx_and_write_queues_empty(ssk)) { 1089 subflow->stale = 1; 1090 __mptcp_retransmit_pending_data(sk); 1091 MPTCP_INC_STATS(net, MPTCP_MIB_SUBFLOWSTALE); 1092 } 1093 unlock_sock_fast(ssk, slow); 1094 1095 /* always try to push the pending data regardless of re-injections: 1096 * we can possibly use backup subflows now, and subflow selection 1097 * is cheap under the msk socket lock 1098 */ 1099 __mptcp_push_pending(sk, 0); 1100 return; 1101 } 1102 } 1103 } 1104 1105 void mptcp_pm_subflow_chk_stale(const struct mptcp_sock *msk, struct sock *ssk) 1106 { 1107 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1108 u32 rcv_tstamp = READ_ONCE(tcp_sk(ssk)->rcv_tstamp); 1109 1110 /* keep track of rtx periods with no progress */ 1111 if (!subflow->stale_count) { 1112 subflow->stale_rcv_tstamp = rcv_tstamp; 1113 subflow->stale_count++; 1114 } else if (subflow->stale_rcv_tstamp == rcv_tstamp) { 1115 if (subflow->stale_count < U8_MAX) 1116 subflow->stale_count++; 1117 mptcp_pm_subflows_chk_stale(msk, ssk); 1118 } else { 1119 subflow->stale_count = 0; 1120 mptcp_subflow_set_active(subflow); 1121 } 1122 } 1123 1124 void mptcp_pm_worker(struct mptcp_sock *msk) 1125 { 1126 struct mptcp_pm_data *pm = &msk->pm; 1127 1128 msk_owned_by_me(msk); 1129 1130 if (!(pm->status & MPTCP_PM_WORK_MASK)) 1131 return; 1132 1133 spin_lock_bh(&msk->pm.lock); 1134 1135 pr_debug("msk=%p status=%x\n", msk, pm->status); 1136 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_SEND_ACK)) { 1137 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_SEND_ACK); 1138 mptcp_pm_addr_send_ack(msk); 1139 } 1140 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) { 1141 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED); 1142 mptcp_pm_rm_addr_recv(msk); 1143 } 1144 __mptcp_pm_kernel_worker(msk); 1145 1146 spin_unlock_bh(&msk->pm.lock); 1147 } 1148 1149 void mptcp_pm_destroy(struct mptcp_sock *msk) 1150 { 1151 spin_lock_bh(&msk->pm.lock); 1152 msk->pm.status |= BIT(MPTCP_PM_DESTROYING); 1153 spin_unlock_bh(&msk->pm.lock); 1154 1155 mptcp_pm_free_announced_list(msk); 1156 1157 /* Free the userspace local address list unconditionally: the socket 1158 * can be reused (mptcp_disconnect()) and re-selected to a different PM 1159 */ 1160 mptcp_userspace_pm_free_local_addr_list(msk); 1161 } 1162 1163 void mptcp_pm_data_reset(struct mptcp_sock *msk) 1164 { 1165 u8 pm_type = mptcp_get_pm_type(sock_net((struct sock *)msk)); 1166 struct mptcp_pm_data *pm = &msk->pm; 1167 1168 memset(&pm->reset, 0, sizeof(pm->reset)); 1169 pm->rm_list_tx.nr = 0; 1170 pm->rm_list_rx.nr = 0; 1171 WRITE_ONCE(pm->pm_type, pm_type); 1172 1173 if (pm_type == MPTCP_PM_TYPE_KERNEL) { 1174 bool subflows_allowed = !!mptcp_pm_get_limit_extra_subflows(msk); 1175 1176 /* pm->work_pending must be only be set to 'true' when 1177 * pm->pm_type is set to MPTCP_PM_TYPE_KERNEL 1178 */ 1179 WRITE_ONCE(pm->work_pending, 1180 (!!mptcp_pm_get_endp_subflow_max(msk) && 1181 subflows_allowed) || 1182 !!mptcp_pm_get_endp_signal_max(msk)); 1183 WRITE_ONCE(pm->accept_addr, 1184 !!mptcp_pm_get_limit_add_addr_accepted(msk) && 1185 subflows_allowed); 1186 WRITE_ONCE(pm->accept_subflow, subflows_allowed); 1187 1188 bitmap_fill(pm->id_avail_bitmap, MPTCP_PM_MAX_ADDR_ID + 1); 1189 } 1190 } 1191 1192 void mptcp_pm_data_init(struct mptcp_sock *msk) 1193 { 1194 spin_lock_init(&msk->pm.lock); 1195 INIT_LIST_HEAD(&msk->pm.anno_list); 1196 INIT_LIST_HEAD(&msk->pm.userspace_pm_local_addr_list); 1197 mptcp_pm_data_reset(msk); 1198 } 1199 1200 void __init mptcp_pm_init(void) 1201 { 1202 mptcp_pm_kernel_register(); 1203 mptcp_pm_userspace_register(); 1204 mptcp_pm_nl_init(); 1205 } 1206 1207 /* Must be called with rcu read lock held */ 1208 struct mptcp_pm_ops *mptcp_pm_find(const char *name) 1209 { 1210 struct mptcp_pm_ops *pm_ops; 1211 1212 list_for_each_entry_rcu(pm_ops, &mptcp_pm_list, list) { 1213 if (!strcmp(pm_ops->name, name)) 1214 return pm_ops; 1215 } 1216 1217 return NULL; 1218 } 1219 1220 int mptcp_pm_validate(struct mptcp_pm_ops *pm_ops) 1221 { 1222 return 0; 1223 } 1224 1225 int mptcp_pm_register(struct mptcp_pm_ops *pm_ops) 1226 { 1227 int ret; 1228 1229 ret = mptcp_pm_validate(pm_ops); 1230 if (ret) 1231 return ret; 1232 1233 spin_lock(&mptcp_pm_list_lock); 1234 if (mptcp_pm_find(pm_ops->name)) { 1235 spin_unlock(&mptcp_pm_list_lock); 1236 return -EEXIST; 1237 } 1238 list_add_tail_rcu(&pm_ops->list, &mptcp_pm_list); 1239 spin_unlock(&mptcp_pm_list_lock); 1240 1241 pr_debug("%s registered\n", pm_ops->name); 1242 return 0; 1243 } 1244 1245 void mptcp_pm_unregister(struct mptcp_pm_ops *pm_ops) 1246 { 1247 /* skip unregistering the default path manager */ 1248 if (WARN_ON_ONCE(pm_ops == &mptcp_pm_kernel)) 1249 return; 1250 1251 spin_lock(&mptcp_pm_list_lock); 1252 list_del_rcu(&pm_ops->list); 1253 spin_unlock(&mptcp_pm_list_lock); 1254 } 1255 1256 /* Build string with list of available path manager values. 1257 * Similar to tcp_get_available_congestion_control() 1258 */ 1259 void mptcp_pm_get_available(char *buf, size_t maxlen) 1260 { 1261 struct mptcp_pm_ops *pm_ops; 1262 size_t offs = 0; 1263 1264 rcu_read_lock(); 1265 list_for_each_entry_rcu(pm_ops, &mptcp_pm_list, list) { 1266 offs += snprintf(buf + offs, maxlen - offs, "%s%s", 1267 offs == 0 ? "" : " ", pm_ops->name); 1268 1269 if (WARN_ON_ONCE(offs >= maxlen)) 1270 break; 1271 } 1272 rcu_read_unlock(); 1273 } 1274