1 // SPDX-License-Identifier: GPL-2.0 2 /* Multipath TCP 3 * 4 * Copyright (c) 2017 - 2019, Intel Corporation. 5 */ 6 7 #define pr_fmt(fmt) "MPTCP: " fmt 8 9 #include <linux/kernel.h> 10 #include <linux/module.h> 11 #include <linux/netdevice.h> 12 #include <linux/sched/signal.h> 13 #include <linux/atomic.h> 14 #include <net/sock.h> 15 #include <net/inet_common.h> 16 #include <net/inet_hashtables.h> 17 #include <net/protocol.h> 18 #include <net/tcp.h> 19 #include <net/tcp_states.h> 20 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 21 #include <net/transp_v6.h> 22 #endif 23 #include <net/mptcp.h> 24 #include <net/xfrm.h> 25 #include "protocol.h" 26 #include "mib.h" 27 28 #define CREATE_TRACE_POINTS 29 #include <trace/events/mptcp.h> 30 31 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 32 struct mptcp6_sock { 33 struct mptcp_sock msk; 34 struct ipv6_pinfo np; 35 }; 36 #endif 37 38 struct mptcp_skb_cb { 39 u64 map_seq; 40 u64 end_seq; 41 u32 offset; 42 u8 has_rxtstamp:1; 43 }; 44 45 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0])) 46 47 enum { 48 MPTCP_CMSG_TS = BIT(0), 49 }; 50 51 static struct percpu_counter mptcp_sockets_allocated; 52 53 static void __mptcp_destroy_sock(struct sock *sk); 54 static void __mptcp_check_send_data_fin(struct sock *sk); 55 56 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions); 57 static struct net_device mptcp_napi_dev; 58 59 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not 60 * completed yet or has failed, return the subflow socket. 61 * Otherwise return NULL. 62 */ 63 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk) 64 { 65 if (!msk->subflow || READ_ONCE(msk->can_ack)) 66 return NULL; 67 68 return msk->subflow; 69 } 70 71 /* Returns end sequence number of the receiver's advertised window */ 72 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 73 { 74 return READ_ONCE(msk->wnd_end); 75 } 76 77 static bool mptcp_is_tcpsk(struct sock *sk) 78 { 79 struct socket *sock = sk->sk_socket; 80 81 if (unlikely(sk->sk_prot == &tcp_prot)) { 82 /* we are being invoked after mptcp_accept() has 83 * accepted a non-mp-capable flow: sk is a tcp_sk, 84 * not an mptcp one. 85 * 86 * Hand the socket over to tcp so all further socket ops 87 * bypass mptcp. 88 */ 89 sock->ops = &inet_stream_ops; 90 return true; 91 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 92 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) { 93 sock->ops = &inet6_stream_ops; 94 return true; 95 #endif 96 } 97 98 return false; 99 } 100 101 static int __mptcp_socket_create(struct mptcp_sock *msk) 102 { 103 struct mptcp_subflow_context *subflow; 104 struct sock *sk = (struct sock *)msk; 105 struct socket *ssock; 106 int err; 107 108 err = mptcp_subflow_create_socket(sk, &ssock); 109 if (err) 110 return err; 111 112 msk->first = ssock->sk; 113 msk->subflow = ssock; 114 subflow = mptcp_subflow_ctx(ssock->sk); 115 list_add(&subflow->node, &msk->conn_list); 116 sock_hold(ssock->sk); 117 subflow->request_mptcp = 1; 118 mptcp_sock_graft(msk->first, sk->sk_socket); 119 120 return 0; 121 } 122 123 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 124 { 125 sk_drops_add(sk, skb); 126 __kfree_skb(skb); 127 } 128 129 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 130 struct sk_buff *from) 131 { 132 bool fragstolen; 133 int delta; 134 135 if (MPTCP_SKB_CB(from)->offset || 136 !skb_try_coalesce(to, from, &fragstolen, &delta)) 137 return false; 138 139 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx", 140 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 141 to->len, MPTCP_SKB_CB(from)->end_seq); 142 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 143 kfree_skb_partial(from, fragstolen); 144 atomic_add(delta, &sk->sk_rmem_alloc); 145 sk_mem_charge(sk, delta); 146 return true; 147 } 148 149 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 150 struct sk_buff *from) 151 { 152 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 153 return false; 154 155 return mptcp_try_coalesce((struct sock *)msk, to, from); 156 } 157 158 /* "inspired" by tcp_data_queue_ofo(), main differences: 159 * - use mptcp seqs 160 * - don't cope with sacks 161 */ 162 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 163 { 164 struct sock *sk = (struct sock *)msk; 165 struct rb_node **p, *parent; 166 u64 seq, end_seq, max_seq; 167 struct sk_buff *skb1; 168 169 seq = MPTCP_SKB_CB(skb)->map_seq; 170 end_seq = MPTCP_SKB_CB(skb)->end_seq; 171 max_seq = READ_ONCE(msk->rcv_wnd_sent); 172 173 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq, 174 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 175 if (after64(end_seq, max_seq)) { 176 /* out of window */ 177 mptcp_drop(sk, skb); 178 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 179 (unsigned long long)end_seq - (unsigned long)max_seq, 180 (unsigned long long)msk->rcv_wnd_sent); 181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 182 return; 183 } 184 185 p = &msk->out_of_order_queue.rb_node; 186 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 187 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 188 rb_link_node(&skb->rbnode, NULL, p); 189 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 190 msk->ooo_last_skb = skb; 191 goto end; 192 } 193 194 /* with 2 subflows, adding at end of ooo queue is quite likely 195 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 196 */ 197 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 198 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 199 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 200 return; 201 } 202 203 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 204 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 205 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 206 parent = &msk->ooo_last_skb->rbnode; 207 p = &parent->rb_right; 208 goto insert; 209 } 210 211 /* Find place to insert this segment. Handle overlaps on the way. */ 212 parent = NULL; 213 while (*p) { 214 parent = *p; 215 skb1 = rb_to_skb(parent); 216 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 217 p = &parent->rb_left; 218 continue; 219 } 220 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 221 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 222 /* All the bits are present. Drop. */ 223 mptcp_drop(sk, skb); 224 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 225 return; 226 } 227 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 228 /* partial overlap: 229 * | skb | 230 * | skb1 | 231 * continue traversing 232 */ 233 } else { 234 /* skb's seq == skb1's seq and skb covers skb1. 235 * Replace skb1 with skb. 236 */ 237 rb_replace_node(&skb1->rbnode, &skb->rbnode, 238 &msk->out_of_order_queue); 239 mptcp_drop(sk, skb1); 240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 241 goto merge_right; 242 } 243 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 245 return; 246 } 247 p = &parent->rb_right; 248 } 249 250 insert: 251 /* Insert segment into RB tree. */ 252 rb_link_node(&skb->rbnode, parent, p); 253 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 254 255 merge_right: 256 /* Remove other segments covered by skb. */ 257 while ((skb1 = skb_rb_next(skb)) != NULL) { 258 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 259 break; 260 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 261 mptcp_drop(sk, skb1); 262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 263 } 264 /* If there is no skb after us, we are the last_skb ! */ 265 if (!skb1) 266 msk->ooo_last_skb = skb; 267 268 end: 269 skb_condense(skb); 270 skb_set_owner_r(skb, sk); 271 } 272 273 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 274 struct sk_buff *skb, unsigned int offset, 275 size_t copy_len) 276 { 277 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 278 struct sock *sk = (struct sock *)msk; 279 struct sk_buff *tail; 280 bool has_rxtstamp; 281 282 __skb_unlink(skb, &ssk->sk_receive_queue); 283 284 skb_ext_reset(skb); 285 skb_orphan(skb); 286 287 /* try to fetch required memory from subflow */ 288 if (!sk_rmem_schedule(sk, skb, skb->truesize)) { 289 int amount = sk_mem_pages(skb->truesize) << SK_MEM_QUANTUM_SHIFT; 290 291 if (ssk->sk_forward_alloc < amount) 292 goto drop; 293 294 ssk->sk_forward_alloc -= amount; 295 sk->sk_forward_alloc += amount; 296 } 297 298 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 299 300 /* the skb map_seq accounts for the skb offset: 301 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 302 * value 303 */ 304 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 305 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 306 MPTCP_SKB_CB(skb)->offset = offset; 307 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 308 309 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 310 /* in sequence */ 311 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 312 tail = skb_peek_tail(&sk->sk_receive_queue); 313 if (tail && mptcp_try_coalesce(sk, tail, skb)) 314 return true; 315 316 skb_set_owner_r(skb, sk); 317 __skb_queue_tail(&sk->sk_receive_queue, skb); 318 return true; 319 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 320 mptcp_data_queue_ofo(msk, skb); 321 return false; 322 } 323 324 /* old data, keep it simple and drop the whole pkt, sender 325 * will retransmit as needed, if needed. 326 */ 327 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 328 drop: 329 mptcp_drop(sk, skb); 330 return false; 331 } 332 333 static void mptcp_stop_timer(struct sock *sk) 334 { 335 struct inet_connection_sock *icsk = inet_csk(sk); 336 337 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 338 mptcp_sk(sk)->timer_ival = 0; 339 } 340 341 static void mptcp_close_wake_up(struct sock *sk) 342 { 343 if (sock_flag(sk, SOCK_DEAD)) 344 return; 345 346 sk->sk_state_change(sk); 347 if (sk->sk_shutdown == SHUTDOWN_MASK || 348 sk->sk_state == TCP_CLOSE) 349 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 350 else 351 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 352 } 353 354 static bool mptcp_pending_data_fin_ack(struct sock *sk) 355 { 356 struct mptcp_sock *msk = mptcp_sk(sk); 357 358 return !__mptcp_check_fallback(msk) && 359 ((1 << sk->sk_state) & 360 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 361 msk->write_seq == READ_ONCE(msk->snd_una); 362 } 363 364 static void mptcp_check_data_fin_ack(struct sock *sk) 365 { 366 struct mptcp_sock *msk = mptcp_sk(sk); 367 368 /* Look for an acknowledged DATA_FIN */ 369 if (mptcp_pending_data_fin_ack(sk)) { 370 WRITE_ONCE(msk->snd_data_fin_enable, 0); 371 372 switch (sk->sk_state) { 373 case TCP_FIN_WAIT1: 374 inet_sk_state_store(sk, TCP_FIN_WAIT2); 375 break; 376 case TCP_CLOSING: 377 case TCP_LAST_ACK: 378 inet_sk_state_store(sk, TCP_CLOSE); 379 break; 380 } 381 382 mptcp_close_wake_up(sk); 383 } 384 } 385 386 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 387 { 388 struct mptcp_sock *msk = mptcp_sk(sk); 389 390 if (READ_ONCE(msk->rcv_data_fin) && 391 ((1 << sk->sk_state) & 392 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 393 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 394 395 if (msk->ack_seq == rcv_data_fin_seq) { 396 if (seq) 397 *seq = rcv_data_fin_seq; 398 399 return true; 400 } 401 } 402 403 return false; 404 } 405 406 static void mptcp_set_datafin_timeout(const struct sock *sk) 407 { 408 struct inet_connection_sock *icsk = inet_csk(sk); 409 410 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX, 411 TCP_RTO_MIN << icsk->icsk_retransmits); 412 } 413 414 static void __mptcp_set_timeout(struct sock *sk, long tout) 415 { 416 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 417 } 418 419 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 420 { 421 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 422 423 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 424 inet_csk(ssk)->icsk_timeout - jiffies : 0; 425 } 426 427 static void mptcp_set_timeout(struct sock *sk) 428 { 429 struct mptcp_subflow_context *subflow; 430 long tout = 0; 431 432 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 433 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 434 __mptcp_set_timeout(sk, tout); 435 } 436 437 static bool tcp_can_send_ack(const struct sock *ssk) 438 { 439 return !((1 << inet_sk_state_load(ssk)) & 440 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 441 } 442 443 static void mptcp_send_ack(struct mptcp_sock *msk) 444 { 445 struct mptcp_subflow_context *subflow; 446 447 mptcp_for_each_subflow(msk, subflow) { 448 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 449 bool slow; 450 451 slow = lock_sock_fast(ssk); 452 if (tcp_can_send_ack(ssk)) 453 tcp_send_ack(ssk); 454 unlock_sock_fast(ssk, slow); 455 } 456 } 457 458 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk) 459 { 460 bool slow; 461 462 slow = lock_sock_fast(ssk); 463 if (tcp_can_send_ack(ssk)) 464 tcp_cleanup_rbuf(ssk, 1); 465 unlock_sock_fast(ssk, slow); 466 } 467 468 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 469 { 470 const struct inet_connection_sock *icsk = inet_csk(ssk); 471 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 472 const struct tcp_sock *tp = tcp_sk(ssk); 473 474 return (ack_pending & ICSK_ACK_SCHED) && 475 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 476 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 477 (rx_empty && ack_pending & 478 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 479 } 480 481 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk) 482 { 483 int old_space = READ_ONCE(msk->old_wspace); 484 struct mptcp_subflow_context *subflow; 485 struct sock *sk = (struct sock *)msk; 486 int space = __mptcp_space(sk); 487 bool cleanup, rx_empty; 488 489 cleanup = (space > 0) && (space >= (old_space << 1)); 490 rx_empty = !__mptcp_rmem(sk); 491 492 mptcp_for_each_subflow(msk, subflow) { 493 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 494 495 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 496 mptcp_subflow_cleanup_rbuf(ssk); 497 } 498 } 499 500 static bool mptcp_check_data_fin(struct sock *sk) 501 { 502 struct mptcp_sock *msk = mptcp_sk(sk); 503 u64 rcv_data_fin_seq; 504 bool ret = false; 505 506 if (__mptcp_check_fallback(msk)) 507 return ret; 508 509 /* Need to ack a DATA_FIN received from a peer while this side 510 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 511 * msk->rcv_data_fin was set when parsing the incoming options 512 * at the subflow level and the msk lock was not held, so this 513 * is the first opportunity to act on the DATA_FIN and change 514 * the msk state. 515 * 516 * If we are caught up to the sequence number of the incoming 517 * DATA_FIN, send the DATA_ACK now and do state transition. If 518 * not caught up, do nothing and let the recv code send DATA_ACK 519 * when catching up. 520 */ 521 522 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 523 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 524 WRITE_ONCE(msk->rcv_data_fin, 0); 525 526 sk->sk_shutdown |= RCV_SHUTDOWN; 527 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 528 set_bit(MPTCP_DATA_READY, &msk->flags); 529 530 switch (sk->sk_state) { 531 case TCP_ESTABLISHED: 532 inet_sk_state_store(sk, TCP_CLOSE_WAIT); 533 break; 534 case TCP_FIN_WAIT1: 535 inet_sk_state_store(sk, TCP_CLOSING); 536 break; 537 case TCP_FIN_WAIT2: 538 inet_sk_state_store(sk, TCP_CLOSE); 539 break; 540 default: 541 /* Other states not expected */ 542 WARN_ON_ONCE(1); 543 break; 544 } 545 546 ret = true; 547 mptcp_send_ack(msk); 548 mptcp_close_wake_up(sk); 549 } 550 return ret; 551 } 552 553 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 554 struct sock *ssk, 555 unsigned int *bytes) 556 { 557 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 558 struct sock *sk = (struct sock *)msk; 559 unsigned int moved = 0; 560 bool more_data_avail; 561 struct tcp_sock *tp; 562 bool done = false; 563 int sk_rbuf; 564 565 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 566 567 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 568 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 569 570 if (unlikely(ssk_rbuf > sk_rbuf)) { 571 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf); 572 sk_rbuf = ssk_rbuf; 573 } 574 } 575 576 pr_debug("msk=%p ssk=%p", msk, ssk); 577 tp = tcp_sk(ssk); 578 do { 579 u32 map_remaining, offset; 580 u32 seq = tp->copied_seq; 581 struct sk_buff *skb; 582 bool fin; 583 584 /* try to move as much data as available */ 585 map_remaining = subflow->map_data_len - 586 mptcp_subflow_get_map_offset(subflow); 587 588 skb = skb_peek(&ssk->sk_receive_queue); 589 if (!skb) { 590 /* if no data is found, a racing workqueue/recvmsg 591 * already processed the new data, stop here or we 592 * can enter an infinite loop 593 */ 594 if (!moved) 595 done = true; 596 break; 597 } 598 599 if (__mptcp_check_fallback(msk)) { 600 /* if we are running under the workqueue, TCP could have 601 * collapsed skbs between dummy map creation and now 602 * be sure to adjust the size 603 */ 604 map_remaining = skb->len; 605 subflow->map_data_len = skb->len; 606 } 607 608 offset = seq - TCP_SKB_CB(skb)->seq; 609 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 610 if (fin) { 611 done = true; 612 seq++; 613 } 614 615 if (offset < skb->len) { 616 size_t len = skb->len - offset; 617 618 if (tp->urg_data) 619 done = true; 620 621 if (__mptcp_move_skb(msk, ssk, skb, offset, len)) 622 moved += len; 623 seq += len; 624 625 if (WARN_ON_ONCE(map_remaining < len)) 626 break; 627 } else { 628 WARN_ON_ONCE(!fin); 629 sk_eat_skb(ssk, skb); 630 done = true; 631 } 632 633 WRITE_ONCE(tp->copied_seq, seq); 634 more_data_avail = mptcp_subflow_data_available(ssk); 635 636 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) { 637 done = true; 638 break; 639 } 640 } while (more_data_avail); 641 642 *bytes += moved; 643 return done; 644 } 645 646 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 647 { 648 struct sock *sk = (struct sock *)msk; 649 struct sk_buff *skb, *tail; 650 bool moved = false; 651 struct rb_node *p; 652 u64 end_seq; 653 654 p = rb_first(&msk->out_of_order_queue); 655 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 656 while (p) { 657 skb = rb_to_skb(p); 658 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 659 break; 660 661 p = rb_next(p); 662 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 663 664 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 665 msk->ack_seq))) { 666 mptcp_drop(sk, skb); 667 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 668 continue; 669 } 670 671 end_seq = MPTCP_SKB_CB(skb)->end_seq; 672 tail = skb_peek_tail(&sk->sk_receive_queue); 673 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 674 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 675 676 /* skip overlapping data, if any */ 677 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d", 678 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 679 delta); 680 MPTCP_SKB_CB(skb)->offset += delta; 681 __skb_queue_tail(&sk->sk_receive_queue, skb); 682 } 683 msk->ack_seq = end_seq; 684 moved = true; 685 } 686 return moved; 687 } 688 689 /* In most cases we will be able to lock the mptcp socket. If its already 690 * owned, we need to defer to the work queue to avoid ABBA deadlock. 691 */ 692 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 693 { 694 struct sock *sk = (struct sock *)msk; 695 unsigned int moved = 0; 696 697 __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 698 __mptcp_ofo_queue(msk); 699 if (unlikely(ssk->sk_err)) { 700 if (!sock_owned_by_user(sk)) 701 __mptcp_error_report(sk); 702 else 703 set_bit(MPTCP_ERROR_REPORT, &msk->flags); 704 } 705 706 /* If the moves have caught up with the DATA_FIN sequence number 707 * it's time to ack the DATA_FIN and change socket state, but 708 * this is not a good place to change state. Let the workqueue 709 * do it. 710 */ 711 if (mptcp_pending_data_fin(sk, NULL)) 712 mptcp_schedule_work(sk); 713 return moved > 0; 714 } 715 716 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 717 { 718 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 719 struct mptcp_sock *msk = mptcp_sk(sk); 720 int sk_rbuf, ssk_rbuf; 721 722 /* The peer can send data while we are shutting down this 723 * subflow at msk destruction time, but we must avoid enqueuing 724 * more data to the msk receive queue 725 */ 726 if (unlikely(subflow->disposable)) 727 return; 728 729 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 730 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 731 if (unlikely(ssk_rbuf > sk_rbuf)) 732 sk_rbuf = ssk_rbuf; 733 734 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/ 735 if (__mptcp_rmem(sk) > sk_rbuf) { 736 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 737 return; 738 } 739 740 /* Wake-up the reader only for in-sequence data */ 741 mptcp_data_lock(sk); 742 if (move_skbs_to_msk(msk, ssk)) { 743 set_bit(MPTCP_DATA_READY, &msk->flags); 744 sk->sk_data_ready(sk); 745 } 746 mptcp_data_unlock(sk); 747 } 748 749 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk) 750 { 751 struct mptcp_subflow_context *subflow; 752 bool ret = false; 753 754 if (likely(list_empty(&msk->join_list))) 755 return false; 756 757 spin_lock_bh(&msk->join_list_lock); 758 list_for_each_entry(subflow, &msk->join_list, node) { 759 u32 sseq = READ_ONCE(subflow->setsockopt_seq); 760 761 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow)); 762 if (READ_ONCE(msk->setsockopt_seq) != sseq) 763 ret = true; 764 } 765 list_splice_tail_init(&msk->join_list, &msk->conn_list); 766 spin_unlock_bh(&msk->join_list_lock); 767 768 return ret; 769 } 770 771 void __mptcp_flush_join_list(struct mptcp_sock *msk) 772 { 773 if (likely(!mptcp_do_flush_join_list(msk))) 774 return; 775 776 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags)) 777 mptcp_schedule_work((struct sock *)msk); 778 } 779 780 static void mptcp_flush_join_list(struct mptcp_sock *msk) 781 { 782 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags); 783 784 might_sleep(); 785 786 if (!mptcp_do_flush_join_list(msk) && !sync_needed) 787 return; 788 789 mptcp_sockopt_sync_all(msk); 790 } 791 792 static bool mptcp_timer_pending(struct sock *sk) 793 { 794 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 795 } 796 797 static void mptcp_reset_timer(struct sock *sk) 798 { 799 struct inet_connection_sock *icsk = inet_csk(sk); 800 unsigned long tout; 801 802 /* prevent rescheduling on close */ 803 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 804 return; 805 806 tout = mptcp_sk(sk)->timer_ival; 807 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 808 } 809 810 bool mptcp_schedule_work(struct sock *sk) 811 { 812 if (inet_sk_state_load(sk) != TCP_CLOSE && 813 schedule_work(&mptcp_sk(sk)->work)) { 814 /* each subflow already holds a reference to the sk, and the 815 * workqueue is invoked by a subflow, so sk can't go away here. 816 */ 817 sock_hold(sk); 818 return true; 819 } 820 return false; 821 } 822 823 void mptcp_subflow_eof(struct sock *sk) 824 { 825 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags)) 826 mptcp_schedule_work(sk); 827 } 828 829 static void mptcp_check_for_eof(struct mptcp_sock *msk) 830 { 831 struct mptcp_subflow_context *subflow; 832 struct sock *sk = (struct sock *)msk; 833 int receivers = 0; 834 835 mptcp_for_each_subflow(msk, subflow) 836 receivers += !subflow->rx_eof; 837 if (receivers) 838 return; 839 840 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) { 841 /* hopefully temporary hack: propagate shutdown status 842 * to msk, when all subflows agree on it 843 */ 844 sk->sk_shutdown |= RCV_SHUTDOWN; 845 846 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 847 set_bit(MPTCP_DATA_READY, &msk->flags); 848 sk->sk_data_ready(sk); 849 } 850 851 switch (sk->sk_state) { 852 case TCP_ESTABLISHED: 853 inet_sk_state_store(sk, TCP_CLOSE_WAIT); 854 break; 855 case TCP_FIN_WAIT1: 856 inet_sk_state_store(sk, TCP_CLOSING); 857 break; 858 case TCP_FIN_WAIT2: 859 inet_sk_state_store(sk, TCP_CLOSE); 860 break; 861 default: 862 return; 863 } 864 mptcp_close_wake_up(sk); 865 } 866 867 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk) 868 { 869 struct mptcp_subflow_context *subflow; 870 struct sock *sk = (struct sock *)msk; 871 872 sock_owned_by_me(sk); 873 874 mptcp_for_each_subflow(msk, subflow) { 875 if (READ_ONCE(subflow->data_avail)) 876 return mptcp_subflow_tcp_sock(subflow); 877 } 878 879 return NULL; 880 } 881 882 static bool mptcp_skb_can_collapse_to(u64 write_seq, 883 const struct sk_buff *skb, 884 const struct mptcp_ext *mpext) 885 { 886 if (!tcp_skb_can_collapse_to(skb)) 887 return false; 888 889 /* can collapse only if MPTCP level sequence is in order and this 890 * mapping has not been xmitted yet 891 */ 892 return mpext && mpext->data_seq + mpext->data_len == write_seq && 893 !mpext->frozen; 894 } 895 896 /* we can append data to the given data frag if: 897 * - there is space available in the backing page_frag 898 * - the data frag tail matches the current page_frag free offset 899 * - the data frag end sequence number matches the current write seq 900 */ 901 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 902 const struct page_frag *pfrag, 903 const struct mptcp_data_frag *df) 904 { 905 return df && pfrag->page == df->page && 906 pfrag->size - pfrag->offset > 0 && 907 pfrag->offset == (df->offset + df->data_len) && 908 df->data_seq + df->data_len == msk->write_seq; 909 } 910 911 static int mptcp_wmem_with_overhead(int size) 912 { 913 return size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT); 914 } 915 916 static void __mptcp_wmem_reserve(struct sock *sk, int size) 917 { 918 int amount = mptcp_wmem_with_overhead(size); 919 struct mptcp_sock *msk = mptcp_sk(sk); 920 921 WARN_ON_ONCE(msk->wmem_reserved); 922 if (WARN_ON_ONCE(amount < 0)) 923 amount = 0; 924 925 if (amount <= sk->sk_forward_alloc) 926 goto reserve; 927 928 /* under memory pressure try to reserve at most a single page 929 * otherwise try to reserve the full estimate and fallback 930 * to a single page before entering the error path 931 */ 932 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) || 933 !sk_wmem_schedule(sk, amount)) { 934 if (amount <= PAGE_SIZE) 935 goto nomem; 936 937 amount = PAGE_SIZE; 938 if (!sk_wmem_schedule(sk, amount)) 939 goto nomem; 940 } 941 942 reserve: 943 msk->wmem_reserved = amount; 944 sk->sk_forward_alloc -= amount; 945 return; 946 947 nomem: 948 /* we will wait for memory on next allocation */ 949 msk->wmem_reserved = -1; 950 } 951 952 static void __mptcp_update_wmem(struct sock *sk) 953 { 954 struct mptcp_sock *msk = mptcp_sk(sk); 955 956 #ifdef CONFIG_LOCKDEP 957 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock)); 958 #endif 959 960 if (!msk->wmem_reserved) 961 return; 962 963 if (msk->wmem_reserved < 0) 964 msk->wmem_reserved = 0; 965 if (msk->wmem_reserved > 0) { 966 sk->sk_forward_alloc += msk->wmem_reserved; 967 msk->wmem_reserved = 0; 968 } 969 } 970 971 static bool mptcp_wmem_alloc(struct sock *sk, int size) 972 { 973 struct mptcp_sock *msk = mptcp_sk(sk); 974 975 /* check for pre-existing error condition */ 976 if (msk->wmem_reserved < 0) 977 return false; 978 979 if (msk->wmem_reserved >= size) 980 goto account; 981 982 mptcp_data_lock(sk); 983 if (!sk_wmem_schedule(sk, size)) { 984 mptcp_data_unlock(sk); 985 return false; 986 } 987 988 sk->sk_forward_alloc -= size; 989 msk->wmem_reserved += size; 990 mptcp_data_unlock(sk); 991 992 account: 993 msk->wmem_reserved -= size; 994 return true; 995 } 996 997 static void mptcp_wmem_uncharge(struct sock *sk, int size) 998 { 999 struct mptcp_sock *msk = mptcp_sk(sk); 1000 1001 if (msk->wmem_reserved < 0) 1002 msk->wmem_reserved = 0; 1003 msk->wmem_reserved += size; 1004 } 1005 1006 static void mptcp_mem_reclaim_partial(struct sock *sk) 1007 { 1008 struct mptcp_sock *msk = mptcp_sk(sk); 1009 1010 /* if we are experiencing a transint allocation error, 1011 * the forward allocation memory has been already 1012 * released 1013 */ 1014 if (msk->wmem_reserved < 0) 1015 return; 1016 1017 mptcp_data_lock(sk); 1018 sk->sk_forward_alloc += msk->wmem_reserved; 1019 sk_mem_reclaim_partial(sk); 1020 msk->wmem_reserved = sk->sk_forward_alloc; 1021 sk->sk_forward_alloc = 0; 1022 mptcp_data_unlock(sk); 1023 } 1024 1025 static void dfrag_uncharge(struct sock *sk, int len) 1026 { 1027 sk_mem_uncharge(sk, len); 1028 sk_wmem_queued_add(sk, -len); 1029 } 1030 1031 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 1032 { 1033 int len = dfrag->data_len + dfrag->overhead; 1034 1035 list_del(&dfrag->list); 1036 dfrag_uncharge(sk, len); 1037 put_page(dfrag->page); 1038 } 1039 1040 static void __mptcp_clean_una(struct sock *sk) 1041 { 1042 struct mptcp_sock *msk = mptcp_sk(sk); 1043 struct mptcp_data_frag *dtmp, *dfrag; 1044 bool cleaned = false; 1045 u64 snd_una; 1046 1047 /* on fallback we just need to ignore snd_una, as this is really 1048 * plain TCP 1049 */ 1050 if (__mptcp_check_fallback(msk)) 1051 msk->snd_una = READ_ONCE(msk->snd_nxt); 1052 1053 snd_una = msk->snd_una; 1054 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 1055 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 1056 break; 1057 1058 if (unlikely(dfrag == msk->first_pending)) { 1059 /* in recovery mode can see ack after the current snd head */ 1060 if (WARN_ON_ONCE(!msk->recovery)) 1061 break; 1062 1063 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1064 } 1065 1066 dfrag_clear(sk, dfrag); 1067 cleaned = true; 1068 } 1069 1070 dfrag = mptcp_rtx_head(sk); 1071 if (dfrag && after64(snd_una, dfrag->data_seq)) { 1072 u64 delta = snd_una - dfrag->data_seq; 1073 1074 /* prevent wrap around in recovery mode */ 1075 if (unlikely(delta > dfrag->already_sent)) { 1076 if (WARN_ON_ONCE(!msk->recovery)) 1077 goto out; 1078 if (WARN_ON_ONCE(delta > dfrag->data_len)) 1079 goto out; 1080 dfrag->already_sent += delta - dfrag->already_sent; 1081 } 1082 1083 dfrag->data_seq += delta; 1084 dfrag->offset += delta; 1085 dfrag->data_len -= delta; 1086 dfrag->already_sent -= delta; 1087 1088 dfrag_uncharge(sk, delta); 1089 cleaned = true; 1090 } 1091 1092 /* all retransmitted data acked, recovery completed */ 1093 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 1094 msk->recovery = false; 1095 1096 out: 1097 if (cleaned) { 1098 if (tcp_under_memory_pressure(sk)) { 1099 __mptcp_update_wmem(sk); 1100 sk_mem_reclaim_partial(sk); 1101 } 1102 } 1103 1104 if (snd_una == READ_ONCE(msk->snd_nxt) && !msk->recovery) { 1105 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 1106 mptcp_stop_timer(sk); 1107 } else { 1108 mptcp_reset_timer(sk); 1109 } 1110 } 1111 1112 static void __mptcp_clean_una_wakeup(struct sock *sk) 1113 { 1114 #ifdef CONFIG_LOCKDEP 1115 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock)); 1116 #endif 1117 __mptcp_clean_una(sk); 1118 mptcp_write_space(sk); 1119 } 1120 1121 static void mptcp_clean_una_wakeup(struct sock *sk) 1122 { 1123 mptcp_data_lock(sk); 1124 __mptcp_clean_una_wakeup(sk); 1125 mptcp_data_unlock(sk); 1126 } 1127 1128 static void mptcp_enter_memory_pressure(struct sock *sk) 1129 { 1130 struct mptcp_subflow_context *subflow; 1131 struct mptcp_sock *msk = mptcp_sk(sk); 1132 bool first = true; 1133 1134 sk_stream_moderate_sndbuf(sk); 1135 mptcp_for_each_subflow(msk, subflow) { 1136 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1137 1138 if (first) 1139 tcp_enter_memory_pressure(ssk); 1140 sk_stream_moderate_sndbuf(ssk); 1141 first = false; 1142 } 1143 } 1144 1145 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1146 * data 1147 */ 1148 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1149 { 1150 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1151 pfrag, sk->sk_allocation))) 1152 return true; 1153 1154 mptcp_enter_memory_pressure(sk); 1155 return false; 1156 } 1157 1158 static struct mptcp_data_frag * 1159 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1160 int orig_offset) 1161 { 1162 int offset = ALIGN(orig_offset, sizeof(long)); 1163 struct mptcp_data_frag *dfrag; 1164 1165 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1166 dfrag->data_len = 0; 1167 dfrag->data_seq = msk->write_seq; 1168 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1169 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1170 dfrag->already_sent = 0; 1171 dfrag->page = pfrag->page; 1172 1173 return dfrag; 1174 } 1175 1176 struct mptcp_sendmsg_info { 1177 int mss_now; 1178 int size_goal; 1179 u16 limit; 1180 u16 sent; 1181 unsigned int flags; 1182 }; 1183 1184 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq, 1185 int avail_size) 1186 { 1187 u64 window_end = mptcp_wnd_end(msk); 1188 1189 if (__mptcp_check_fallback(msk)) 1190 return avail_size; 1191 1192 if (!before64(data_seq + avail_size, window_end)) { 1193 u64 allowed_size = window_end - data_seq; 1194 1195 return min_t(unsigned int, allowed_size, avail_size); 1196 } 1197 1198 return avail_size; 1199 } 1200 1201 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1202 { 1203 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1204 1205 if (!mpext) 1206 return false; 1207 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1208 return true; 1209 } 1210 1211 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1212 { 1213 struct sk_buff *skb; 1214 1215 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1216 if (likely(skb)) { 1217 if (likely(__mptcp_add_ext(skb, gfp))) { 1218 skb_reserve(skb, MAX_TCP_HEADER); 1219 skb->reserved_tailroom = skb->end - skb->tail; 1220 return skb; 1221 } 1222 __kfree_skb(skb); 1223 } else { 1224 mptcp_enter_memory_pressure(sk); 1225 } 1226 return NULL; 1227 } 1228 1229 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1230 { 1231 struct sk_buff *skb; 1232 1233 if (ssk->sk_tx_skb_cache) { 1234 skb = ssk->sk_tx_skb_cache; 1235 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) && 1236 !__mptcp_add_ext(skb, gfp))) 1237 return false; 1238 return true; 1239 } 1240 1241 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1242 if (!skb) 1243 return false; 1244 1245 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1246 ssk->sk_tx_skb_cache = skb; 1247 return true; 1248 } 1249 kfree_skb(skb); 1250 return false; 1251 } 1252 1253 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk) 1254 { 1255 return !ssk->sk_tx_skb_cache && 1256 tcp_under_memory_pressure(sk); 1257 } 1258 1259 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk) 1260 { 1261 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) 1262 mptcp_mem_reclaim_partial(sk); 1263 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation); 1264 } 1265 1266 /* note: this always recompute the csum on the whole skb, even 1267 * if we just appended a single frag. More status info needed 1268 */ 1269 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1270 { 1271 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1272 __wsum csum = ~csum_unfold(mpext->csum); 1273 int offset = skb->len - added; 1274 1275 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1276 } 1277 1278 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1279 struct mptcp_data_frag *dfrag, 1280 struct mptcp_sendmsg_info *info) 1281 { 1282 u64 data_seq = dfrag->data_seq + info->sent; 1283 struct mptcp_sock *msk = mptcp_sk(sk); 1284 bool zero_window_probe = false; 1285 struct mptcp_ext *mpext = NULL; 1286 struct sk_buff *skb, *tail; 1287 bool can_collapse = false; 1288 int size_bias = 0; 1289 int avail_size; 1290 size_t ret = 0; 1291 1292 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u", 1293 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1294 1295 /* compute send limit */ 1296 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1297 avail_size = info->size_goal; 1298 skb = tcp_write_queue_tail(ssk); 1299 if (skb) { 1300 /* Limit the write to the size available in the 1301 * current skb, if any, so that we create at most a new skb. 1302 * Explicitly tells TCP internals to avoid collapsing on later 1303 * queue management operation, to avoid breaking the ext <-> 1304 * SSN association set here 1305 */ 1306 mpext = skb_ext_find(skb, SKB_EXT_MPTCP); 1307 can_collapse = (info->size_goal - skb->len > 0) && 1308 mptcp_skb_can_collapse_to(data_seq, skb, mpext); 1309 if (!can_collapse) { 1310 TCP_SKB_CB(skb)->eor = 1; 1311 } else { 1312 size_bias = skb->len; 1313 avail_size = info->size_goal - skb->len; 1314 } 1315 } 1316 1317 /* Zero window and all data acked? Probe. */ 1318 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size); 1319 if (avail_size == 0) { 1320 u64 snd_una = READ_ONCE(msk->snd_una); 1321 1322 if (skb || snd_una != msk->snd_nxt) 1323 return 0; 1324 zero_window_probe = true; 1325 data_seq = snd_una - 1; 1326 avail_size = 1; 1327 } 1328 1329 if (WARN_ON_ONCE(info->sent > info->limit || 1330 info->limit > dfrag->data_len)) 1331 return 0; 1332 1333 ret = info->limit - info->sent; 1334 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags, 1335 dfrag->page, dfrag->offset + info->sent, &ret); 1336 if (!tail) { 1337 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk)); 1338 return -ENOMEM; 1339 } 1340 1341 /* if the tail skb is still the cached one, collapsing really happened. 1342 */ 1343 if (skb == tail) { 1344 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH; 1345 mpext->data_len += ret; 1346 WARN_ON_ONCE(!can_collapse); 1347 WARN_ON_ONCE(zero_window_probe); 1348 goto out; 1349 } 1350 1351 mpext = skb_ext_find(tail, SKB_EXT_MPTCP); 1352 if (WARN_ON_ONCE(!mpext)) { 1353 /* should never reach here, stream corrupted */ 1354 return -EINVAL; 1355 } 1356 1357 memset(mpext, 0, sizeof(*mpext)); 1358 mpext->data_seq = data_seq; 1359 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1360 mpext->data_len = ret; 1361 mpext->use_map = 1; 1362 mpext->dsn64 = 1; 1363 1364 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d", 1365 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1366 mpext->dsn64); 1367 1368 if (zero_window_probe) { 1369 mptcp_subflow_ctx(ssk)->rel_write_seq += ret; 1370 mpext->frozen = 1; 1371 if (READ_ONCE(msk->csum_enabled)) 1372 mptcp_update_data_checksum(tail, ret); 1373 tcp_push_pending_frames(ssk); 1374 return 0; 1375 } 1376 out: 1377 if (READ_ONCE(msk->csum_enabled)) 1378 mptcp_update_data_checksum(tail, ret); 1379 mptcp_subflow_ctx(ssk)->rel_write_seq += ret; 1380 return ret; 1381 } 1382 1383 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1384 sizeof(struct tcphdr) - \ 1385 MAX_TCP_OPTION_SPACE - \ 1386 sizeof(struct ipv6hdr) - \ 1387 sizeof(struct frag_hdr)) 1388 1389 struct subflow_send_info { 1390 struct sock *ssk; 1391 u64 ratio; 1392 }; 1393 1394 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1395 { 1396 if (!subflow->stale) 1397 return; 1398 1399 subflow->stale = 0; 1400 } 1401 1402 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1403 { 1404 if (unlikely(subflow->stale)) { 1405 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1406 1407 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1408 return false; 1409 1410 mptcp_subflow_set_active(subflow); 1411 } 1412 return __mptcp_subflow_active(subflow); 1413 } 1414 1415 /* implement the mptcp packet scheduler; 1416 * returns the subflow that will transmit the next DSS 1417 * additionally updates the rtx timeout 1418 */ 1419 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1420 { 1421 struct subflow_send_info send_info[2]; 1422 struct mptcp_subflow_context *subflow; 1423 struct sock *sk = (struct sock *)msk; 1424 int i, nr_active = 0; 1425 struct sock *ssk; 1426 long tout = 0; 1427 u64 ratio; 1428 u32 pace; 1429 1430 sock_owned_by_me(sk); 1431 1432 if (__mptcp_check_fallback(msk)) { 1433 if (!msk->first) 1434 return NULL; 1435 return sk_stream_memory_free(msk->first) ? msk->first : NULL; 1436 } 1437 1438 /* re-use last subflow, if the burst allow that */ 1439 if (msk->last_snd && msk->snd_burst > 0 && 1440 sk_stream_memory_free(msk->last_snd) && 1441 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) { 1442 mptcp_set_timeout(sk); 1443 return msk->last_snd; 1444 } 1445 1446 /* pick the subflow with the lower wmem/wspace ratio */ 1447 for (i = 0; i < 2; ++i) { 1448 send_info[i].ssk = NULL; 1449 send_info[i].ratio = -1; 1450 } 1451 mptcp_for_each_subflow(msk, subflow) { 1452 trace_mptcp_subflow_get_send(subflow); 1453 ssk = mptcp_subflow_tcp_sock(subflow); 1454 if (!mptcp_subflow_active(subflow)) 1455 continue; 1456 1457 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1458 nr_active += !subflow->backup; 1459 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd) 1460 continue; 1461 1462 pace = READ_ONCE(ssk->sk_pacing_rate); 1463 if (!pace) 1464 continue; 1465 1466 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, 1467 pace); 1468 if (ratio < send_info[subflow->backup].ratio) { 1469 send_info[subflow->backup].ssk = ssk; 1470 send_info[subflow->backup].ratio = ratio; 1471 } 1472 } 1473 __mptcp_set_timeout(sk, tout); 1474 1475 /* pick the best backup if no other subflow is active */ 1476 if (!nr_active) 1477 send_info[0].ssk = send_info[1].ssk; 1478 1479 if (send_info[0].ssk) { 1480 msk->last_snd = send_info[0].ssk; 1481 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE, 1482 tcp_sk(msk->last_snd)->snd_wnd); 1483 return msk->last_snd; 1484 } 1485 1486 return NULL; 1487 } 1488 1489 static void mptcp_push_release(struct sock *sk, struct sock *ssk, 1490 struct mptcp_sendmsg_info *info) 1491 { 1492 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1493 release_sock(ssk); 1494 } 1495 1496 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1497 { 1498 struct sock *prev_ssk = NULL, *ssk = NULL; 1499 struct mptcp_sock *msk = mptcp_sk(sk); 1500 struct mptcp_sendmsg_info info = { 1501 .flags = flags, 1502 }; 1503 struct mptcp_data_frag *dfrag; 1504 int len, copied = 0; 1505 1506 while ((dfrag = mptcp_send_head(sk))) { 1507 info.sent = dfrag->already_sent; 1508 info.limit = dfrag->data_len; 1509 len = dfrag->data_len - dfrag->already_sent; 1510 while (len > 0) { 1511 int ret = 0; 1512 1513 prev_ssk = ssk; 1514 mptcp_flush_join_list(msk); 1515 ssk = mptcp_subflow_get_send(msk); 1516 1517 /* try to keep the subflow socket lock across 1518 * consecutive xmit on the same socket 1519 */ 1520 if (ssk != prev_ssk && prev_ssk) 1521 mptcp_push_release(sk, prev_ssk, &info); 1522 if (!ssk) 1523 goto out; 1524 1525 if (ssk != prev_ssk || !prev_ssk) 1526 lock_sock(ssk); 1527 1528 /* keep it simple and always provide a new skb for the 1529 * subflow, even if we will not use it when collapsing 1530 * on the pending one 1531 */ 1532 if (!mptcp_alloc_tx_skb(sk, ssk)) { 1533 mptcp_push_release(sk, ssk, &info); 1534 goto out; 1535 } 1536 1537 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 1538 if (ret <= 0) { 1539 mptcp_push_release(sk, ssk, &info); 1540 goto out; 1541 } 1542 1543 info.sent += ret; 1544 dfrag->already_sent += ret; 1545 msk->snd_nxt += ret; 1546 msk->snd_burst -= ret; 1547 msk->tx_pending_data -= ret; 1548 copied += ret; 1549 len -= ret; 1550 } 1551 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1552 } 1553 1554 /* at this point we held the socket lock for the last subflow we used */ 1555 if (ssk) 1556 mptcp_push_release(sk, ssk, &info); 1557 1558 out: 1559 /* ensure the rtx timer is running */ 1560 if (!mptcp_timer_pending(sk)) 1561 mptcp_reset_timer(sk); 1562 if (copied) 1563 __mptcp_check_send_data_fin(sk); 1564 } 1565 1566 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk) 1567 { 1568 struct mptcp_sock *msk = mptcp_sk(sk); 1569 struct mptcp_sendmsg_info info; 1570 struct mptcp_data_frag *dfrag; 1571 struct sock *xmit_ssk; 1572 int len, copied = 0; 1573 bool first = true; 1574 1575 info.flags = 0; 1576 while ((dfrag = mptcp_send_head(sk))) { 1577 info.sent = dfrag->already_sent; 1578 info.limit = dfrag->data_len; 1579 len = dfrag->data_len - dfrag->already_sent; 1580 while (len > 0) { 1581 int ret = 0; 1582 1583 /* the caller already invoked the packet scheduler, 1584 * check for a different subflow usage only after 1585 * spooling the first chunk of data 1586 */ 1587 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk)); 1588 if (!xmit_ssk) 1589 goto out; 1590 if (xmit_ssk != ssk) { 1591 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk)); 1592 goto out; 1593 } 1594 1595 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) { 1596 __mptcp_update_wmem(sk); 1597 sk_mem_reclaim_partial(sk); 1598 } 1599 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC)) 1600 goto out; 1601 1602 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 1603 if (ret <= 0) 1604 goto out; 1605 1606 info.sent += ret; 1607 dfrag->already_sent += ret; 1608 msk->snd_nxt += ret; 1609 msk->snd_burst -= ret; 1610 msk->tx_pending_data -= ret; 1611 copied += ret; 1612 len -= ret; 1613 first = false; 1614 } 1615 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1616 } 1617 1618 out: 1619 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1620 * not going to flush it via release_sock() 1621 */ 1622 __mptcp_update_wmem(sk); 1623 if (copied) { 1624 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1625 info.size_goal); 1626 if (!mptcp_timer_pending(sk)) 1627 mptcp_reset_timer(sk); 1628 1629 if (msk->snd_data_fin_enable && 1630 msk->snd_nxt + 1 == msk->write_seq) 1631 mptcp_schedule_work(sk); 1632 } 1633 } 1634 1635 static void mptcp_set_nospace(struct sock *sk) 1636 { 1637 /* enable autotune */ 1638 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1639 1640 /* will be cleared on avail space */ 1641 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags); 1642 } 1643 1644 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1645 { 1646 struct mptcp_sock *msk = mptcp_sk(sk); 1647 struct page_frag *pfrag; 1648 size_t copied = 0; 1649 int ret = 0; 1650 long timeo; 1651 1652 /* we don't support FASTOPEN yet */ 1653 if (msg->msg_flags & MSG_FASTOPEN) 1654 return -EOPNOTSUPP; 1655 1656 /* silently ignore everything else */ 1657 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL; 1658 1659 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len))); 1660 1661 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1662 1663 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1664 ret = sk_stream_wait_connect(sk, &timeo); 1665 if (ret) 1666 goto out; 1667 } 1668 1669 pfrag = sk_page_frag(sk); 1670 1671 while (msg_data_left(msg)) { 1672 int total_ts, frag_truesize = 0; 1673 struct mptcp_data_frag *dfrag; 1674 bool dfrag_collapsed; 1675 size_t psize, offset; 1676 1677 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) { 1678 ret = -EPIPE; 1679 goto out; 1680 } 1681 1682 /* reuse tail pfrag, if possible, or carve a new one from the 1683 * page allocator 1684 */ 1685 dfrag = mptcp_pending_tail(sk); 1686 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1687 if (!dfrag_collapsed) { 1688 if (!sk_stream_memory_free(sk)) 1689 goto wait_for_memory; 1690 1691 if (!mptcp_page_frag_refill(sk, pfrag)) 1692 goto wait_for_memory; 1693 1694 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1695 frag_truesize = dfrag->overhead; 1696 } 1697 1698 /* we do not bound vs wspace, to allow a single packet. 1699 * memory accounting will prevent execessive memory usage 1700 * anyway 1701 */ 1702 offset = dfrag->offset + dfrag->data_len; 1703 psize = pfrag->size - offset; 1704 psize = min_t(size_t, psize, msg_data_left(msg)); 1705 total_ts = psize + frag_truesize; 1706 1707 if (!mptcp_wmem_alloc(sk, total_ts)) 1708 goto wait_for_memory; 1709 1710 if (copy_page_from_iter(dfrag->page, offset, psize, 1711 &msg->msg_iter) != psize) { 1712 mptcp_wmem_uncharge(sk, psize + frag_truesize); 1713 ret = -EFAULT; 1714 goto out; 1715 } 1716 1717 /* data successfully copied into the write queue */ 1718 copied += psize; 1719 dfrag->data_len += psize; 1720 frag_truesize += psize; 1721 pfrag->offset += frag_truesize; 1722 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1723 msk->tx_pending_data += psize; 1724 1725 /* charge data on mptcp pending queue to the msk socket 1726 * Note: we charge such data both to sk and ssk 1727 */ 1728 sk_wmem_queued_add(sk, frag_truesize); 1729 if (!dfrag_collapsed) { 1730 get_page(dfrag->page); 1731 list_add_tail(&dfrag->list, &msk->rtx_queue); 1732 if (!msk->first_pending) 1733 WRITE_ONCE(msk->first_pending, dfrag); 1734 } 1735 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk, 1736 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1737 !dfrag_collapsed); 1738 1739 continue; 1740 1741 wait_for_memory: 1742 mptcp_set_nospace(sk); 1743 __mptcp_push_pending(sk, msg->msg_flags); 1744 ret = sk_stream_wait_memory(sk, &timeo); 1745 if (ret) 1746 goto out; 1747 } 1748 1749 if (copied) 1750 __mptcp_push_pending(sk, msg->msg_flags); 1751 1752 out: 1753 release_sock(sk); 1754 return copied ? : ret; 1755 } 1756 1757 static void mptcp_wait_data(struct sock *sk, long *timeo) 1758 { 1759 DEFINE_WAIT_FUNC(wait, woken_wake_function); 1760 struct mptcp_sock *msk = mptcp_sk(sk); 1761 1762 add_wait_queue(sk_sleep(sk), &wait); 1763 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 1764 1765 sk_wait_event(sk, timeo, 1766 test_bit(MPTCP_DATA_READY, &msk->flags), &wait); 1767 1768 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 1769 remove_wait_queue(sk_sleep(sk), &wait); 1770 } 1771 1772 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk, 1773 struct msghdr *msg, 1774 size_t len, int flags, 1775 struct scm_timestamping_internal *tss, 1776 int *cmsg_flags) 1777 { 1778 struct sk_buff *skb, *tmp; 1779 int copied = 0; 1780 1781 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) { 1782 u32 offset = MPTCP_SKB_CB(skb)->offset; 1783 u32 data_len = skb->len - offset; 1784 u32 count = min_t(size_t, len - copied, data_len); 1785 int err; 1786 1787 if (!(flags & MSG_TRUNC)) { 1788 err = skb_copy_datagram_msg(skb, offset, msg, count); 1789 if (unlikely(err < 0)) { 1790 if (!copied) 1791 return err; 1792 break; 1793 } 1794 } 1795 1796 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 1797 tcp_update_recv_tstamps(skb, tss); 1798 *cmsg_flags |= MPTCP_CMSG_TS; 1799 } 1800 1801 copied += count; 1802 1803 if (count < data_len) { 1804 if (!(flags & MSG_PEEK)) 1805 MPTCP_SKB_CB(skb)->offset += count; 1806 break; 1807 } 1808 1809 if (!(flags & MSG_PEEK)) { 1810 /* we will bulk release the skb memory later */ 1811 skb->destructor = NULL; 1812 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize); 1813 __skb_unlink(skb, &msk->receive_queue); 1814 __kfree_skb(skb); 1815 } 1816 1817 if (copied >= len) 1818 break; 1819 } 1820 1821 return copied; 1822 } 1823 1824 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1825 * 1826 * Only difference: Use highest rtt estimate of the subflows in use. 1827 */ 1828 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1829 { 1830 struct mptcp_subflow_context *subflow; 1831 struct sock *sk = (struct sock *)msk; 1832 u32 time, advmss = 1; 1833 u64 rtt_us, mstamp; 1834 1835 sock_owned_by_me(sk); 1836 1837 if (copied <= 0) 1838 return; 1839 1840 msk->rcvq_space.copied += copied; 1841 1842 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1843 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1844 1845 rtt_us = msk->rcvq_space.rtt_us; 1846 if (rtt_us && time < (rtt_us >> 3)) 1847 return; 1848 1849 rtt_us = 0; 1850 mptcp_for_each_subflow(msk, subflow) { 1851 const struct tcp_sock *tp; 1852 u64 sf_rtt_us; 1853 u32 sf_advmss; 1854 1855 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1856 1857 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1858 sf_advmss = READ_ONCE(tp->advmss); 1859 1860 rtt_us = max(sf_rtt_us, rtt_us); 1861 advmss = max(sf_advmss, advmss); 1862 } 1863 1864 msk->rcvq_space.rtt_us = rtt_us; 1865 if (time < (rtt_us >> 3) || rtt_us == 0) 1866 return; 1867 1868 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 1869 goto new_measure; 1870 1871 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && 1872 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 1873 int rcvmem, rcvbuf; 1874 u64 rcvwin, grow; 1875 1876 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 1877 1878 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 1879 1880 do_div(grow, msk->rcvq_space.space); 1881 rcvwin += (grow << 1); 1882 1883 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER); 1884 while (tcp_win_from_space(sk, rcvmem) < advmss) 1885 rcvmem += 128; 1886 1887 do_div(rcvwin, advmss); 1888 rcvbuf = min_t(u64, rcvwin * rcvmem, 1889 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 1890 1891 if (rcvbuf > sk->sk_rcvbuf) { 1892 u32 window_clamp; 1893 1894 window_clamp = tcp_win_from_space(sk, rcvbuf); 1895 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 1896 1897 /* Make subflows follow along. If we do not do this, we 1898 * get drops at subflow level if skbs can't be moved to 1899 * the mptcp rx queue fast enough (announced rcv_win can 1900 * exceed ssk->sk_rcvbuf). 1901 */ 1902 mptcp_for_each_subflow(msk, subflow) { 1903 struct sock *ssk; 1904 bool slow; 1905 1906 ssk = mptcp_subflow_tcp_sock(subflow); 1907 slow = lock_sock_fast(ssk); 1908 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 1909 tcp_sk(ssk)->window_clamp = window_clamp; 1910 tcp_cleanup_rbuf(ssk, 1); 1911 unlock_sock_fast(ssk, slow); 1912 } 1913 } 1914 } 1915 1916 msk->rcvq_space.space = msk->rcvq_space.copied; 1917 new_measure: 1918 msk->rcvq_space.copied = 0; 1919 msk->rcvq_space.time = mstamp; 1920 } 1921 1922 static void __mptcp_update_rmem(struct sock *sk) 1923 { 1924 struct mptcp_sock *msk = mptcp_sk(sk); 1925 1926 if (!msk->rmem_released) 1927 return; 1928 1929 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc); 1930 sk_mem_uncharge(sk, msk->rmem_released); 1931 WRITE_ONCE(msk->rmem_released, 0); 1932 } 1933 1934 static void __mptcp_splice_receive_queue(struct sock *sk) 1935 { 1936 struct mptcp_sock *msk = mptcp_sk(sk); 1937 1938 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue); 1939 } 1940 1941 static bool __mptcp_move_skbs(struct mptcp_sock *msk) 1942 { 1943 struct sock *sk = (struct sock *)msk; 1944 unsigned int moved = 0; 1945 bool ret, done; 1946 1947 mptcp_flush_join_list(msk); 1948 do { 1949 struct sock *ssk = mptcp_subflow_recv_lookup(msk); 1950 bool slowpath; 1951 1952 /* we can have data pending in the subflows only if the msk 1953 * receive buffer was full at subflow_data_ready() time, 1954 * that is an unlikely slow path. 1955 */ 1956 if (likely(!ssk)) 1957 break; 1958 1959 slowpath = lock_sock_fast(ssk); 1960 mptcp_data_lock(sk); 1961 __mptcp_update_rmem(sk); 1962 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 1963 mptcp_data_unlock(sk); 1964 1965 if (unlikely(ssk->sk_err)) 1966 __mptcp_error_report(sk); 1967 unlock_sock_fast(ssk, slowpath); 1968 } while (!done); 1969 1970 /* acquire the data lock only if some input data is pending */ 1971 ret = moved > 0; 1972 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) || 1973 !skb_queue_empty_lockless(&sk->sk_receive_queue)) { 1974 mptcp_data_lock(sk); 1975 __mptcp_update_rmem(sk); 1976 ret |= __mptcp_ofo_queue(msk); 1977 __mptcp_splice_receive_queue(sk); 1978 mptcp_data_unlock(sk); 1979 } 1980 if (ret) 1981 mptcp_check_data_fin((struct sock *)msk); 1982 return !skb_queue_empty(&msk->receive_queue); 1983 } 1984 1985 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 1986 int nonblock, int flags, int *addr_len) 1987 { 1988 struct mptcp_sock *msk = mptcp_sk(sk); 1989 struct scm_timestamping_internal tss; 1990 int copied = 0, cmsg_flags = 0; 1991 int target; 1992 long timeo; 1993 1994 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 1995 if (unlikely(flags & MSG_ERRQUEUE)) 1996 return inet_recv_error(sk, msg, len, addr_len); 1997 1998 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk)); 1999 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2000 copied = -ENOTCONN; 2001 goto out_err; 2002 } 2003 2004 timeo = sock_rcvtimeo(sk, nonblock); 2005 2006 len = min_t(size_t, len, INT_MAX); 2007 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2008 2009 while (copied < len) { 2010 int bytes_read; 2011 2012 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags); 2013 if (unlikely(bytes_read < 0)) { 2014 if (!copied) 2015 copied = bytes_read; 2016 goto out_err; 2017 } 2018 2019 copied += bytes_read; 2020 2021 /* be sure to advertise window change */ 2022 mptcp_cleanup_rbuf(msk); 2023 2024 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk)) 2025 continue; 2026 2027 /* only the master socket status is relevant here. The exit 2028 * conditions mirror closely tcp_recvmsg() 2029 */ 2030 if (copied >= target) 2031 break; 2032 2033 if (copied) { 2034 if (sk->sk_err || 2035 sk->sk_state == TCP_CLOSE || 2036 (sk->sk_shutdown & RCV_SHUTDOWN) || 2037 !timeo || 2038 signal_pending(current)) 2039 break; 2040 } else { 2041 if (sk->sk_err) { 2042 copied = sock_error(sk); 2043 break; 2044 } 2045 2046 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags)) 2047 mptcp_check_for_eof(msk); 2048 2049 if (sk->sk_shutdown & RCV_SHUTDOWN) { 2050 /* race breaker: the shutdown could be after the 2051 * previous receive queue check 2052 */ 2053 if (__mptcp_move_skbs(msk)) 2054 continue; 2055 break; 2056 } 2057 2058 if (sk->sk_state == TCP_CLOSE) { 2059 copied = -ENOTCONN; 2060 break; 2061 } 2062 2063 if (!timeo) { 2064 copied = -EAGAIN; 2065 break; 2066 } 2067 2068 if (signal_pending(current)) { 2069 copied = sock_intr_errno(timeo); 2070 break; 2071 } 2072 } 2073 2074 pr_debug("block timeout %ld", timeo); 2075 mptcp_wait_data(sk, &timeo); 2076 } 2077 2078 if (skb_queue_empty_lockless(&sk->sk_receive_queue) && 2079 skb_queue_empty(&msk->receive_queue)) { 2080 /* entire backlog drained, clear DATA_READY. */ 2081 clear_bit(MPTCP_DATA_READY, &msk->flags); 2082 2083 /* .. race-breaker: ssk might have gotten new data 2084 * after last __mptcp_move_skbs() returned false. 2085 */ 2086 if (unlikely(__mptcp_move_skbs(msk))) 2087 set_bit(MPTCP_DATA_READY, &msk->flags); 2088 } 2089 2090 out_err: 2091 if (cmsg_flags && copied >= 0) { 2092 if (cmsg_flags & MPTCP_CMSG_TS) 2093 tcp_recv_timestamp(msg, sk, &tss); 2094 } 2095 2096 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d", 2097 msk, test_bit(MPTCP_DATA_READY, &msk->flags), 2098 skb_queue_empty_lockless(&sk->sk_receive_queue), copied); 2099 if (!(flags & MSG_PEEK)) 2100 mptcp_rcv_space_adjust(msk, copied); 2101 2102 release_sock(sk); 2103 return copied; 2104 } 2105 2106 static void mptcp_retransmit_timer(struct timer_list *t) 2107 { 2108 struct inet_connection_sock *icsk = from_timer(icsk, t, 2109 icsk_retransmit_timer); 2110 struct sock *sk = &icsk->icsk_inet.sk; 2111 struct mptcp_sock *msk = mptcp_sk(sk); 2112 2113 bh_lock_sock(sk); 2114 if (!sock_owned_by_user(sk)) { 2115 /* we need a process context to retransmit */ 2116 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2117 mptcp_schedule_work(sk); 2118 } else { 2119 /* delegate our work to tcp_release_cb() */ 2120 set_bit(MPTCP_RETRANSMIT, &msk->flags); 2121 } 2122 bh_unlock_sock(sk); 2123 sock_put(sk); 2124 } 2125 2126 static void mptcp_timeout_timer(struct timer_list *t) 2127 { 2128 struct sock *sk = from_timer(sk, t, sk_timer); 2129 2130 mptcp_schedule_work(sk); 2131 sock_put(sk); 2132 } 2133 2134 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2135 * level. 2136 * 2137 * A backup subflow is returned only if that is the only kind available. 2138 */ 2139 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2140 { 2141 struct sock *backup = NULL, *pick = NULL; 2142 struct mptcp_subflow_context *subflow; 2143 int min_stale_count = INT_MAX; 2144 2145 sock_owned_by_me((const struct sock *)msk); 2146 2147 if (__mptcp_check_fallback(msk)) 2148 return NULL; 2149 2150 mptcp_for_each_subflow(msk, subflow) { 2151 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2152 2153 if (!__mptcp_subflow_active(subflow)) 2154 continue; 2155 2156 /* still data outstanding at TCP level? skip this */ 2157 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2158 mptcp_pm_subflow_chk_stale(msk, ssk); 2159 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2160 continue; 2161 } 2162 2163 if (subflow->backup) { 2164 if (!backup) 2165 backup = ssk; 2166 continue; 2167 } 2168 2169 if (!pick) 2170 pick = ssk; 2171 } 2172 2173 if (pick) 2174 return pick; 2175 2176 /* use backup only if there are no progresses anywhere */ 2177 return min_stale_count > 1 ? backup : NULL; 2178 } 2179 2180 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk) 2181 { 2182 if (msk->subflow) { 2183 iput(SOCK_INODE(msk->subflow)); 2184 msk->subflow = NULL; 2185 } 2186 } 2187 2188 bool __mptcp_retransmit_pending_data(struct sock *sk) 2189 { 2190 struct mptcp_data_frag *cur, *rtx_head; 2191 struct mptcp_sock *msk = mptcp_sk(sk); 2192 2193 if (__mptcp_check_fallback(mptcp_sk(sk))) 2194 return false; 2195 2196 if (tcp_rtx_and_write_queues_empty(sk)) 2197 return false; 2198 2199 /* the closing socket has some data untransmitted and/or unacked: 2200 * some data in the mptcp rtx queue has not really xmitted yet. 2201 * keep it simple and re-inject the whole mptcp level rtx queue 2202 */ 2203 mptcp_data_lock(sk); 2204 __mptcp_clean_una_wakeup(sk); 2205 rtx_head = mptcp_rtx_head(sk); 2206 if (!rtx_head) { 2207 mptcp_data_unlock(sk); 2208 return false; 2209 } 2210 2211 /* will accept ack for reijected data before re-sending them */ 2212 if (!msk->recovery || after64(msk->snd_nxt, msk->recovery_snd_nxt)) 2213 msk->recovery_snd_nxt = msk->snd_nxt; 2214 msk->recovery = true; 2215 mptcp_data_unlock(sk); 2216 2217 msk->first_pending = rtx_head; 2218 msk->tx_pending_data += msk->snd_nxt - rtx_head->data_seq; 2219 msk->snd_nxt = rtx_head->data_seq; 2220 msk->snd_burst = 0; 2221 2222 /* be sure to clear the "sent status" on all re-injected fragments */ 2223 list_for_each_entry(cur, &msk->rtx_queue, list) { 2224 if (!cur->already_sent) 2225 break; 2226 cur->already_sent = 0; 2227 } 2228 2229 return true; 2230 } 2231 2232 /* subflow sockets can be either outgoing (connect) or incoming 2233 * (accept). 2234 * 2235 * Outgoing subflows use in-kernel sockets. 2236 * Incoming subflows do not have their own 'struct socket' allocated, 2237 * so we need to use tcp_close() after detaching them from the mptcp 2238 * parent socket. 2239 */ 2240 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2241 struct mptcp_subflow_context *subflow) 2242 { 2243 struct mptcp_sock *msk = mptcp_sk(sk); 2244 bool need_push; 2245 2246 list_del(&subflow->node); 2247 2248 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2249 2250 /* if we are invoked by the msk cleanup code, the subflow is 2251 * already orphaned 2252 */ 2253 if (ssk->sk_socket) 2254 sock_orphan(ssk); 2255 2256 need_push = __mptcp_retransmit_pending_data(sk); 2257 subflow->disposable = 1; 2258 2259 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2260 * the ssk has been already destroyed, we just need to release the 2261 * reference owned by msk; 2262 */ 2263 if (!inet_csk(ssk)->icsk_ulp_ops) { 2264 kfree_rcu(subflow, rcu); 2265 } else { 2266 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2267 __tcp_close(ssk, 0); 2268 2269 /* close acquired an extra ref */ 2270 __sock_put(ssk); 2271 } 2272 release_sock(ssk); 2273 2274 sock_put(ssk); 2275 2276 if (ssk == msk->last_snd) 2277 msk->last_snd = NULL; 2278 2279 if (ssk == msk->first) 2280 msk->first = NULL; 2281 2282 if (msk->subflow && ssk == msk->subflow->sk) 2283 mptcp_dispose_initial_subflow(msk); 2284 2285 if (need_push) 2286 __mptcp_push_pending(sk, 0); 2287 } 2288 2289 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2290 struct mptcp_subflow_context *subflow) 2291 { 2292 if (sk->sk_state == TCP_ESTABLISHED) 2293 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2294 __mptcp_close_ssk(sk, ssk, subflow); 2295 } 2296 2297 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2298 { 2299 return 0; 2300 } 2301 2302 static void __mptcp_close_subflow(struct mptcp_sock *msk) 2303 { 2304 struct mptcp_subflow_context *subflow, *tmp; 2305 2306 might_sleep(); 2307 2308 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) { 2309 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2310 2311 if (inet_sk_state_load(ssk) != TCP_CLOSE) 2312 continue; 2313 2314 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2315 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2316 continue; 2317 2318 mptcp_close_ssk((struct sock *)msk, ssk, subflow); 2319 } 2320 } 2321 2322 static bool mptcp_check_close_timeout(const struct sock *sk) 2323 { 2324 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp; 2325 struct mptcp_subflow_context *subflow; 2326 2327 if (delta >= TCP_TIMEWAIT_LEN) 2328 return true; 2329 2330 /* if all subflows are in closed status don't bother with additional 2331 * timeout 2332 */ 2333 mptcp_for_each_subflow(mptcp_sk(sk), subflow) { 2334 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) != 2335 TCP_CLOSE) 2336 return false; 2337 } 2338 return true; 2339 } 2340 2341 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2342 { 2343 struct mptcp_subflow_context *subflow, *tmp; 2344 struct sock *sk = &msk->sk.icsk_inet.sk; 2345 2346 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2347 return; 2348 2349 mptcp_token_destroy(msk); 2350 2351 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) { 2352 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2353 bool slow; 2354 2355 slow = lock_sock_fast(tcp_sk); 2356 if (tcp_sk->sk_state != TCP_CLOSE) { 2357 tcp_send_active_reset(tcp_sk, GFP_ATOMIC); 2358 tcp_set_state(tcp_sk, TCP_CLOSE); 2359 } 2360 unlock_sock_fast(tcp_sk, slow); 2361 } 2362 2363 inet_sk_state_store(sk, TCP_CLOSE); 2364 sk->sk_shutdown = SHUTDOWN_MASK; 2365 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2366 set_bit(MPTCP_DATA_READY, &msk->flags); 2367 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2368 2369 mptcp_close_wake_up(sk); 2370 } 2371 2372 static void __mptcp_retrans(struct sock *sk) 2373 { 2374 struct mptcp_sock *msk = mptcp_sk(sk); 2375 struct mptcp_sendmsg_info info = {}; 2376 struct mptcp_data_frag *dfrag; 2377 size_t copied = 0; 2378 struct sock *ssk; 2379 int ret; 2380 2381 mptcp_clean_una_wakeup(sk); 2382 dfrag = mptcp_rtx_head(sk); 2383 if (!dfrag) { 2384 if (mptcp_data_fin_enabled(msk)) { 2385 struct inet_connection_sock *icsk = inet_csk(sk); 2386 2387 icsk->icsk_retransmits++; 2388 mptcp_set_datafin_timeout(sk); 2389 mptcp_send_ack(msk); 2390 2391 goto reset_timer; 2392 } 2393 2394 return; 2395 } 2396 2397 ssk = mptcp_subflow_get_retrans(msk); 2398 if (!ssk) 2399 goto reset_timer; 2400 2401 lock_sock(ssk); 2402 2403 /* limit retransmission to the bytes already sent on some subflows */ 2404 info.sent = 0; 2405 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent; 2406 while (info.sent < info.limit) { 2407 if (!mptcp_alloc_tx_skb(sk, ssk)) 2408 break; 2409 2410 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2411 if (ret <= 0) 2412 break; 2413 2414 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2415 copied += ret; 2416 info.sent += ret; 2417 } 2418 if (copied) { 2419 dfrag->already_sent = max(dfrag->already_sent, info.sent); 2420 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2421 info.size_goal); 2422 } 2423 2424 release_sock(ssk); 2425 2426 reset_timer: 2427 if (!mptcp_timer_pending(sk)) 2428 mptcp_reset_timer(sk); 2429 } 2430 2431 static void mptcp_worker(struct work_struct *work) 2432 { 2433 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2434 struct sock *sk = &msk->sk.icsk_inet.sk; 2435 int state; 2436 2437 lock_sock(sk); 2438 state = sk->sk_state; 2439 if (unlikely(state == TCP_CLOSE)) 2440 goto unlock; 2441 2442 mptcp_check_data_fin_ack(sk); 2443 mptcp_flush_join_list(msk); 2444 2445 mptcp_check_fastclose(msk); 2446 2447 if (msk->pm.status) 2448 mptcp_pm_nl_work(msk); 2449 2450 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags)) 2451 mptcp_check_for_eof(msk); 2452 2453 __mptcp_check_send_data_fin(sk); 2454 mptcp_check_data_fin(sk); 2455 2456 /* There is no point in keeping around an orphaned sk timedout or 2457 * closed, but we need the msk around to reply to incoming DATA_FIN, 2458 * even if it is orphaned and in FIN_WAIT2 state 2459 */ 2460 if (sock_flag(sk, SOCK_DEAD) && 2461 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) { 2462 inet_sk_state_store(sk, TCP_CLOSE); 2463 __mptcp_destroy_sock(sk); 2464 goto unlock; 2465 } 2466 2467 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2468 __mptcp_close_subflow(msk); 2469 2470 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2471 __mptcp_retrans(sk); 2472 2473 unlock: 2474 release_sock(sk); 2475 sock_put(sk); 2476 } 2477 2478 static int __mptcp_init_sock(struct sock *sk) 2479 { 2480 struct mptcp_sock *msk = mptcp_sk(sk); 2481 2482 spin_lock_init(&msk->join_list_lock); 2483 2484 INIT_LIST_HEAD(&msk->conn_list); 2485 INIT_LIST_HEAD(&msk->join_list); 2486 INIT_LIST_HEAD(&msk->rtx_queue); 2487 INIT_WORK(&msk->work, mptcp_worker); 2488 __skb_queue_head_init(&msk->receive_queue); 2489 msk->out_of_order_queue = RB_ROOT; 2490 msk->first_pending = NULL; 2491 msk->wmem_reserved = 0; 2492 WRITE_ONCE(msk->rmem_released, 0); 2493 msk->tx_pending_data = 0; 2494 msk->timer_ival = TCP_RTO_MIN; 2495 2496 msk->first = NULL; 2497 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2498 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2499 msk->recovery = false; 2500 2501 mptcp_pm_data_init(msk); 2502 2503 /* re-use the csk retrans timer for MPTCP-level retrans */ 2504 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2505 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0); 2506 2507 return 0; 2508 } 2509 2510 static int mptcp_init_sock(struct sock *sk) 2511 { 2512 struct inet_connection_sock *icsk = inet_csk(sk); 2513 struct net *net = sock_net(sk); 2514 int ret; 2515 2516 ret = __mptcp_init_sock(sk); 2517 if (ret) 2518 return ret; 2519 2520 if (!mptcp_is_enabled(net)) 2521 return -ENOPROTOOPT; 2522 2523 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2524 return -ENOMEM; 2525 2526 ret = __mptcp_socket_create(mptcp_sk(sk)); 2527 if (ret) 2528 return ret; 2529 2530 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2531 * propagate the correct value 2532 */ 2533 tcp_assign_congestion_control(sk); 2534 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name); 2535 2536 /* no need to keep a reference to the ops, the name will suffice */ 2537 tcp_cleanup_congestion_control(sk); 2538 icsk->icsk_ca_ops = NULL; 2539 2540 sk_sockets_allocated_inc(sk); 2541 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1]; 2542 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1]; 2543 2544 return 0; 2545 } 2546 2547 static void __mptcp_clear_xmit(struct sock *sk) 2548 { 2549 struct mptcp_sock *msk = mptcp_sk(sk); 2550 struct mptcp_data_frag *dtmp, *dfrag; 2551 2552 WRITE_ONCE(msk->first_pending, NULL); 2553 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2554 dfrag_clear(sk, dfrag); 2555 } 2556 2557 static void mptcp_cancel_work(struct sock *sk) 2558 { 2559 struct mptcp_sock *msk = mptcp_sk(sk); 2560 2561 if (cancel_work_sync(&msk->work)) 2562 __sock_put(sk); 2563 } 2564 2565 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2566 { 2567 lock_sock(ssk); 2568 2569 switch (ssk->sk_state) { 2570 case TCP_LISTEN: 2571 if (!(how & RCV_SHUTDOWN)) 2572 break; 2573 fallthrough; 2574 case TCP_SYN_SENT: 2575 tcp_disconnect(ssk, O_NONBLOCK); 2576 break; 2577 default: 2578 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2579 pr_debug("Fallback"); 2580 ssk->sk_shutdown |= how; 2581 tcp_shutdown(ssk, how); 2582 } else { 2583 pr_debug("Sending DATA_FIN on subflow %p", ssk); 2584 tcp_send_ack(ssk); 2585 if (!mptcp_timer_pending(sk)) 2586 mptcp_reset_timer(sk); 2587 } 2588 break; 2589 } 2590 2591 release_sock(ssk); 2592 } 2593 2594 static const unsigned char new_state[16] = { 2595 /* current state: new state: action: */ 2596 [0 /* (Invalid) */] = TCP_CLOSE, 2597 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2598 [TCP_SYN_SENT] = TCP_CLOSE, 2599 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2600 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2601 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2602 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2603 [TCP_CLOSE] = TCP_CLOSE, 2604 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2605 [TCP_LAST_ACK] = TCP_LAST_ACK, 2606 [TCP_LISTEN] = TCP_CLOSE, 2607 [TCP_CLOSING] = TCP_CLOSING, 2608 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2609 }; 2610 2611 static int mptcp_close_state(struct sock *sk) 2612 { 2613 int next = (int)new_state[sk->sk_state]; 2614 int ns = next & TCP_STATE_MASK; 2615 2616 inet_sk_state_store(sk, ns); 2617 2618 return next & TCP_ACTION_FIN; 2619 } 2620 2621 static void __mptcp_check_send_data_fin(struct sock *sk) 2622 { 2623 struct mptcp_subflow_context *subflow; 2624 struct mptcp_sock *msk = mptcp_sk(sk); 2625 2626 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu", 2627 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2628 msk->snd_nxt, msk->write_seq); 2629 2630 /* we still need to enqueue subflows or not really shutting down, 2631 * skip this 2632 */ 2633 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2634 mptcp_send_head(sk)) 2635 return; 2636 2637 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2638 2639 /* fallback socket will not get data_fin/ack, can move to the next 2640 * state now 2641 */ 2642 if (__mptcp_check_fallback(msk)) { 2643 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) { 2644 inet_sk_state_store(sk, TCP_CLOSE); 2645 mptcp_close_wake_up(sk); 2646 } else if (sk->sk_state == TCP_FIN_WAIT1) { 2647 inet_sk_state_store(sk, TCP_FIN_WAIT2); 2648 } 2649 } 2650 2651 mptcp_flush_join_list(msk); 2652 mptcp_for_each_subflow(msk, subflow) { 2653 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2654 2655 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2656 } 2657 } 2658 2659 static void __mptcp_wr_shutdown(struct sock *sk) 2660 { 2661 struct mptcp_sock *msk = mptcp_sk(sk); 2662 2663 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d", 2664 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2665 !!mptcp_send_head(sk)); 2666 2667 /* will be ignored by fallback sockets */ 2668 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2669 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2670 2671 __mptcp_check_send_data_fin(sk); 2672 } 2673 2674 static void __mptcp_destroy_sock(struct sock *sk) 2675 { 2676 struct mptcp_subflow_context *subflow, *tmp; 2677 struct mptcp_sock *msk = mptcp_sk(sk); 2678 LIST_HEAD(conn_list); 2679 2680 pr_debug("msk=%p", msk); 2681 2682 might_sleep(); 2683 2684 /* be sure to always acquire the join list lock, to sync vs 2685 * mptcp_finish_join(). 2686 */ 2687 spin_lock_bh(&msk->join_list_lock); 2688 list_splice_tail_init(&msk->join_list, &msk->conn_list); 2689 spin_unlock_bh(&msk->join_list_lock); 2690 list_splice_init(&msk->conn_list, &conn_list); 2691 2692 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer); 2693 sk_stop_timer(sk, &sk->sk_timer); 2694 msk->pm.status = 0; 2695 2696 list_for_each_entry_safe(subflow, tmp, &conn_list, node) { 2697 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2698 __mptcp_close_ssk(sk, ssk, subflow); 2699 } 2700 2701 sk->sk_prot->destroy(sk); 2702 2703 WARN_ON_ONCE(msk->wmem_reserved); 2704 WARN_ON_ONCE(msk->rmem_released); 2705 sk_stream_kill_queues(sk); 2706 xfrm_sk_free_policy(sk); 2707 2708 sk_refcnt_debug_release(sk); 2709 mptcp_dispose_initial_subflow(msk); 2710 sock_put(sk); 2711 } 2712 2713 static void mptcp_close(struct sock *sk, long timeout) 2714 { 2715 struct mptcp_subflow_context *subflow; 2716 bool do_cancel_work = false; 2717 2718 lock_sock(sk); 2719 sk->sk_shutdown = SHUTDOWN_MASK; 2720 2721 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 2722 inet_sk_state_store(sk, TCP_CLOSE); 2723 goto cleanup; 2724 } 2725 2726 if (mptcp_close_state(sk)) 2727 __mptcp_wr_shutdown(sk); 2728 2729 sk_stream_wait_close(sk, timeout); 2730 2731 cleanup: 2732 /* orphan all the subflows */ 2733 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32; 2734 mptcp_for_each_subflow(mptcp_sk(sk), subflow) { 2735 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2736 bool slow = lock_sock_fast(ssk); 2737 2738 sock_orphan(ssk); 2739 unlock_sock_fast(ssk, slow); 2740 } 2741 sock_orphan(sk); 2742 2743 sock_hold(sk); 2744 pr_debug("msk=%p state=%d", sk, sk->sk_state); 2745 if (sk->sk_state == TCP_CLOSE) { 2746 __mptcp_destroy_sock(sk); 2747 do_cancel_work = true; 2748 } else { 2749 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN); 2750 } 2751 release_sock(sk); 2752 if (do_cancel_work) 2753 mptcp_cancel_work(sk); 2754 2755 if (mptcp_sk(sk)->token) 2756 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL); 2757 2758 sock_put(sk); 2759 } 2760 2761 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 2762 { 2763 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2764 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 2765 struct ipv6_pinfo *msk6 = inet6_sk(msk); 2766 2767 msk->sk_v6_daddr = ssk->sk_v6_daddr; 2768 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 2769 2770 if (msk6 && ssk6) { 2771 msk6->saddr = ssk6->saddr; 2772 msk6->flow_label = ssk6->flow_label; 2773 } 2774 #endif 2775 2776 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 2777 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 2778 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 2779 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 2780 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 2781 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 2782 } 2783 2784 static int mptcp_disconnect(struct sock *sk, int flags) 2785 { 2786 struct mptcp_subflow_context *subflow; 2787 struct mptcp_sock *msk = mptcp_sk(sk); 2788 2789 mptcp_do_flush_join_list(msk); 2790 2791 mptcp_for_each_subflow(msk, subflow) { 2792 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2793 2794 lock_sock(ssk); 2795 tcp_disconnect(ssk, flags); 2796 release_sock(ssk); 2797 } 2798 return 0; 2799 } 2800 2801 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2802 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 2803 { 2804 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo); 2805 2806 return (struct ipv6_pinfo *)(((u8 *)sk) + offset); 2807 } 2808 #endif 2809 2810 struct sock *mptcp_sk_clone(const struct sock *sk, 2811 const struct mptcp_options_received *mp_opt, 2812 struct request_sock *req) 2813 { 2814 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 2815 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 2816 struct mptcp_sock *msk; 2817 u64 ack_seq; 2818 2819 if (!nsk) 2820 return NULL; 2821 2822 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2823 if (nsk->sk_family == AF_INET6) 2824 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 2825 #endif 2826 2827 __mptcp_init_sock(nsk); 2828 2829 msk = mptcp_sk(nsk); 2830 msk->local_key = subflow_req->local_key; 2831 msk->token = subflow_req->token; 2832 msk->subflow = NULL; 2833 WRITE_ONCE(msk->fully_established, false); 2834 if (mp_opt->csum_reqd) 2835 WRITE_ONCE(msk->csum_enabled, true); 2836 2837 msk->write_seq = subflow_req->idsn + 1; 2838 msk->snd_nxt = msk->write_seq; 2839 msk->snd_una = msk->write_seq; 2840 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd; 2841 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 2842 2843 if (mp_opt->mp_capable) { 2844 msk->can_ack = true; 2845 msk->remote_key = mp_opt->sndr_key; 2846 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq); 2847 ack_seq++; 2848 WRITE_ONCE(msk->ack_seq, ack_seq); 2849 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq); 2850 } 2851 2852 sock_reset_flag(nsk, SOCK_RCU_FREE); 2853 /* will be fully established after successful MPC subflow creation */ 2854 inet_sk_state_store(nsk, TCP_SYN_RECV); 2855 2856 security_inet_csk_clone(nsk, req); 2857 bh_unlock_sock(nsk); 2858 2859 /* keep a single reference */ 2860 __sock_put(nsk); 2861 return nsk; 2862 } 2863 2864 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 2865 { 2866 const struct tcp_sock *tp = tcp_sk(ssk); 2867 2868 msk->rcvq_space.copied = 0; 2869 msk->rcvq_space.rtt_us = 0; 2870 2871 msk->rcvq_space.time = tp->tcp_mstamp; 2872 2873 /* initial rcv_space offering made to peer */ 2874 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 2875 TCP_INIT_CWND * tp->advmss); 2876 if (msk->rcvq_space.space == 0) 2877 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 2878 2879 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 2880 } 2881 2882 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err, 2883 bool kern) 2884 { 2885 struct mptcp_sock *msk = mptcp_sk(sk); 2886 struct socket *listener; 2887 struct sock *newsk; 2888 2889 listener = __mptcp_nmpc_socket(msk); 2890 if (WARN_ON_ONCE(!listener)) { 2891 *err = -EINVAL; 2892 return NULL; 2893 } 2894 2895 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk)); 2896 newsk = inet_csk_accept(listener->sk, flags, err, kern); 2897 if (!newsk) 2898 return NULL; 2899 2900 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk)); 2901 if (sk_is_mptcp(newsk)) { 2902 struct mptcp_subflow_context *subflow; 2903 struct sock *new_mptcp_sock; 2904 2905 subflow = mptcp_subflow_ctx(newsk); 2906 new_mptcp_sock = subflow->conn; 2907 2908 /* is_mptcp should be false if subflow->conn is missing, see 2909 * subflow_syn_recv_sock() 2910 */ 2911 if (WARN_ON_ONCE(!new_mptcp_sock)) { 2912 tcp_sk(newsk)->is_mptcp = 0; 2913 return newsk; 2914 } 2915 2916 /* acquire the 2nd reference for the owning socket */ 2917 sock_hold(new_mptcp_sock); 2918 newsk = new_mptcp_sock; 2919 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 2920 } else { 2921 MPTCP_INC_STATS(sock_net(sk), 2922 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK); 2923 } 2924 2925 return newsk; 2926 } 2927 2928 void mptcp_destroy_common(struct mptcp_sock *msk) 2929 { 2930 struct sock *sk = (struct sock *)msk; 2931 2932 __mptcp_clear_xmit(sk); 2933 2934 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */ 2935 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue); 2936 2937 skb_rbtree_purge(&msk->out_of_order_queue); 2938 mptcp_token_destroy(msk); 2939 mptcp_pm_free_anno_list(msk); 2940 } 2941 2942 static void mptcp_destroy(struct sock *sk) 2943 { 2944 struct mptcp_sock *msk = mptcp_sk(sk); 2945 2946 mptcp_destroy_common(msk); 2947 sk_sockets_allocated_dec(sk); 2948 } 2949 2950 void __mptcp_data_acked(struct sock *sk) 2951 { 2952 if (!sock_owned_by_user(sk)) 2953 __mptcp_clean_una(sk); 2954 else 2955 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags); 2956 2957 if (mptcp_pending_data_fin_ack(sk)) 2958 mptcp_schedule_work(sk); 2959 } 2960 2961 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 2962 { 2963 if (!mptcp_send_head(sk)) 2964 return; 2965 2966 if (!sock_owned_by_user(sk)) { 2967 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk)); 2968 2969 if (xmit_ssk == ssk) 2970 __mptcp_subflow_push_pending(sk, ssk); 2971 else if (xmit_ssk) 2972 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk)); 2973 } else { 2974 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags); 2975 } 2976 } 2977 2978 /* processes deferred events and flush wmem */ 2979 static void mptcp_release_cb(struct sock *sk) 2980 { 2981 for (;;) { 2982 unsigned long flags = 0; 2983 2984 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags)) 2985 flags |= BIT(MPTCP_PUSH_PENDING); 2986 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags)) 2987 flags |= BIT(MPTCP_RETRANSMIT); 2988 if (!flags) 2989 break; 2990 2991 /* the following actions acquire the subflow socket lock 2992 * 2993 * 1) can't be invoked in atomic scope 2994 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 2995 * datapath acquires the msk socket spinlock while helding 2996 * the subflow socket lock 2997 */ 2998 2999 spin_unlock_bh(&sk->sk_lock.slock); 3000 if (flags & BIT(MPTCP_PUSH_PENDING)) 3001 __mptcp_push_pending(sk, 0); 3002 if (flags & BIT(MPTCP_RETRANSMIT)) 3003 __mptcp_retrans(sk); 3004 3005 cond_resched(); 3006 spin_lock_bh(&sk->sk_lock.slock); 3007 } 3008 3009 /* be sure to set the current sk state before tacking actions 3010 * depending on sk_state 3011 */ 3012 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags)) 3013 __mptcp_set_connected(sk); 3014 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags)) 3015 __mptcp_clean_una_wakeup(sk); 3016 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags)) 3017 __mptcp_error_report(sk); 3018 3019 /* push_pending may touch wmem_reserved, ensure we do the cleanup 3020 * later 3021 */ 3022 __mptcp_update_wmem(sk); 3023 __mptcp_update_rmem(sk); 3024 } 3025 3026 void mptcp_subflow_process_delegated(struct sock *ssk) 3027 { 3028 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3029 struct sock *sk = subflow->conn; 3030 3031 mptcp_data_lock(sk); 3032 if (!sock_owned_by_user(sk)) 3033 __mptcp_subflow_push_pending(sk, ssk); 3034 else 3035 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags); 3036 mptcp_data_unlock(sk); 3037 mptcp_subflow_delegated_done(subflow); 3038 } 3039 3040 static int mptcp_hash(struct sock *sk) 3041 { 3042 /* should never be called, 3043 * we hash the TCP subflows not the master socket 3044 */ 3045 WARN_ON_ONCE(1); 3046 return 0; 3047 } 3048 3049 static void mptcp_unhash(struct sock *sk) 3050 { 3051 /* called from sk_common_release(), but nothing to do here */ 3052 } 3053 3054 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3055 { 3056 struct mptcp_sock *msk = mptcp_sk(sk); 3057 struct socket *ssock; 3058 3059 ssock = __mptcp_nmpc_socket(msk); 3060 pr_debug("msk=%p, subflow=%p", msk, ssock); 3061 if (WARN_ON_ONCE(!ssock)) 3062 return -EINVAL; 3063 3064 return inet_csk_get_port(ssock->sk, snum); 3065 } 3066 3067 void mptcp_finish_connect(struct sock *ssk) 3068 { 3069 struct mptcp_subflow_context *subflow; 3070 struct mptcp_sock *msk; 3071 struct sock *sk; 3072 u64 ack_seq; 3073 3074 subflow = mptcp_subflow_ctx(ssk); 3075 sk = subflow->conn; 3076 msk = mptcp_sk(sk); 3077 3078 pr_debug("msk=%p, token=%u", sk, subflow->token); 3079 3080 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq); 3081 ack_seq++; 3082 subflow->map_seq = ack_seq; 3083 subflow->map_subflow_seq = 1; 3084 3085 /* the socket is not connected yet, no msk/subflow ops can access/race 3086 * accessing the field below 3087 */ 3088 WRITE_ONCE(msk->remote_key, subflow->remote_key); 3089 WRITE_ONCE(msk->local_key, subflow->local_key); 3090 WRITE_ONCE(msk->write_seq, subflow->idsn + 1); 3091 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3092 WRITE_ONCE(msk->ack_seq, ack_seq); 3093 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq); 3094 WRITE_ONCE(msk->can_ack, 1); 3095 WRITE_ONCE(msk->snd_una, msk->write_seq); 3096 3097 mptcp_pm_new_connection(msk, ssk, 0); 3098 3099 mptcp_rcv_space_init(msk, ssk); 3100 } 3101 3102 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3103 { 3104 write_lock_bh(&sk->sk_callback_lock); 3105 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3106 sk_set_socket(sk, parent); 3107 sk->sk_uid = SOCK_INODE(parent)->i_uid; 3108 write_unlock_bh(&sk->sk_callback_lock); 3109 } 3110 3111 bool mptcp_finish_join(struct sock *ssk) 3112 { 3113 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3114 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3115 struct sock *parent = (void *)msk; 3116 struct socket *parent_sock; 3117 bool ret; 3118 3119 pr_debug("msk=%p, subflow=%p", msk, subflow); 3120 3121 /* mptcp socket already closing? */ 3122 if (!mptcp_is_fully_established(parent)) { 3123 subflow->reset_reason = MPTCP_RST_EMPTCP; 3124 return false; 3125 } 3126 3127 if (!msk->pm.server_side) 3128 goto out; 3129 3130 if (!mptcp_pm_allow_new_subflow(msk)) { 3131 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3132 return false; 3133 } 3134 3135 /* active connections are already on conn_list, and we can't acquire 3136 * msk lock here. 3137 * use the join list lock as synchronization point and double-check 3138 * msk status to avoid racing with __mptcp_destroy_sock() 3139 */ 3140 spin_lock_bh(&msk->join_list_lock); 3141 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED; 3142 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) { 3143 list_add_tail(&subflow->node, &msk->join_list); 3144 sock_hold(ssk); 3145 } 3146 spin_unlock_bh(&msk->join_list_lock); 3147 if (!ret) { 3148 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3149 return false; 3150 } 3151 3152 /* attach to msk socket only after we are sure he will deal with us 3153 * at close time 3154 */ 3155 parent_sock = READ_ONCE(parent->sk_socket); 3156 if (parent_sock && !ssk->sk_socket) 3157 mptcp_sock_graft(ssk, parent_sock); 3158 subflow->map_seq = READ_ONCE(msk->ack_seq); 3159 out: 3160 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 3161 return true; 3162 } 3163 3164 static void mptcp_shutdown(struct sock *sk, int how) 3165 { 3166 pr_debug("sk=%p, how=%d", sk, how); 3167 3168 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3169 __mptcp_wr_shutdown(sk); 3170 } 3171 3172 static struct proto mptcp_prot = { 3173 .name = "MPTCP", 3174 .owner = THIS_MODULE, 3175 .init = mptcp_init_sock, 3176 .disconnect = mptcp_disconnect, 3177 .close = mptcp_close, 3178 .accept = mptcp_accept, 3179 .setsockopt = mptcp_setsockopt, 3180 .getsockopt = mptcp_getsockopt, 3181 .shutdown = mptcp_shutdown, 3182 .destroy = mptcp_destroy, 3183 .sendmsg = mptcp_sendmsg, 3184 .recvmsg = mptcp_recvmsg, 3185 .release_cb = mptcp_release_cb, 3186 .hash = mptcp_hash, 3187 .unhash = mptcp_unhash, 3188 .get_port = mptcp_get_port, 3189 .sockets_allocated = &mptcp_sockets_allocated, 3190 .memory_allocated = &tcp_memory_allocated, 3191 .memory_pressure = &tcp_memory_pressure, 3192 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3193 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3194 .sysctl_mem = sysctl_tcp_mem, 3195 .obj_size = sizeof(struct mptcp_sock), 3196 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3197 .no_autobind = true, 3198 }; 3199 3200 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3201 { 3202 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3203 struct socket *ssock; 3204 int err; 3205 3206 lock_sock(sock->sk); 3207 ssock = __mptcp_nmpc_socket(msk); 3208 if (!ssock) { 3209 err = -EINVAL; 3210 goto unlock; 3211 } 3212 3213 err = ssock->ops->bind(ssock, uaddr, addr_len); 3214 if (!err) 3215 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3216 3217 unlock: 3218 release_sock(sock->sk); 3219 return err; 3220 } 3221 3222 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk, 3223 struct mptcp_subflow_context *subflow) 3224 { 3225 subflow->request_mptcp = 0; 3226 __mptcp_do_fallback(msk); 3227 } 3228 3229 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr, 3230 int addr_len, int flags) 3231 { 3232 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3233 struct mptcp_subflow_context *subflow; 3234 struct socket *ssock; 3235 int err; 3236 3237 lock_sock(sock->sk); 3238 if (sock->state != SS_UNCONNECTED && msk->subflow) { 3239 /* pending connection or invalid state, let existing subflow 3240 * cope with that 3241 */ 3242 ssock = msk->subflow; 3243 goto do_connect; 3244 } 3245 3246 ssock = __mptcp_nmpc_socket(msk); 3247 if (!ssock) { 3248 err = -EINVAL; 3249 goto unlock; 3250 } 3251 3252 mptcp_token_destroy(msk); 3253 inet_sk_state_store(sock->sk, TCP_SYN_SENT); 3254 subflow = mptcp_subflow_ctx(ssock->sk); 3255 #ifdef CONFIG_TCP_MD5SIG 3256 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3257 * TCP option space. 3258 */ 3259 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info)) 3260 mptcp_subflow_early_fallback(msk, subflow); 3261 #endif 3262 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) { 3263 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT); 3264 mptcp_subflow_early_fallback(msk, subflow); 3265 } 3266 if (likely(!__mptcp_check_fallback(msk))) 3267 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE); 3268 3269 do_connect: 3270 err = ssock->ops->connect(ssock, uaddr, addr_len, flags); 3271 sock->state = ssock->state; 3272 3273 /* on successful connect, the msk state will be moved to established by 3274 * subflow_finish_connect() 3275 */ 3276 if (!err || err == -EINPROGRESS) 3277 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3278 else 3279 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk)); 3280 3281 unlock: 3282 release_sock(sock->sk); 3283 return err; 3284 } 3285 3286 static int mptcp_listen(struct socket *sock, int backlog) 3287 { 3288 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3289 struct socket *ssock; 3290 int err; 3291 3292 pr_debug("msk=%p", msk); 3293 3294 lock_sock(sock->sk); 3295 ssock = __mptcp_nmpc_socket(msk); 3296 if (!ssock) { 3297 err = -EINVAL; 3298 goto unlock; 3299 } 3300 3301 mptcp_token_destroy(msk); 3302 inet_sk_state_store(sock->sk, TCP_LISTEN); 3303 sock_set_flag(sock->sk, SOCK_RCU_FREE); 3304 3305 err = ssock->ops->listen(ssock, backlog); 3306 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk)); 3307 if (!err) 3308 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3309 3310 unlock: 3311 release_sock(sock->sk); 3312 return err; 3313 } 3314 3315 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3316 int flags, bool kern) 3317 { 3318 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3319 struct socket *ssock; 3320 int err; 3321 3322 pr_debug("msk=%p", msk); 3323 3324 lock_sock(sock->sk); 3325 if (sock->sk->sk_state != TCP_LISTEN) 3326 goto unlock_fail; 3327 3328 ssock = __mptcp_nmpc_socket(msk); 3329 if (!ssock) 3330 goto unlock_fail; 3331 3332 clear_bit(MPTCP_DATA_READY, &msk->flags); 3333 sock_hold(ssock->sk); 3334 release_sock(sock->sk); 3335 3336 err = ssock->ops->accept(sock, newsock, flags, kern); 3337 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) { 3338 struct mptcp_sock *msk = mptcp_sk(newsock->sk); 3339 struct mptcp_subflow_context *subflow; 3340 struct sock *newsk = newsock->sk; 3341 3342 lock_sock(newsk); 3343 3344 /* PM/worker can now acquire the first subflow socket 3345 * lock without racing with listener queue cleanup, 3346 * we can notify it, if needed. 3347 * 3348 * Even if remote has reset the initial subflow by now 3349 * the refcnt is still at least one. 3350 */ 3351 subflow = mptcp_subflow_ctx(msk->first); 3352 list_add(&subflow->node, &msk->conn_list); 3353 sock_hold(msk->first); 3354 if (mptcp_is_fully_established(newsk)) 3355 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL); 3356 3357 mptcp_copy_inaddrs(newsk, msk->first); 3358 mptcp_rcv_space_init(msk, msk->first); 3359 mptcp_propagate_sndbuf(newsk, msk->first); 3360 3361 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3362 * This is needed so NOSPACE flag can be set from tcp stack. 3363 */ 3364 mptcp_flush_join_list(msk); 3365 mptcp_for_each_subflow(msk, subflow) { 3366 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3367 3368 if (!ssk->sk_socket) 3369 mptcp_sock_graft(ssk, newsock); 3370 } 3371 release_sock(newsk); 3372 } 3373 3374 if (inet_csk_listen_poll(ssock->sk)) 3375 set_bit(MPTCP_DATA_READY, &msk->flags); 3376 sock_put(ssock->sk); 3377 return err; 3378 3379 unlock_fail: 3380 release_sock(sock->sk); 3381 return -EINVAL; 3382 } 3383 3384 static __poll_t mptcp_check_readable(struct mptcp_sock *msk) 3385 { 3386 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 3387 0; 3388 } 3389 3390 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3391 { 3392 struct sock *sk = (struct sock *)msk; 3393 3394 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN)) 3395 return EPOLLOUT | EPOLLWRNORM; 3396 3397 if (sk_stream_is_writeable(sk)) 3398 return EPOLLOUT | EPOLLWRNORM; 3399 3400 mptcp_set_nospace(sk); 3401 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */ 3402 if (sk_stream_is_writeable(sk)) 3403 return EPOLLOUT | EPOLLWRNORM; 3404 3405 return 0; 3406 } 3407 3408 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3409 struct poll_table_struct *wait) 3410 { 3411 struct sock *sk = sock->sk; 3412 struct mptcp_sock *msk; 3413 __poll_t mask = 0; 3414 int state; 3415 3416 msk = mptcp_sk(sk); 3417 sock_poll_wait(file, sock, wait); 3418 3419 state = inet_sk_state_load(sk); 3420 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags); 3421 if (state == TCP_LISTEN) 3422 return mptcp_check_readable(msk); 3423 3424 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 3425 mask |= mptcp_check_readable(msk); 3426 mask |= mptcp_check_writeable(msk); 3427 } 3428 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 3429 mask |= EPOLLHUP; 3430 if (sk->sk_shutdown & RCV_SHUTDOWN) 3431 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 3432 3433 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 3434 smp_rmb(); 3435 if (sk->sk_err) 3436 mask |= EPOLLERR; 3437 3438 return mask; 3439 } 3440 3441 static const struct proto_ops mptcp_stream_ops = { 3442 .family = PF_INET, 3443 .owner = THIS_MODULE, 3444 .release = inet_release, 3445 .bind = mptcp_bind, 3446 .connect = mptcp_stream_connect, 3447 .socketpair = sock_no_socketpair, 3448 .accept = mptcp_stream_accept, 3449 .getname = inet_getname, 3450 .poll = mptcp_poll, 3451 .ioctl = inet_ioctl, 3452 .gettstamp = sock_gettstamp, 3453 .listen = mptcp_listen, 3454 .shutdown = inet_shutdown, 3455 .setsockopt = sock_common_setsockopt, 3456 .getsockopt = sock_common_getsockopt, 3457 .sendmsg = inet_sendmsg, 3458 .recvmsg = inet_recvmsg, 3459 .mmap = sock_no_mmap, 3460 .sendpage = inet_sendpage, 3461 }; 3462 3463 static struct inet_protosw mptcp_protosw = { 3464 .type = SOCK_STREAM, 3465 .protocol = IPPROTO_MPTCP, 3466 .prot = &mptcp_prot, 3467 .ops = &mptcp_stream_ops, 3468 .flags = INET_PROTOSW_ICSK, 3469 }; 3470 3471 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 3472 { 3473 struct mptcp_delegated_action *delegated; 3474 struct mptcp_subflow_context *subflow; 3475 int work_done = 0; 3476 3477 delegated = container_of(napi, struct mptcp_delegated_action, napi); 3478 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 3479 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3480 3481 bh_lock_sock_nested(ssk); 3482 if (!sock_owned_by_user(ssk) && 3483 mptcp_subflow_has_delegated_action(subflow)) 3484 mptcp_subflow_process_delegated(ssk); 3485 /* ... elsewhere tcp_release_cb_override already processed 3486 * the action or will do at next release_sock(). 3487 * In both case must dequeue the subflow here - on the same 3488 * CPU that scheduled it. 3489 */ 3490 bh_unlock_sock(ssk); 3491 sock_put(ssk); 3492 3493 if (++work_done == budget) 3494 return budget; 3495 } 3496 3497 /* always provide a 0 'work_done' argument, so that napi_complete_done 3498 * will not try accessing the NULL napi->dev ptr 3499 */ 3500 napi_complete_done(napi, 0); 3501 return work_done; 3502 } 3503 3504 void __init mptcp_proto_init(void) 3505 { 3506 struct mptcp_delegated_action *delegated; 3507 int cpu; 3508 3509 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 3510 3511 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 3512 panic("Failed to allocate MPTCP pcpu counter\n"); 3513 3514 init_dummy_netdev(&mptcp_napi_dev); 3515 for_each_possible_cpu(cpu) { 3516 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 3517 INIT_LIST_HEAD(&delegated->head); 3518 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll, 3519 NAPI_POLL_WEIGHT); 3520 napi_enable(&delegated->napi); 3521 } 3522 3523 mptcp_subflow_init(); 3524 mptcp_pm_init(); 3525 mptcp_token_init(); 3526 3527 if (proto_register(&mptcp_prot, 1) != 0) 3528 panic("Failed to register MPTCP proto.\n"); 3529 3530 inet_register_protosw(&mptcp_protosw); 3531 3532 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 3533 } 3534 3535 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3536 static const struct proto_ops mptcp_v6_stream_ops = { 3537 .family = PF_INET6, 3538 .owner = THIS_MODULE, 3539 .release = inet6_release, 3540 .bind = mptcp_bind, 3541 .connect = mptcp_stream_connect, 3542 .socketpair = sock_no_socketpair, 3543 .accept = mptcp_stream_accept, 3544 .getname = inet6_getname, 3545 .poll = mptcp_poll, 3546 .ioctl = inet6_ioctl, 3547 .gettstamp = sock_gettstamp, 3548 .listen = mptcp_listen, 3549 .shutdown = inet_shutdown, 3550 .setsockopt = sock_common_setsockopt, 3551 .getsockopt = sock_common_getsockopt, 3552 .sendmsg = inet6_sendmsg, 3553 .recvmsg = inet6_recvmsg, 3554 .mmap = sock_no_mmap, 3555 .sendpage = inet_sendpage, 3556 #ifdef CONFIG_COMPAT 3557 .compat_ioctl = inet6_compat_ioctl, 3558 #endif 3559 }; 3560 3561 static struct proto mptcp_v6_prot; 3562 3563 static void mptcp_v6_destroy(struct sock *sk) 3564 { 3565 mptcp_destroy(sk); 3566 inet6_destroy_sock(sk); 3567 } 3568 3569 static struct inet_protosw mptcp_v6_protosw = { 3570 .type = SOCK_STREAM, 3571 .protocol = IPPROTO_MPTCP, 3572 .prot = &mptcp_v6_prot, 3573 .ops = &mptcp_v6_stream_ops, 3574 .flags = INET_PROTOSW_ICSK, 3575 }; 3576 3577 int __init mptcp_proto_v6_init(void) 3578 { 3579 int err; 3580 3581 mptcp_v6_prot = mptcp_prot; 3582 strcpy(mptcp_v6_prot.name, "MPTCPv6"); 3583 mptcp_v6_prot.slab = NULL; 3584 mptcp_v6_prot.destroy = mptcp_v6_destroy; 3585 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 3586 3587 err = proto_register(&mptcp_v6_prot, 1); 3588 if (err) 3589 return err; 3590 3591 err = inet6_register_protosw(&mptcp_v6_protosw); 3592 if (err) 3593 proto_unregister(&mptcp_v6_prot); 3594 3595 return err; 3596 } 3597 #endif 3598