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