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