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