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 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 29 struct mptcp6_sock { 30 struct mptcp_sock msk; 31 struct ipv6_pinfo np; 32 }; 33 #endif 34 35 struct mptcp_skb_cb { 36 u64 map_seq; 37 u64 end_seq; 38 u32 offset; 39 }; 40 41 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0])) 42 43 static struct percpu_counter mptcp_sockets_allocated; 44 45 static void __mptcp_destroy_sock(struct sock *sk); 46 static void __mptcp_check_send_data_fin(struct sock *sk); 47 48 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions); 49 static struct net_device mptcp_napi_dev; 50 51 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not 52 * completed yet or has failed, return the subflow socket. 53 * Otherwise return NULL. 54 */ 55 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk) 56 { 57 if (!msk->subflow || READ_ONCE(msk->can_ack)) 58 return NULL; 59 60 return msk->subflow; 61 } 62 63 /* Returns end sequence number of the receiver's advertised window */ 64 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 65 { 66 return READ_ONCE(msk->wnd_end); 67 } 68 69 static bool mptcp_is_tcpsk(struct sock *sk) 70 { 71 struct socket *sock = sk->sk_socket; 72 73 if (unlikely(sk->sk_prot == &tcp_prot)) { 74 /* we are being invoked after mptcp_accept() has 75 * accepted a non-mp-capable flow: sk is a tcp_sk, 76 * not an mptcp one. 77 * 78 * Hand the socket over to tcp so all further socket ops 79 * bypass mptcp. 80 */ 81 sock->ops = &inet_stream_ops; 82 return true; 83 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 84 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) { 85 sock->ops = &inet6_stream_ops; 86 return true; 87 #endif 88 } 89 90 return false; 91 } 92 93 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk) 94 { 95 sock_owned_by_me((const struct sock *)msk); 96 97 if (likely(!__mptcp_check_fallback(msk))) 98 return NULL; 99 100 return msk->first; 101 } 102 103 static int __mptcp_socket_create(struct mptcp_sock *msk) 104 { 105 struct mptcp_subflow_context *subflow; 106 struct sock *sk = (struct sock *)msk; 107 struct socket *ssock; 108 int err; 109 110 err = mptcp_subflow_create_socket(sk, &ssock); 111 if (err) 112 return err; 113 114 msk->first = ssock->sk; 115 msk->subflow = ssock; 116 subflow = mptcp_subflow_ctx(ssock->sk); 117 list_add(&subflow->node, &msk->conn_list); 118 sock_hold(ssock->sk); 119 subflow->request_mptcp = 1; 120 mptcp_sock_graft(msk->first, sk->sk_socket); 121 122 return 0; 123 } 124 125 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 126 { 127 sk_drops_add(sk, skb); 128 __kfree_skb(skb); 129 } 130 131 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 132 struct sk_buff *from) 133 { 134 bool fragstolen; 135 int delta; 136 137 if (MPTCP_SKB_CB(from)->offset || 138 !skb_try_coalesce(to, from, &fragstolen, &delta)) 139 return false; 140 141 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx", 142 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 143 to->len, MPTCP_SKB_CB(from)->end_seq); 144 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 145 kfree_skb_partial(from, fragstolen); 146 atomic_add(delta, &sk->sk_rmem_alloc); 147 sk_mem_charge(sk, delta); 148 return true; 149 } 150 151 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 152 struct sk_buff *from) 153 { 154 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 155 return false; 156 157 return mptcp_try_coalesce((struct sock *)msk, to, from); 158 } 159 160 /* "inspired" by tcp_data_queue_ofo(), main differences: 161 * - use mptcp seqs 162 * - don't cope with sacks 163 */ 164 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 165 { 166 struct sock *sk = (struct sock *)msk; 167 struct rb_node **p, *parent; 168 u64 seq, end_seq, max_seq; 169 struct sk_buff *skb1; 170 171 seq = MPTCP_SKB_CB(skb)->map_seq; 172 end_seq = MPTCP_SKB_CB(skb)->end_seq; 173 max_seq = READ_ONCE(msk->rcv_wnd_sent); 174 175 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq, 176 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 177 if (after64(end_seq, max_seq)) { 178 /* out of window */ 179 mptcp_drop(sk, skb); 180 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 181 (unsigned long long)end_seq - (unsigned long)max_seq, 182 (unsigned long long)msk->rcv_wnd_sent); 183 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 184 return; 185 } 186 187 p = &msk->out_of_order_queue.rb_node; 188 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 189 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 190 rb_link_node(&skb->rbnode, NULL, p); 191 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 192 msk->ooo_last_skb = skb; 193 goto end; 194 } 195 196 /* with 2 subflows, adding at end of ooo queue is quite likely 197 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 198 */ 199 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 202 return; 203 } 204 205 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 206 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 207 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 208 parent = &msk->ooo_last_skb->rbnode; 209 p = &parent->rb_right; 210 goto insert; 211 } 212 213 /* Find place to insert this segment. Handle overlaps on the way. */ 214 parent = NULL; 215 while (*p) { 216 parent = *p; 217 skb1 = rb_to_skb(parent); 218 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 219 p = &parent->rb_left; 220 continue; 221 } 222 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 223 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 224 /* All the bits are present. Drop. */ 225 mptcp_drop(sk, skb); 226 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 227 return; 228 } 229 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 230 /* partial overlap: 231 * | skb | 232 * | skb1 | 233 * continue traversing 234 */ 235 } else { 236 /* skb's seq == skb1's seq and skb covers skb1. 237 * Replace skb1 with skb. 238 */ 239 rb_replace_node(&skb1->rbnode, &skb->rbnode, 240 &msk->out_of_order_queue); 241 mptcp_drop(sk, skb1); 242 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 243 goto merge_right; 244 } 245 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 247 return; 248 } 249 p = &parent->rb_right; 250 } 251 252 insert: 253 /* Insert segment into RB tree. */ 254 rb_link_node(&skb->rbnode, parent, p); 255 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 256 257 merge_right: 258 /* Remove other segments covered by skb. */ 259 while ((skb1 = skb_rb_next(skb)) != NULL) { 260 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 261 break; 262 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 263 mptcp_drop(sk, skb1); 264 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 265 } 266 /* If there is no skb after us, we are the last_skb ! */ 267 if (!skb1) 268 msk->ooo_last_skb = skb; 269 270 end: 271 skb_condense(skb); 272 skb_set_owner_r(skb, sk); 273 } 274 275 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 276 struct sk_buff *skb, unsigned int offset, 277 size_t copy_len) 278 { 279 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 280 struct sock *sk = (struct sock *)msk; 281 struct sk_buff *tail; 282 283 __skb_unlink(skb, &ssk->sk_receive_queue); 284 285 skb_ext_reset(skb); 286 skb_orphan(skb); 287 288 /* try to fetch required memory from subflow */ 289 if (!sk_rmem_schedule(sk, skb, skb->truesize)) { 290 if (ssk->sk_forward_alloc < skb->truesize) 291 goto drop; 292 __sk_mem_reclaim(ssk, skb->truesize); 293 if (!sk_rmem_schedule(sk, skb, skb->truesize)) 294 goto drop; 295 } 296 297 /* the skb map_seq accounts for the skb offset: 298 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 299 * value 300 */ 301 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 302 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 303 MPTCP_SKB_CB(skb)->offset = offset; 304 305 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 306 /* in sequence */ 307 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 308 tail = skb_peek_tail(&sk->sk_receive_queue); 309 if (tail && mptcp_try_coalesce(sk, tail, skb)) 310 return true; 311 312 skb_set_owner_r(skb, sk); 313 __skb_queue_tail(&sk->sk_receive_queue, skb); 314 return true; 315 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 316 mptcp_data_queue_ofo(msk, skb); 317 return false; 318 } 319 320 /* old data, keep it simple and drop the whole pkt, sender 321 * will retransmit as needed, if needed. 322 */ 323 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 324 drop: 325 mptcp_drop(sk, skb); 326 return false; 327 } 328 329 static void mptcp_stop_timer(struct sock *sk) 330 { 331 struct inet_connection_sock *icsk = inet_csk(sk); 332 333 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 334 mptcp_sk(sk)->timer_ival = 0; 335 } 336 337 static void mptcp_close_wake_up(struct sock *sk) 338 { 339 if (sock_flag(sk, SOCK_DEAD)) 340 return; 341 342 sk->sk_state_change(sk); 343 if (sk->sk_shutdown == SHUTDOWN_MASK || 344 sk->sk_state == TCP_CLOSE) 345 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 346 else 347 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 348 } 349 350 static bool mptcp_pending_data_fin_ack(struct sock *sk) 351 { 352 struct mptcp_sock *msk = mptcp_sk(sk); 353 354 return !__mptcp_check_fallback(msk) && 355 ((1 << sk->sk_state) & 356 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 357 msk->write_seq == READ_ONCE(msk->snd_una); 358 } 359 360 static void mptcp_check_data_fin_ack(struct sock *sk) 361 { 362 struct mptcp_sock *msk = mptcp_sk(sk); 363 364 /* Look for an acknowledged DATA_FIN */ 365 if (mptcp_pending_data_fin_ack(sk)) { 366 WRITE_ONCE(msk->snd_data_fin_enable, 0); 367 368 switch (sk->sk_state) { 369 case TCP_FIN_WAIT1: 370 inet_sk_state_store(sk, TCP_FIN_WAIT2); 371 break; 372 case TCP_CLOSING: 373 case TCP_LAST_ACK: 374 inet_sk_state_store(sk, TCP_CLOSE); 375 break; 376 } 377 378 mptcp_close_wake_up(sk); 379 } 380 } 381 382 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 383 { 384 struct mptcp_sock *msk = mptcp_sk(sk); 385 386 if (READ_ONCE(msk->rcv_data_fin) && 387 ((1 << sk->sk_state) & 388 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 389 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 390 391 if (msk->ack_seq == rcv_data_fin_seq) { 392 if (seq) 393 *seq = rcv_data_fin_seq; 394 395 return true; 396 } 397 } 398 399 return false; 400 } 401 402 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk) 403 { 404 long tout = ssk && inet_csk(ssk)->icsk_pending ? 405 inet_csk(ssk)->icsk_timeout - jiffies : 0; 406 407 if (tout <= 0) 408 tout = mptcp_sk(sk)->timer_ival; 409 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 410 } 411 412 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 413 { 414 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 415 416 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */ 417 if (subflow->request_join && !subflow->fully_established) 418 return false; 419 420 /* only send if our side has not closed yet */ 421 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)); 422 } 423 424 static bool tcp_can_send_ack(const struct sock *ssk) 425 { 426 return !((1 << inet_sk_state_load(ssk)) & 427 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 428 } 429 430 static void mptcp_send_ack(struct mptcp_sock *msk) 431 { 432 struct mptcp_subflow_context *subflow; 433 434 mptcp_for_each_subflow(msk, subflow) { 435 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 436 437 lock_sock(ssk); 438 if (tcp_can_send_ack(ssk)) 439 tcp_send_ack(ssk); 440 release_sock(ssk); 441 } 442 } 443 444 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk) 445 { 446 int ret; 447 448 lock_sock(ssk); 449 ret = tcp_can_send_ack(ssk); 450 if (ret) 451 tcp_cleanup_rbuf(ssk, 1); 452 release_sock(ssk); 453 return ret; 454 } 455 456 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk) 457 { 458 struct sock *ack_hint = READ_ONCE(msk->ack_hint); 459 int old_space = READ_ONCE(msk->old_wspace); 460 struct mptcp_subflow_context *subflow; 461 struct sock *sk = (struct sock *)msk; 462 bool cleanup; 463 464 /* this is a simple superset of what tcp_cleanup_rbuf() implements 465 * so that we don't have to acquire the ssk socket lock most of the time 466 * to do actually nothing 467 */ 468 cleanup = __mptcp_space(sk) - old_space >= max(0, old_space); 469 if (!cleanup) 470 return; 471 472 /* if the hinted ssk is still active, try to use it */ 473 if (likely(ack_hint)) { 474 mptcp_for_each_subflow(msk, subflow) { 475 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 476 477 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk)) 478 return; 479 } 480 } 481 482 /* otherwise pick the first active subflow */ 483 mptcp_for_each_subflow(msk, subflow) 484 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow))) 485 return; 486 } 487 488 static bool mptcp_check_data_fin(struct sock *sk) 489 { 490 struct mptcp_sock *msk = mptcp_sk(sk); 491 u64 rcv_data_fin_seq; 492 bool ret = false; 493 494 if (__mptcp_check_fallback(msk) || !msk->first) 495 return ret; 496 497 /* Need to ack a DATA_FIN received from a peer while this side 498 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 499 * msk->rcv_data_fin was set when parsing the incoming options 500 * at the subflow level and the msk lock was not held, so this 501 * is the first opportunity to act on the DATA_FIN and change 502 * the msk state. 503 * 504 * If we are caught up to the sequence number of the incoming 505 * DATA_FIN, send the DATA_ACK now and do state transition. If 506 * not caught up, do nothing and let the recv code send DATA_ACK 507 * when catching up. 508 */ 509 510 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 511 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 512 WRITE_ONCE(msk->rcv_data_fin, 0); 513 514 sk->sk_shutdown |= RCV_SHUTDOWN; 515 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 516 set_bit(MPTCP_DATA_READY, &msk->flags); 517 518 switch (sk->sk_state) { 519 case TCP_ESTABLISHED: 520 inet_sk_state_store(sk, TCP_CLOSE_WAIT); 521 break; 522 case TCP_FIN_WAIT1: 523 inet_sk_state_store(sk, TCP_CLOSING); 524 break; 525 case TCP_FIN_WAIT2: 526 inet_sk_state_store(sk, TCP_CLOSE); 527 break; 528 default: 529 /* Other states not expected */ 530 WARN_ON_ONCE(1); 531 break; 532 } 533 534 ret = true; 535 mptcp_set_timeout(sk, NULL); 536 mptcp_send_ack(msk); 537 mptcp_close_wake_up(sk); 538 } 539 return ret; 540 } 541 542 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 543 struct sock *ssk, 544 unsigned int *bytes) 545 { 546 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 547 struct sock *sk = (struct sock *)msk; 548 unsigned int moved = 0; 549 bool more_data_avail; 550 struct tcp_sock *tp; 551 bool done = false; 552 int sk_rbuf; 553 554 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 555 556 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 557 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 558 559 if (unlikely(ssk_rbuf > sk_rbuf)) { 560 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf); 561 sk_rbuf = ssk_rbuf; 562 } 563 } 564 565 pr_debug("msk=%p ssk=%p", msk, ssk); 566 tp = tcp_sk(ssk); 567 do { 568 u32 map_remaining, offset; 569 u32 seq = tp->copied_seq; 570 struct sk_buff *skb; 571 bool fin; 572 573 /* try to move as much data as available */ 574 map_remaining = subflow->map_data_len - 575 mptcp_subflow_get_map_offset(subflow); 576 577 skb = skb_peek(&ssk->sk_receive_queue); 578 if (!skb) { 579 /* if no data is found, a racing workqueue/recvmsg 580 * already processed the new data, stop here or we 581 * can enter an infinite loop 582 */ 583 if (!moved) 584 done = true; 585 break; 586 } 587 588 if (__mptcp_check_fallback(msk)) { 589 /* if we are running under the workqueue, TCP could have 590 * collapsed skbs between dummy map creation and now 591 * be sure to adjust the size 592 */ 593 map_remaining = skb->len; 594 subflow->map_data_len = skb->len; 595 } 596 597 offset = seq - TCP_SKB_CB(skb)->seq; 598 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 599 if (fin) { 600 done = true; 601 seq++; 602 } 603 604 if (offset < skb->len) { 605 size_t len = skb->len - offset; 606 607 if (tp->urg_data) 608 done = true; 609 610 if (__mptcp_move_skb(msk, ssk, skb, offset, len)) 611 moved += len; 612 seq += len; 613 614 if (WARN_ON_ONCE(map_remaining < len)) 615 break; 616 } else { 617 WARN_ON_ONCE(!fin); 618 sk_eat_skb(ssk, skb); 619 done = true; 620 } 621 622 WRITE_ONCE(tp->copied_seq, seq); 623 more_data_avail = mptcp_subflow_data_available(ssk); 624 625 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) { 626 done = true; 627 break; 628 } 629 } while (more_data_avail); 630 WRITE_ONCE(msk->ack_hint, ssk); 631 632 *bytes += moved; 633 return done; 634 } 635 636 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 637 { 638 struct sock *sk = (struct sock *)msk; 639 struct sk_buff *skb, *tail; 640 bool moved = false; 641 struct rb_node *p; 642 u64 end_seq; 643 644 p = rb_first(&msk->out_of_order_queue); 645 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 646 while (p) { 647 skb = rb_to_skb(p); 648 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 649 break; 650 651 p = rb_next(p); 652 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 653 654 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 655 msk->ack_seq))) { 656 mptcp_drop(sk, skb); 657 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 658 continue; 659 } 660 661 end_seq = MPTCP_SKB_CB(skb)->end_seq; 662 tail = skb_peek_tail(&sk->sk_receive_queue); 663 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 664 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 665 666 /* skip overlapping data, if any */ 667 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d", 668 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 669 delta); 670 MPTCP_SKB_CB(skb)->offset += delta; 671 __skb_queue_tail(&sk->sk_receive_queue, skb); 672 } 673 msk->ack_seq = end_seq; 674 moved = true; 675 } 676 return moved; 677 } 678 679 /* In most cases we will be able to lock the mptcp socket. If its already 680 * owned, we need to defer to the work queue to avoid ABBA deadlock. 681 */ 682 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 683 { 684 struct sock *sk = (struct sock *)msk; 685 unsigned int moved = 0; 686 687 if (inet_sk_state_load(sk) == TCP_CLOSE) 688 return; 689 690 mptcp_data_lock(sk); 691 692 __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 693 __mptcp_ofo_queue(msk); 694 695 /* If the moves have caught up with the DATA_FIN sequence number 696 * it's time to ack the DATA_FIN and change socket state, but 697 * this is not a good place to change state. Let the workqueue 698 * do it. 699 */ 700 if (mptcp_pending_data_fin(sk, NULL)) 701 mptcp_schedule_work(sk); 702 mptcp_data_unlock(sk); 703 } 704 705 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 706 { 707 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 708 struct mptcp_sock *msk = mptcp_sk(sk); 709 int sk_rbuf, ssk_rbuf; 710 bool wake; 711 712 /* The peer can send data while we are shutting down this 713 * subflow at msk destruction time, but we must avoid enqueuing 714 * more data to the msk receive queue 715 */ 716 if (unlikely(subflow->disposable)) 717 return; 718 719 /* move_skbs_to_msk below can legitly clear the data_avail flag, 720 * but we will need later to properly woke the reader, cache its 721 * value 722 */ 723 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL; 724 if (wake) 725 set_bit(MPTCP_DATA_READY, &msk->flags); 726 727 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 728 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 729 if (unlikely(ssk_rbuf > sk_rbuf)) 730 sk_rbuf = ssk_rbuf; 731 732 /* over limit? can't append more skbs to msk */ 733 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) 734 goto wake; 735 736 move_skbs_to_msk(msk, ssk); 737 738 wake: 739 if (wake) 740 sk->sk_data_ready(sk); 741 } 742 743 void __mptcp_flush_join_list(struct mptcp_sock *msk) 744 { 745 struct mptcp_subflow_context *subflow; 746 747 if (likely(list_empty(&msk->join_list))) 748 return; 749 750 spin_lock_bh(&msk->join_list_lock); 751 list_for_each_entry(subflow, &msk->join_list, node) 752 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow)); 753 list_splice_tail_init(&msk->join_list, &msk->conn_list); 754 spin_unlock_bh(&msk->join_list_lock); 755 } 756 757 static bool mptcp_timer_pending(struct sock *sk) 758 { 759 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 760 } 761 762 static void mptcp_reset_timer(struct sock *sk) 763 { 764 struct inet_connection_sock *icsk = inet_csk(sk); 765 unsigned long tout; 766 767 /* prevent rescheduling on close */ 768 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 769 return; 770 771 /* should never be called with mptcp level timer cleared */ 772 tout = READ_ONCE(mptcp_sk(sk)->timer_ival); 773 if (WARN_ON_ONCE(!tout)) 774 tout = TCP_RTO_MIN; 775 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 776 } 777 778 bool mptcp_schedule_work(struct sock *sk) 779 { 780 if (inet_sk_state_load(sk) != TCP_CLOSE && 781 schedule_work(&mptcp_sk(sk)->work)) { 782 /* each subflow already holds a reference to the sk, and the 783 * workqueue is invoked by a subflow, so sk can't go away here. 784 */ 785 sock_hold(sk); 786 return true; 787 } 788 return false; 789 } 790 791 void mptcp_subflow_eof(struct sock *sk) 792 { 793 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags)) 794 mptcp_schedule_work(sk); 795 } 796 797 static void mptcp_check_for_eof(struct mptcp_sock *msk) 798 { 799 struct mptcp_subflow_context *subflow; 800 struct sock *sk = (struct sock *)msk; 801 int receivers = 0; 802 803 mptcp_for_each_subflow(msk, subflow) 804 receivers += !subflow->rx_eof; 805 if (receivers) 806 return; 807 808 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) { 809 /* hopefully temporary hack: propagate shutdown status 810 * to msk, when all subflows agree on it 811 */ 812 sk->sk_shutdown |= RCV_SHUTDOWN; 813 814 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 815 set_bit(MPTCP_DATA_READY, &msk->flags); 816 sk->sk_data_ready(sk); 817 } 818 819 switch (sk->sk_state) { 820 case TCP_ESTABLISHED: 821 inet_sk_state_store(sk, TCP_CLOSE_WAIT); 822 break; 823 case TCP_FIN_WAIT1: 824 inet_sk_state_store(sk, TCP_CLOSING); 825 break; 826 case TCP_FIN_WAIT2: 827 inet_sk_state_store(sk, TCP_CLOSE); 828 break; 829 default: 830 return; 831 } 832 mptcp_close_wake_up(sk); 833 } 834 835 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk) 836 { 837 struct mptcp_subflow_context *subflow; 838 struct sock *sk = (struct sock *)msk; 839 840 sock_owned_by_me(sk); 841 842 mptcp_for_each_subflow(msk, subflow) { 843 if (subflow->data_avail) 844 return mptcp_subflow_tcp_sock(subflow); 845 } 846 847 return NULL; 848 } 849 850 static bool mptcp_skb_can_collapse_to(u64 write_seq, 851 const struct sk_buff *skb, 852 const struct mptcp_ext *mpext) 853 { 854 if (!tcp_skb_can_collapse_to(skb)) 855 return false; 856 857 /* can collapse only if MPTCP level sequence is in order and this 858 * mapping has not been xmitted yet 859 */ 860 return mpext && mpext->data_seq + mpext->data_len == write_seq && 861 !mpext->frozen; 862 } 863 864 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 865 const struct page_frag *pfrag, 866 const struct mptcp_data_frag *df) 867 { 868 return df && pfrag->page == df->page && 869 pfrag->size - pfrag->offset > 0 && 870 df->data_seq + df->data_len == msk->write_seq; 871 } 872 873 static int mptcp_wmem_with_overhead(struct sock *sk, int size) 874 { 875 struct mptcp_sock *msk = mptcp_sk(sk); 876 int ret, skbs; 877 878 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT); 879 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache; 880 if (skbs < msk->skb_tx_cache.qlen) 881 return ret; 882 883 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER); 884 } 885 886 static void __mptcp_wmem_reserve(struct sock *sk, int size) 887 { 888 int amount = mptcp_wmem_with_overhead(sk, size); 889 struct mptcp_sock *msk = mptcp_sk(sk); 890 891 WARN_ON_ONCE(msk->wmem_reserved); 892 if (WARN_ON_ONCE(amount < 0)) 893 amount = 0; 894 895 if (amount <= sk->sk_forward_alloc) 896 goto reserve; 897 898 /* under memory pressure try to reserve at most a single page 899 * otherwise try to reserve the full estimate and fallback 900 * to a single page before entering the error path 901 */ 902 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) || 903 !sk_wmem_schedule(sk, amount)) { 904 if (amount <= PAGE_SIZE) 905 goto nomem; 906 907 amount = PAGE_SIZE; 908 if (!sk_wmem_schedule(sk, amount)) 909 goto nomem; 910 } 911 912 reserve: 913 msk->wmem_reserved = amount; 914 sk->sk_forward_alloc -= amount; 915 return; 916 917 nomem: 918 /* we will wait for memory on next allocation */ 919 msk->wmem_reserved = -1; 920 } 921 922 static void __mptcp_update_wmem(struct sock *sk) 923 { 924 struct mptcp_sock *msk = mptcp_sk(sk); 925 926 if (!msk->wmem_reserved) 927 return; 928 929 if (msk->wmem_reserved < 0) 930 msk->wmem_reserved = 0; 931 if (msk->wmem_reserved > 0) { 932 sk->sk_forward_alloc += msk->wmem_reserved; 933 msk->wmem_reserved = 0; 934 } 935 } 936 937 static bool mptcp_wmem_alloc(struct sock *sk, int size) 938 { 939 struct mptcp_sock *msk = mptcp_sk(sk); 940 941 /* check for pre-existing error condition */ 942 if (msk->wmem_reserved < 0) 943 return false; 944 945 if (msk->wmem_reserved >= size) 946 goto account; 947 948 mptcp_data_lock(sk); 949 if (!sk_wmem_schedule(sk, size)) { 950 mptcp_data_unlock(sk); 951 return false; 952 } 953 954 sk->sk_forward_alloc -= size; 955 msk->wmem_reserved += size; 956 mptcp_data_unlock(sk); 957 958 account: 959 msk->wmem_reserved -= size; 960 return true; 961 } 962 963 static void mptcp_wmem_uncharge(struct sock *sk, int size) 964 { 965 struct mptcp_sock *msk = mptcp_sk(sk); 966 967 if (msk->wmem_reserved < 0) 968 msk->wmem_reserved = 0; 969 msk->wmem_reserved += size; 970 } 971 972 static void mptcp_mem_reclaim_partial(struct sock *sk) 973 { 974 struct mptcp_sock *msk = mptcp_sk(sk); 975 976 /* if we are experiencing a transint allocation error, 977 * the forward allocation memory has been already 978 * released 979 */ 980 if (msk->wmem_reserved < 0) 981 return; 982 983 mptcp_data_lock(sk); 984 sk->sk_forward_alloc += msk->wmem_reserved; 985 sk_mem_reclaim_partial(sk); 986 msk->wmem_reserved = sk->sk_forward_alloc; 987 sk->sk_forward_alloc = 0; 988 mptcp_data_unlock(sk); 989 } 990 991 static void dfrag_uncharge(struct sock *sk, int len) 992 { 993 sk_mem_uncharge(sk, len); 994 sk_wmem_queued_add(sk, -len); 995 } 996 997 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 998 { 999 int len = dfrag->data_len + dfrag->overhead; 1000 1001 list_del(&dfrag->list); 1002 dfrag_uncharge(sk, len); 1003 put_page(dfrag->page); 1004 } 1005 1006 static void __mptcp_clean_una(struct sock *sk) 1007 { 1008 struct mptcp_sock *msk = mptcp_sk(sk); 1009 struct mptcp_data_frag *dtmp, *dfrag; 1010 bool cleaned = false; 1011 u64 snd_una; 1012 1013 /* on fallback we just need to ignore snd_una, as this is really 1014 * plain TCP 1015 */ 1016 if (__mptcp_check_fallback(msk)) 1017 msk->snd_una = READ_ONCE(msk->snd_nxt); 1018 1019 snd_una = msk->snd_una; 1020 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 1021 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 1022 break; 1023 1024 if (WARN_ON_ONCE(dfrag == msk->first_pending)) 1025 break; 1026 dfrag_clear(sk, dfrag); 1027 cleaned = true; 1028 } 1029 1030 dfrag = mptcp_rtx_head(sk); 1031 if (dfrag && after64(snd_una, dfrag->data_seq)) { 1032 u64 delta = snd_una - dfrag->data_seq; 1033 1034 if (WARN_ON_ONCE(delta > dfrag->already_sent)) 1035 goto out; 1036 1037 dfrag->data_seq += delta; 1038 dfrag->offset += delta; 1039 dfrag->data_len -= delta; 1040 dfrag->already_sent -= delta; 1041 1042 dfrag_uncharge(sk, delta); 1043 cleaned = true; 1044 } 1045 1046 out: 1047 if (cleaned) { 1048 if (tcp_under_memory_pressure(sk)) { 1049 __mptcp_update_wmem(sk); 1050 sk_mem_reclaim_partial(sk); 1051 } 1052 } 1053 1054 if (snd_una == READ_ONCE(msk->snd_nxt)) { 1055 if (msk->timer_ival) 1056 mptcp_stop_timer(sk); 1057 } else { 1058 mptcp_reset_timer(sk); 1059 } 1060 } 1061 1062 static void mptcp_enter_memory_pressure(struct sock *sk) 1063 { 1064 struct mptcp_subflow_context *subflow; 1065 struct mptcp_sock *msk = mptcp_sk(sk); 1066 bool first = true; 1067 1068 sk_stream_moderate_sndbuf(sk); 1069 mptcp_for_each_subflow(msk, subflow) { 1070 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1071 1072 if (first) 1073 tcp_enter_memory_pressure(ssk); 1074 sk_stream_moderate_sndbuf(ssk); 1075 first = false; 1076 } 1077 } 1078 1079 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1080 * data 1081 */ 1082 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1083 { 1084 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1085 pfrag, sk->sk_allocation))) 1086 return true; 1087 1088 mptcp_enter_memory_pressure(sk); 1089 return false; 1090 } 1091 1092 static struct mptcp_data_frag * 1093 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1094 int orig_offset) 1095 { 1096 int offset = ALIGN(orig_offset, sizeof(long)); 1097 struct mptcp_data_frag *dfrag; 1098 1099 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1100 dfrag->data_len = 0; 1101 dfrag->data_seq = msk->write_seq; 1102 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1103 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1104 dfrag->already_sent = 0; 1105 dfrag->page = pfrag->page; 1106 1107 return dfrag; 1108 } 1109 1110 struct mptcp_sendmsg_info { 1111 int mss_now; 1112 int size_goal; 1113 u16 limit; 1114 u16 sent; 1115 unsigned int flags; 1116 }; 1117 1118 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq, 1119 int avail_size) 1120 { 1121 u64 window_end = mptcp_wnd_end(msk); 1122 1123 if (__mptcp_check_fallback(msk)) 1124 return avail_size; 1125 1126 if (!before64(data_seq + avail_size, window_end)) { 1127 u64 allowed_size = window_end - data_seq; 1128 1129 return min_t(unsigned int, allowed_size, avail_size); 1130 } 1131 1132 return avail_size; 1133 } 1134 1135 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1136 { 1137 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1138 1139 if (!mpext) 1140 return false; 1141 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1142 return true; 1143 } 1144 1145 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1146 { 1147 struct sk_buff *skb; 1148 1149 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1150 if (likely(skb)) { 1151 if (likely(__mptcp_add_ext(skb, gfp))) { 1152 skb_reserve(skb, MAX_TCP_HEADER); 1153 skb->reserved_tailroom = skb->end - skb->tail; 1154 return skb; 1155 } 1156 __kfree_skb(skb); 1157 } else { 1158 mptcp_enter_memory_pressure(sk); 1159 } 1160 return NULL; 1161 } 1162 1163 static bool mptcp_tx_cache_refill(struct sock *sk, int size, 1164 struct sk_buff_head *skbs, int *total_ts) 1165 { 1166 struct mptcp_sock *msk = mptcp_sk(sk); 1167 struct sk_buff *skb; 1168 int space_needed; 1169 1170 if (unlikely(tcp_under_memory_pressure(sk))) { 1171 mptcp_mem_reclaim_partial(sk); 1172 1173 /* under pressure pre-allocate at most a single skb */ 1174 if (msk->skb_tx_cache.qlen) 1175 return true; 1176 space_needed = msk->size_goal_cache; 1177 } else { 1178 space_needed = msk->tx_pending_data + size - 1179 msk->skb_tx_cache.qlen * msk->size_goal_cache; 1180 } 1181 1182 while (space_needed > 0) { 1183 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation); 1184 if (unlikely(!skb)) { 1185 /* under memory pressure, try to pass the caller a 1186 * single skb to allow forward progress 1187 */ 1188 while (skbs->qlen > 1) { 1189 skb = __skb_dequeue_tail(skbs); 1190 __kfree_skb(skb); 1191 } 1192 return skbs->qlen > 0; 1193 } 1194 1195 *total_ts += skb->truesize; 1196 __skb_queue_tail(skbs, skb); 1197 space_needed -= msk->size_goal_cache; 1198 } 1199 return true; 1200 } 1201 1202 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1203 { 1204 struct mptcp_sock *msk = mptcp_sk(sk); 1205 struct sk_buff *skb; 1206 1207 if (ssk->sk_tx_skb_cache) { 1208 skb = ssk->sk_tx_skb_cache; 1209 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) && 1210 !__mptcp_add_ext(skb, gfp))) 1211 return false; 1212 return true; 1213 } 1214 1215 skb = skb_peek(&msk->skb_tx_cache); 1216 if (skb) { 1217 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1218 skb = __skb_dequeue(&msk->skb_tx_cache); 1219 if (WARN_ON_ONCE(!skb)) 1220 return false; 1221 1222 mptcp_wmem_uncharge(sk, skb->truesize); 1223 ssk->sk_tx_skb_cache = skb; 1224 return true; 1225 } 1226 1227 /* over memory limit, no point to try to allocate a new skb */ 1228 return false; 1229 } 1230 1231 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1232 if (!skb) 1233 return false; 1234 1235 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1236 ssk->sk_tx_skb_cache = skb; 1237 return true; 1238 } 1239 kfree_skb(skb); 1240 return false; 1241 } 1242 1243 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk) 1244 { 1245 return !ssk->sk_tx_skb_cache && 1246 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) && 1247 tcp_under_memory_pressure(sk); 1248 } 1249 1250 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk) 1251 { 1252 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) 1253 mptcp_mem_reclaim_partial(sk); 1254 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation); 1255 } 1256 1257 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1258 struct mptcp_data_frag *dfrag, 1259 struct mptcp_sendmsg_info *info) 1260 { 1261 u64 data_seq = dfrag->data_seq + info->sent; 1262 struct mptcp_sock *msk = mptcp_sk(sk); 1263 bool zero_window_probe = false; 1264 struct mptcp_ext *mpext = NULL; 1265 struct sk_buff *skb, *tail; 1266 bool can_collapse = false; 1267 int size_bias = 0; 1268 int avail_size; 1269 size_t ret = 0; 1270 1271 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d", 1272 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1273 1274 /* compute send limit */ 1275 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1276 avail_size = info->size_goal; 1277 msk->size_goal_cache = info->size_goal; 1278 skb = tcp_write_queue_tail(ssk); 1279 if (skb) { 1280 /* Limit the write to the size available in the 1281 * current skb, if any, so that we create at most a new skb. 1282 * Explicitly tells TCP internals to avoid collapsing on later 1283 * queue management operation, to avoid breaking the ext <-> 1284 * SSN association set here 1285 */ 1286 mpext = skb_ext_find(skb, SKB_EXT_MPTCP); 1287 can_collapse = (info->size_goal - skb->len > 0) && 1288 mptcp_skb_can_collapse_to(data_seq, skb, mpext); 1289 if (!can_collapse) { 1290 TCP_SKB_CB(skb)->eor = 1; 1291 } else { 1292 size_bias = skb->len; 1293 avail_size = info->size_goal - skb->len; 1294 } 1295 } 1296 1297 /* Zero window and all data acked? Probe. */ 1298 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size); 1299 if (avail_size == 0) { 1300 u64 snd_una = READ_ONCE(msk->snd_una); 1301 1302 if (skb || snd_una != msk->snd_nxt) 1303 return 0; 1304 zero_window_probe = true; 1305 data_seq = snd_una - 1; 1306 avail_size = 1; 1307 } 1308 1309 if (WARN_ON_ONCE(info->sent > info->limit || 1310 info->limit > dfrag->data_len)) 1311 return 0; 1312 1313 ret = info->limit - info->sent; 1314 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags, 1315 dfrag->page, dfrag->offset + info->sent, &ret); 1316 if (!tail) { 1317 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk)); 1318 return -ENOMEM; 1319 } 1320 1321 /* if the tail skb is still the cached one, collapsing really happened. 1322 */ 1323 if (skb == tail) { 1324 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH; 1325 mpext->data_len += ret; 1326 WARN_ON_ONCE(!can_collapse); 1327 WARN_ON_ONCE(zero_window_probe); 1328 goto out; 1329 } 1330 1331 mpext = skb_ext_find(tail, SKB_EXT_MPTCP); 1332 if (WARN_ON_ONCE(!mpext)) { 1333 /* should never reach here, stream corrupted */ 1334 return -EINVAL; 1335 } 1336 1337 memset(mpext, 0, sizeof(*mpext)); 1338 mpext->data_seq = data_seq; 1339 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1340 mpext->data_len = ret; 1341 mpext->use_map = 1; 1342 mpext->dsn64 = 1; 1343 1344 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d", 1345 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1346 mpext->dsn64); 1347 1348 if (zero_window_probe) { 1349 mptcp_subflow_ctx(ssk)->rel_write_seq += ret; 1350 mpext->frozen = 1; 1351 ret = 0; 1352 tcp_push_pending_frames(ssk); 1353 } 1354 out: 1355 mptcp_subflow_ctx(ssk)->rel_write_seq += ret; 1356 return ret; 1357 } 1358 1359 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1360 sizeof(struct tcphdr) - \ 1361 MAX_TCP_OPTION_SPACE - \ 1362 sizeof(struct ipv6hdr) - \ 1363 sizeof(struct frag_hdr)) 1364 1365 struct subflow_send_info { 1366 struct sock *ssk; 1367 u64 ratio; 1368 }; 1369 1370 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1371 { 1372 struct subflow_send_info send_info[2]; 1373 struct mptcp_subflow_context *subflow; 1374 int i, nr_active = 0; 1375 struct sock *ssk; 1376 u64 ratio; 1377 u32 pace; 1378 1379 sock_owned_by_me((struct sock *)msk); 1380 1381 if (__mptcp_check_fallback(msk)) { 1382 if (!msk->first) 1383 return NULL; 1384 return sk_stream_memory_free(msk->first) ? msk->first : NULL; 1385 } 1386 1387 /* re-use last subflow, if the burst allow that */ 1388 if (msk->last_snd && msk->snd_burst > 0 && 1389 sk_stream_memory_free(msk->last_snd) && 1390 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) 1391 return msk->last_snd; 1392 1393 /* pick the subflow with the lower wmem/wspace ratio */ 1394 for (i = 0; i < 2; ++i) { 1395 send_info[i].ssk = NULL; 1396 send_info[i].ratio = -1; 1397 } 1398 mptcp_for_each_subflow(msk, subflow) { 1399 ssk = mptcp_subflow_tcp_sock(subflow); 1400 if (!mptcp_subflow_active(subflow)) 1401 continue; 1402 1403 nr_active += !subflow->backup; 1404 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd) 1405 continue; 1406 1407 pace = READ_ONCE(ssk->sk_pacing_rate); 1408 if (!pace) 1409 continue; 1410 1411 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, 1412 pace); 1413 if (ratio < send_info[subflow->backup].ratio) { 1414 send_info[subflow->backup].ssk = ssk; 1415 send_info[subflow->backup].ratio = ratio; 1416 } 1417 } 1418 1419 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld", 1420 msk, nr_active, send_info[0].ssk, send_info[0].ratio, 1421 send_info[1].ssk, send_info[1].ratio); 1422 1423 /* pick the best backup if no other subflow is active */ 1424 if (!nr_active) 1425 send_info[0].ssk = send_info[1].ssk; 1426 1427 if (send_info[0].ssk) { 1428 msk->last_snd = send_info[0].ssk; 1429 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE, 1430 tcp_sk(msk->last_snd)->snd_wnd); 1431 return msk->last_snd; 1432 } 1433 1434 return NULL; 1435 } 1436 1437 static void mptcp_push_release(struct sock *sk, struct sock *ssk, 1438 struct mptcp_sendmsg_info *info) 1439 { 1440 mptcp_set_timeout(sk, ssk); 1441 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1442 release_sock(ssk); 1443 } 1444 1445 static void mptcp_push_pending(struct sock *sk, unsigned int flags) 1446 { 1447 struct sock *prev_ssk = NULL, *ssk = NULL; 1448 struct mptcp_sock *msk = mptcp_sk(sk); 1449 struct mptcp_sendmsg_info info = { 1450 .flags = flags, 1451 }; 1452 struct mptcp_data_frag *dfrag; 1453 int len, copied = 0; 1454 1455 while ((dfrag = mptcp_send_head(sk))) { 1456 info.sent = dfrag->already_sent; 1457 info.limit = dfrag->data_len; 1458 len = dfrag->data_len - dfrag->already_sent; 1459 while (len > 0) { 1460 int ret = 0; 1461 1462 prev_ssk = ssk; 1463 __mptcp_flush_join_list(msk); 1464 ssk = mptcp_subflow_get_send(msk); 1465 1466 /* try to keep the subflow socket lock across 1467 * consecutive xmit on the same socket 1468 */ 1469 if (ssk != prev_ssk && prev_ssk) 1470 mptcp_push_release(sk, prev_ssk, &info); 1471 if (!ssk) 1472 goto out; 1473 1474 if (ssk != prev_ssk || !prev_ssk) 1475 lock_sock(ssk); 1476 1477 /* keep it simple and always provide a new skb for the 1478 * subflow, even if we will not use it when collapsing 1479 * on the pending one 1480 */ 1481 if (!mptcp_alloc_tx_skb(sk, ssk)) { 1482 mptcp_push_release(sk, ssk, &info); 1483 goto out; 1484 } 1485 1486 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 1487 if (ret <= 0) { 1488 mptcp_push_release(sk, ssk, &info); 1489 goto out; 1490 } 1491 1492 info.sent += ret; 1493 dfrag->already_sent += ret; 1494 msk->snd_nxt += ret; 1495 msk->snd_burst -= ret; 1496 msk->tx_pending_data -= ret; 1497 copied += ret; 1498 len -= ret; 1499 } 1500 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1501 } 1502 1503 /* at this point we held the socket lock for the last subflow we used */ 1504 if (ssk) 1505 mptcp_push_release(sk, ssk, &info); 1506 1507 out: 1508 if (copied) { 1509 /* start the timer, if it's not pending */ 1510 if (!mptcp_timer_pending(sk)) 1511 mptcp_reset_timer(sk); 1512 __mptcp_check_send_data_fin(sk); 1513 } 1514 } 1515 1516 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk) 1517 { 1518 struct mptcp_sock *msk = mptcp_sk(sk); 1519 struct mptcp_sendmsg_info info; 1520 struct mptcp_data_frag *dfrag; 1521 struct sock *xmit_ssk; 1522 int len, copied = 0; 1523 bool first = true; 1524 1525 info.flags = 0; 1526 while ((dfrag = mptcp_send_head(sk))) { 1527 info.sent = dfrag->already_sent; 1528 info.limit = dfrag->data_len; 1529 len = dfrag->data_len - dfrag->already_sent; 1530 while (len > 0) { 1531 int ret = 0; 1532 1533 /* the caller already invoked the packet scheduler, 1534 * check for a different subflow usage only after 1535 * spooling the first chunk of data 1536 */ 1537 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk)); 1538 if (!xmit_ssk) 1539 goto out; 1540 if (xmit_ssk != ssk) { 1541 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk)); 1542 goto out; 1543 } 1544 1545 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) { 1546 __mptcp_update_wmem(sk); 1547 sk_mem_reclaim_partial(sk); 1548 } 1549 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC)) 1550 goto out; 1551 1552 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 1553 if (ret <= 0) 1554 goto out; 1555 1556 info.sent += ret; 1557 dfrag->already_sent += ret; 1558 msk->snd_nxt += ret; 1559 msk->snd_burst -= ret; 1560 msk->tx_pending_data -= ret; 1561 copied += ret; 1562 len -= ret; 1563 first = false; 1564 } 1565 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1566 } 1567 1568 out: 1569 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1570 * not going to flush it via release_sock() 1571 */ 1572 __mptcp_update_wmem(sk); 1573 if (copied) { 1574 mptcp_set_timeout(sk, ssk); 1575 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1576 info.size_goal); 1577 if (!mptcp_timer_pending(sk)) 1578 mptcp_reset_timer(sk); 1579 1580 if (msk->snd_data_fin_enable && 1581 msk->snd_nxt + 1 == msk->write_seq) 1582 mptcp_schedule_work(sk); 1583 } 1584 } 1585 1586 static void mptcp_set_nospace(struct sock *sk) 1587 { 1588 /* enable autotune */ 1589 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1590 1591 /* will be cleared on avail space */ 1592 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags); 1593 } 1594 1595 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1596 { 1597 struct mptcp_sock *msk = mptcp_sk(sk); 1598 struct page_frag *pfrag; 1599 size_t copied = 0; 1600 int ret = 0; 1601 long timeo; 1602 1603 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL)) 1604 return -EOPNOTSUPP; 1605 1606 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len))); 1607 1608 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1609 1610 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1611 ret = sk_stream_wait_connect(sk, &timeo); 1612 if (ret) 1613 goto out; 1614 } 1615 1616 pfrag = sk_page_frag(sk); 1617 1618 while (msg_data_left(msg)) { 1619 int total_ts, frag_truesize = 0; 1620 struct mptcp_data_frag *dfrag; 1621 struct sk_buff_head skbs; 1622 bool dfrag_collapsed; 1623 size_t psize, offset; 1624 1625 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) { 1626 ret = -EPIPE; 1627 goto out; 1628 } 1629 1630 /* reuse tail pfrag, if possible, or carve a new one from the 1631 * page allocator 1632 */ 1633 dfrag = mptcp_pending_tail(sk); 1634 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1635 if (!dfrag_collapsed) { 1636 if (!sk_stream_memory_free(sk)) 1637 goto wait_for_memory; 1638 1639 if (!mptcp_page_frag_refill(sk, pfrag)) 1640 goto wait_for_memory; 1641 1642 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1643 frag_truesize = dfrag->overhead; 1644 } 1645 1646 /* we do not bound vs wspace, to allow a single packet. 1647 * memory accounting will prevent execessive memory usage 1648 * anyway 1649 */ 1650 offset = dfrag->offset + dfrag->data_len; 1651 psize = pfrag->size - offset; 1652 psize = min_t(size_t, psize, msg_data_left(msg)); 1653 total_ts = psize + frag_truesize; 1654 __skb_queue_head_init(&skbs); 1655 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts)) 1656 goto wait_for_memory; 1657 1658 if (!mptcp_wmem_alloc(sk, total_ts)) { 1659 __skb_queue_purge(&skbs); 1660 goto wait_for_memory; 1661 } 1662 1663 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache); 1664 if (copy_page_from_iter(dfrag->page, offset, psize, 1665 &msg->msg_iter) != psize) { 1666 mptcp_wmem_uncharge(sk, psize + frag_truesize); 1667 ret = -EFAULT; 1668 goto out; 1669 } 1670 1671 /* data successfully copied into the write queue */ 1672 copied += psize; 1673 dfrag->data_len += psize; 1674 frag_truesize += psize; 1675 pfrag->offset += frag_truesize; 1676 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1677 msk->tx_pending_data += psize; 1678 1679 /* charge data on mptcp pending queue to the msk socket 1680 * Note: we charge such data both to sk and ssk 1681 */ 1682 sk_wmem_queued_add(sk, frag_truesize); 1683 if (!dfrag_collapsed) { 1684 get_page(dfrag->page); 1685 list_add_tail(&dfrag->list, &msk->rtx_queue); 1686 if (!msk->first_pending) 1687 WRITE_ONCE(msk->first_pending, dfrag); 1688 } 1689 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk, 1690 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1691 !dfrag_collapsed); 1692 1693 continue; 1694 1695 wait_for_memory: 1696 mptcp_set_nospace(sk); 1697 mptcp_push_pending(sk, msg->msg_flags); 1698 ret = sk_stream_wait_memory(sk, &timeo); 1699 if (ret) 1700 goto out; 1701 } 1702 1703 if (copied) 1704 mptcp_push_pending(sk, msg->msg_flags); 1705 1706 out: 1707 release_sock(sk); 1708 return copied ? : ret; 1709 } 1710 1711 static void mptcp_wait_data(struct sock *sk, long *timeo) 1712 { 1713 DEFINE_WAIT_FUNC(wait, woken_wake_function); 1714 struct mptcp_sock *msk = mptcp_sk(sk); 1715 1716 add_wait_queue(sk_sleep(sk), &wait); 1717 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk); 1718 1719 sk_wait_event(sk, timeo, 1720 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait); 1721 1722 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk); 1723 remove_wait_queue(sk_sleep(sk), &wait); 1724 } 1725 1726 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk, 1727 struct msghdr *msg, 1728 size_t len) 1729 { 1730 struct sk_buff *skb; 1731 int copied = 0; 1732 1733 while ((skb = skb_peek(&msk->receive_queue)) != NULL) { 1734 u32 offset = MPTCP_SKB_CB(skb)->offset; 1735 u32 data_len = skb->len - offset; 1736 u32 count = min_t(size_t, len - copied, data_len); 1737 int err; 1738 1739 err = skb_copy_datagram_msg(skb, offset, msg, count); 1740 if (unlikely(err < 0)) { 1741 if (!copied) 1742 return err; 1743 break; 1744 } 1745 1746 copied += count; 1747 1748 if (count < data_len) { 1749 MPTCP_SKB_CB(skb)->offset += count; 1750 break; 1751 } 1752 1753 /* we will bulk release the skb memory later */ 1754 skb->destructor = NULL; 1755 msk->rmem_released += skb->truesize; 1756 __skb_unlink(skb, &msk->receive_queue); 1757 __kfree_skb(skb); 1758 1759 if (copied >= len) 1760 break; 1761 } 1762 1763 return copied; 1764 } 1765 1766 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1767 * 1768 * Only difference: Use highest rtt estimate of the subflows in use. 1769 */ 1770 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1771 { 1772 struct mptcp_subflow_context *subflow; 1773 struct sock *sk = (struct sock *)msk; 1774 u32 time, advmss = 1; 1775 u64 rtt_us, mstamp; 1776 1777 sock_owned_by_me(sk); 1778 1779 if (copied <= 0) 1780 return; 1781 1782 msk->rcvq_space.copied += copied; 1783 1784 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1785 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1786 1787 rtt_us = msk->rcvq_space.rtt_us; 1788 if (rtt_us && time < (rtt_us >> 3)) 1789 return; 1790 1791 rtt_us = 0; 1792 mptcp_for_each_subflow(msk, subflow) { 1793 const struct tcp_sock *tp; 1794 u64 sf_rtt_us; 1795 u32 sf_advmss; 1796 1797 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1798 1799 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1800 sf_advmss = READ_ONCE(tp->advmss); 1801 1802 rtt_us = max(sf_rtt_us, rtt_us); 1803 advmss = max(sf_advmss, advmss); 1804 } 1805 1806 msk->rcvq_space.rtt_us = rtt_us; 1807 if (time < (rtt_us >> 3) || rtt_us == 0) 1808 return; 1809 1810 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 1811 goto new_measure; 1812 1813 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && 1814 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 1815 int rcvmem, rcvbuf; 1816 u64 rcvwin, grow; 1817 1818 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 1819 1820 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 1821 1822 do_div(grow, msk->rcvq_space.space); 1823 rcvwin += (grow << 1); 1824 1825 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER); 1826 while (tcp_win_from_space(sk, rcvmem) < advmss) 1827 rcvmem += 128; 1828 1829 do_div(rcvwin, advmss); 1830 rcvbuf = min_t(u64, rcvwin * rcvmem, 1831 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 1832 1833 if (rcvbuf > sk->sk_rcvbuf) { 1834 u32 window_clamp; 1835 1836 window_clamp = tcp_win_from_space(sk, rcvbuf); 1837 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 1838 1839 /* Make subflows follow along. If we do not do this, we 1840 * get drops at subflow level if skbs can't be moved to 1841 * the mptcp rx queue fast enough (announced rcv_win can 1842 * exceed ssk->sk_rcvbuf). 1843 */ 1844 mptcp_for_each_subflow(msk, subflow) { 1845 struct sock *ssk; 1846 bool slow; 1847 1848 ssk = mptcp_subflow_tcp_sock(subflow); 1849 slow = lock_sock_fast(ssk); 1850 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 1851 tcp_sk(ssk)->window_clamp = window_clamp; 1852 tcp_cleanup_rbuf(ssk, 1); 1853 unlock_sock_fast(ssk, slow); 1854 } 1855 } 1856 } 1857 1858 msk->rcvq_space.space = msk->rcvq_space.copied; 1859 new_measure: 1860 msk->rcvq_space.copied = 0; 1861 msk->rcvq_space.time = mstamp; 1862 } 1863 1864 static void __mptcp_update_rmem(struct sock *sk) 1865 { 1866 struct mptcp_sock *msk = mptcp_sk(sk); 1867 1868 if (!msk->rmem_released) 1869 return; 1870 1871 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc); 1872 sk_mem_uncharge(sk, msk->rmem_released); 1873 msk->rmem_released = 0; 1874 } 1875 1876 static void __mptcp_splice_receive_queue(struct sock *sk) 1877 { 1878 struct mptcp_sock *msk = mptcp_sk(sk); 1879 1880 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue); 1881 } 1882 1883 static bool __mptcp_move_skbs(struct mptcp_sock *msk) 1884 { 1885 struct sock *sk = (struct sock *)msk; 1886 unsigned int moved = 0; 1887 bool ret, done; 1888 1889 __mptcp_flush_join_list(msk); 1890 do { 1891 struct sock *ssk = mptcp_subflow_recv_lookup(msk); 1892 bool slowpath; 1893 1894 /* we can have data pending in the subflows only if the msk 1895 * receive buffer was full at subflow_data_ready() time, 1896 * that is an unlikely slow path. 1897 */ 1898 if (likely(!ssk)) 1899 break; 1900 1901 slowpath = lock_sock_fast(ssk); 1902 mptcp_data_lock(sk); 1903 __mptcp_update_rmem(sk); 1904 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 1905 mptcp_data_unlock(sk); 1906 tcp_cleanup_rbuf(ssk, moved); 1907 unlock_sock_fast(ssk, slowpath); 1908 } while (!done); 1909 1910 /* acquire the data lock only if some input data is pending */ 1911 ret = moved > 0; 1912 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) || 1913 !skb_queue_empty_lockless(&sk->sk_receive_queue)) { 1914 mptcp_data_lock(sk); 1915 __mptcp_update_rmem(sk); 1916 ret |= __mptcp_ofo_queue(msk); 1917 __mptcp_splice_receive_queue(sk); 1918 mptcp_data_unlock(sk); 1919 mptcp_cleanup_rbuf(msk); 1920 } 1921 if (ret) 1922 mptcp_check_data_fin((struct sock *)msk); 1923 return !skb_queue_empty(&msk->receive_queue); 1924 } 1925 1926 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 1927 int nonblock, int flags, int *addr_len) 1928 { 1929 struct mptcp_sock *msk = mptcp_sk(sk); 1930 int copied = 0; 1931 int target; 1932 long timeo; 1933 1934 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT)) 1935 return -EOPNOTSUPP; 1936 1937 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk)); 1938 if (unlikely(sk->sk_state == TCP_LISTEN)) { 1939 copied = -ENOTCONN; 1940 goto out_err; 1941 } 1942 1943 timeo = sock_rcvtimeo(sk, nonblock); 1944 1945 len = min_t(size_t, len, INT_MAX); 1946 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 1947 1948 while (copied < len) { 1949 int bytes_read; 1950 1951 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied); 1952 if (unlikely(bytes_read < 0)) { 1953 if (!copied) 1954 copied = bytes_read; 1955 goto out_err; 1956 } 1957 1958 copied += bytes_read; 1959 1960 /* be sure to advertise window change */ 1961 mptcp_cleanup_rbuf(msk); 1962 1963 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk)) 1964 continue; 1965 1966 /* only the master socket status is relevant here. The exit 1967 * conditions mirror closely tcp_recvmsg() 1968 */ 1969 if (copied >= target) 1970 break; 1971 1972 if (copied) { 1973 if (sk->sk_err || 1974 sk->sk_state == TCP_CLOSE || 1975 (sk->sk_shutdown & RCV_SHUTDOWN) || 1976 !timeo || 1977 signal_pending(current)) 1978 break; 1979 } else { 1980 if (sk->sk_err) { 1981 copied = sock_error(sk); 1982 break; 1983 } 1984 1985 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags)) 1986 mptcp_check_for_eof(msk); 1987 1988 if (sk->sk_shutdown & RCV_SHUTDOWN) { 1989 /* race breaker: the shutdown could be after the 1990 * previous receive queue check 1991 */ 1992 if (__mptcp_move_skbs(msk)) 1993 continue; 1994 break; 1995 } 1996 1997 if (sk->sk_state == TCP_CLOSE) { 1998 copied = -ENOTCONN; 1999 break; 2000 } 2001 2002 if (!timeo) { 2003 copied = -EAGAIN; 2004 break; 2005 } 2006 2007 if (signal_pending(current)) { 2008 copied = sock_intr_errno(timeo); 2009 break; 2010 } 2011 } 2012 2013 pr_debug("block timeout %ld", timeo); 2014 mptcp_wait_data(sk, &timeo); 2015 } 2016 2017 if (skb_queue_empty_lockless(&sk->sk_receive_queue) && 2018 skb_queue_empty(&msk->receive_queue)) { 2019 /* entire backlog drained, clear DATA_READY. */ 2020 clear_bit(MPTCP_DATA_READY, &msk->flags); 2021 2022 /* .. race-breaker: ssk might have gotten new data 2023 * after last __mptcp_move_skbs() returned false. 2024 */ 2025 if (unlikely(__mptcp_move_skbs(msk))) 2026 set_bit(MPTCP_DATA_READY, &msk->flags); 2027 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) { 2028 /* data to read but mptcp_wait_data() cleared DATA_READY */ 2029 set_bit(MPTCP_DATA_READY, &msk->flags); 2030 } 2031 out_err: 2032 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d", 2033 msk, test_bit(MPTCP_DATA_READY, &msk->flags), 2034 skb_queue_empty_lockless(&sk->sk_receive_queue), copied); 2035 mptcp_rcv_space_adjust(msk, copied); 2036 2037 release_sock(sk); 2038 return copied; 2039 } 2040 2041 static void mptcp_retransmit_handler(struct sock *sk) 2042 { 2043 struct mptcp_sock *msk = mptcp_sk(sk); 2044 2045 set_bit(MPTCP_WORK_RTX, &msk->flags); 2046 mptcp_schedule_work(sk); 2047 } 2048 2049 static void mptcp_retransmit_timer(struct timer_list *t) 2050 { 2051 struct inet_connection_sock *icsk = from_timer(icsk, t, 2052 icsk_retransmit_timer); 2053 struct sock *sk = &icsk->icsk_inet.sk; 2054 2055 bh_lock_sock(sk); 2056 if (!sock_owned_by_user(sk)) { 2057 mptcp_retransmit_handler(sk); 2058 } else { 2059 /* delegate our work to tcp_release_cb() */ 2060 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, 2061 &sk->sk_tsq_flags)) 2062 sock_hold(sk); 2063 } 2064 bh_unlock_sock(sk); 2065 sock_put(sk); 2066 } 2067 2068 static void mptcp_timeout_timer(struct timer_list *t) 2069 { 2070 struct sock *sk = from_timer(sk, t, sk_timer); 2071 2072 mptcp_schedule_work(sk); 2073 sock_put(sk); 2074 } 2075 2076 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2077 * level. 2078 * 2079 * A backup subflow is returned only if that is the only kind available. 2080 */ 2081 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk) 2082 { 2083 struct mptcp_subflow_context *subflow; 2084 struct sock *backup = NULL; 2085 2086 sock_owned_by_me((const struct sock *)msk); 2087 2088 if (__mptcp_check_fallback(msk)) 2089 return NULL; 2090 2091 mptcp_for_each_subflow(msk, subflow) { 2092 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2093 2094 if (!mptcp_subflow_active(subflow)) 2095 continue; 2096 2097 /* still data outstanding at TCP level? Don't retransmit. */ 2098 if (!tcp_write_queue_empty(ssk)) { 2099 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss) 2100 continue; 2101 return NULL; 2102 } 2103 2104 if (subflow->backup) { 2105 if (!backup) 2106 backup = ssk; 2107 continue; 2108 } 2109 2110 return ssk; 2111 } 2112 2113 return backup; 2114 } 2115 2116 /* subflow sockets can be either outgoing (connect) or incoming 2117 * (accept). 2118 * 2119 * Outgoing subflows use in-kernel sockets. 2120 * Incoming subflows do not have their own 'struct socket' allocated, 2121 * so we need to use tcp_close() after detaching them from the mptcp 2122 * parent socket. 2123 */ 2124 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2125 struct mptcp_subflow_context *subflow) 2126 { 2127 list_del(&subflow->node); 2128 2129 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2130 2131 /* if we are invoked by the msk cleanup code, the subflow is 2132 * already orphaned 2133 */ 2134 if (ssk->sk_socket) 2135 sock_orphan(ssk); 2136 2137 subflow->disposable = 1; 2138 2139 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2140 * the ssk has been already destroyed, we just need to release the 2141 * reference owned by msk; 2142 */ 2143 if (!inet_csk(ssk)->icsk_ulp_ops) { 2144 kfree_rcu(subflow, rcu); 2145 } else { 2146 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2147 __tcp_close(ssk, 0); 2148 2149 /* close acquired an extra ref */ 2150 __sock_put(ssk); 2151 } 2152 release_sock(ssk); 2153 2154 sock_put(ssk); 2155 } 2156 2157 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2158 struct mptcp_subflow_context *subflow) 2159 { 2160 if (sk->sk_state == TCP_ESTABLISHED) 2161 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2162 __mptcp_close_ssk(sk, ssk, subflow); 2163 } 2164 2165 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2166 { 2167 return 0; 2168 } 2169 2170 static void __mptcp_close_subflow(struct mptcp_sock *msk) 2171 { 2172 struct mptcp_subflow_context *subflow, *tmp; 2173 2174 might_sleep(); 2175 2176 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) { 2177 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2178 2179 if (inet_sk_state_load(ssk) != TCP_CLOSE) 2180 continue; 2181 2182 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2183 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2184 continue; 2185 2186 mptcp_close_ssk((struct sock *)msk, ssk, subflow); 2187 } 2188 } 2189 2190 static bool mptcp_check_close_timeout(const struct sock *sk) 2191 { 2192 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp; 2193 struct mptcp_subflow_context *subflow; 2194 2195 if (delta >= TCP_TIMEWAIT_LEN) 2196 return true; 2197 2198 /* if all subflows are in closed status don't bother with additional 2199 * timeout 2200 */ 2201 mptcp_for_each_subflow(mptcp_sk(sk), subflow) { 2202 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) != 2203 TCP_CLOSE) 2204 return false; 2205 } 2206 return true; 2207 } 2208 2209 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2210 { 2211 struct mptcp_subflow_context *subflow, *tmp; 2212 struct sock *sk = &msk->sk.icsk_inet.sk; 2213 2214 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2215 return; 2216 2217 mptcp_token_destroy(msk); 2218 2219 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) { 2220 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2221 2222 lock_sock(tcp_sk); 2223 if (tcp_sk->sk_state != TCP_CLOSE) { 2224 tcp_send_active_reset(tcp_sk, GFP_ATOMIC); 2225 tcp_set_state(tcp_sk, TCP_CLOSE); 2226 } 2227 release_sock(tcp_sk); 2228 } 2229 2230 inet_sk_state_store(sk, TCP_CLOSE); 2231 sk->sk_shutdown = SHUTDOWN_MASK; 2232 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2233 set_bit(MPTCP_DATA_READY, &msk->flags); 2234 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2235 2236 mptcp_close_wake_up(sk); 2237 } 2238 2239 static void mptcp_worker(struct work_struct *work) 2240 { 2241 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2242 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk; 2243 struct mptcp_sendmsg_info info = {}; 2244 struct mptcp_data_frag *dfrag; 2245 size_t copied = 0; 2246 int state, ret; 2247 2248 lock_sock(sk); 2249 state = sk->sk_state; 2250 if (unlikely(state == TCP_CLOSE)) 2251 goto unlock; 2252 2253 mptcp_check_data_fin_ack(sk); 2254 __mptcp_flush_join_list(msk); 2255 2256 mptcp_check_fastclose(msk); 2257 2258 if (msk->pm.status) 2259 mptcp_pm_nl_work(msk); 2260 2261 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags)) 2262 mptcp_check_for_eof(msk); 2263 2264 __mptcp_check_send_data_fin(sk); 2265 mptcp_check_data_fin(sk); 2266 2267 /* if the msk data is completely acked, or the socket timedout, 2268 * there is no point in keeping around an orphaned sk 2269 */ 2270 if (sock_flag(sk, SOCK_DEAD) && 2271 (mptcp_check_close_timeout(sk) || 2272 (state != sk->sk_state && 2273 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) { 2274 inet_sk_state_store(sk, TCP_CLOSE); 2275 __mptcp_destroy_sock(sk); 2276 goto unlock; 2277 } 2278 2279 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2280 __mptcp_close_subflow(msk); 2281 2282 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2283 goto unlock; 2284 2285 __mptcp_clean_una(sk); 2286 dfrag = mptcp_rtx_head(sk); 2287 if (!dfrag) 2288 goto unlock; 2289 2290 ssk = mptcp_subflow_get_retrans(msk); 2291 if (!ssk) 2292 goto reset_unlock; 2293 2294 lock_sock(ssk); 2295 2296 /* limit retransmission to the bytes already sent on some subflows */ 2297 info.sent = 0; 2298 info.limit = dfrag->already_sent; 2299 while (info.sent < dfrag->already_sent) { 2300 if (!mptcp_alloc_tx_skb(sk, ssk)) 2301 break; 2302 2303 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2304 if (ret <= 0) 2305 break; 2306 2307 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2308 copied += ret; 2309 info.sent += ret; 2310 } 2311 if (copied) 2312 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2313 info.size_goal); 2314 2315 mptcp_set_timeout(sk, ssk); 2316 release_sock(ssk); 2317 2318 reset_unlock: 2319 if (!mptcp_timer_pending(sk)) 2320 mptcp_reset_timer(sk); 2321 2322 unlock: 2323 release_sock(sk); 2324 sock_put(sk); 2325 } 2326 2327 static int __mptcp_init_sock(struct sock *sk) 2328 { 2329 struct mptcp_sock *msk = mptcp_sk(sk); 2330 2331 spin_lock_init(&msk->join_list_lock); 2332 2333 INIT_LIST_HEAD(&msk->conn_list); 2334 INIT_LIST_HEAD(&msk->join_list); 2335 INIT_LIST_HEAD(&msk->rtx_queue); 2336 INIT_WORK(&msk->work, mptcp_worker); 2337 __skb_queue_head_init(&msk->receive_queue); 2338 __skb_queue_head_init(&msk->skb_tx_cache); 2339 msk->out_of_order_queue = RB_ROOT; 2340 msk->first_pending = NULL; 2341 msk->wmem_reserved = 0; 2342 msk->rmem_released = 0; 2343 msk->tx_pending_data = 0; 2344 msk->size_goal_cache = TCP_BASE_MSS; 2345 2346 msk->ack_hint = NULL; 2347 msk->first = NULL; 2348 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2349 2350 mptcp_pm_data_init(msk); 2351 2352 /* re-use the csk retrans timer for MPTCP-level retrans */ 2353 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2354 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0); 2355 return 0; 2356 } 2357 2358 static int mptcp_init_sock(struct sock *sk) 2359 { 2360 struct net *net = sock_net(sk); 2361 int ret; 2362 2363 ret = __mptcp_init_sock(sk); 2364 if (ret) 2365 return ret; 2366 2367 if (!mptcp_is_enabled(net)) 2368 return -ENOPROTOOPT; 2369 2370 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2371 return -ENOMEM; 2372 2373 ret = __mptcp_socket_create(mptcp_sk(sk)); 2374 if (ret) 2375 return ret; 2376 2377 sk_sockets_allocated_inc(sk); 2378 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1]; 2379 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1]; 2380 2381 return 0; 2382 } 2383 2384 static void __mptcp_clear_xmit(struct sock *sk) 2385 { 2386 struct mptcp_sock *msk = mptcp_sk(sk); 2387 struct mptcp_data_frag *dtmp, *dfrag; 2388 struct sk_buff *skb; 2389 2390 WRITE_ONCE(msk->first_pending, NULL); 2391 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2392 dfrag_clear(sk, dfrag); 2393 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) { 2394 sk->sk_forward_alloc += skb->truesize; 2395 kfree_skb(skb); 2396 } 2397 } 2398 2399 static void mptcp_cancel_work(struct sock *sk) 2400 { 2401 struct mptcp_sock *msk = mptcp_sk(sk); 2402 2403 if (cancel_work_sync(&msk->work)) 2404 __sock_put(sk); 2405 } 2406 2407 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2408 { 2409 lock_sock(ssk); 2410 2411 switch (ssk->sk_state) { 2412 case TCP_LISTEN: 2413 if (!(how & RCV_SHUTDOWN)) 2414 break; 2415 fallthrough; 2416 case TCP_SYN_SENT: 2417 tcp_disconnect(ssk, O_NONBLOCK); 2418 break; 2419 default: 2420 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2421 pr_debug("Fallback"); 2422 ssk->sk_shutdown |= how; 2423 tcp_shutdown(ssk, how); 2424 } else { 2425 pr_debug("Sending DATA_FIN on subflow %p", ssk); 2426 mptcp_set_timeout(sk, ssk); 2427 tcp_send_ack(ssk); 2428 } 2429 break; 2430 } 2431 2432 release_sock(ssk); 2433 } 2434 2435 static const unsigned char new_state[16] = { 2436 /* current state: new state: action: */ 2437 [0 /* (Invalid) */] = TCP_CLOSE, 2438 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2439 [TCP_SYN_SENT] = TCP_CLOSE, 2440 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2441 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2442 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2443 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2444 [TCP_CLOSE] = TCP_CLOSE, 2445 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2446 [TCP_LAST_ACK] = TCP_LAST_ACK, 2447 [TCP_LISTEN] = TCP_CLOSE, 2448 [TCP_CLOSING] = TCP_CLOSING, 2449 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2450 }; 2451 2452 static int mptcp_close_state(struct sock *sk) 2453 { 2454 int next = (int)new_state[sk->sk_state]; 2455 int ns = next & TCP_STATE_MASK; 2456 2457 inet_sk_state_store(sk, ns); 2458 2459 return next & TCP_ACTION_FIN; 2460 } 2461 2462 static void __mptcp_check_send_data_fin(struct sock *sk) 2463 { 2464 struct mptcp_subflow_context *subflow; 2465 struct mptcp_sock *msk = mptcp_sk(sk); 2466 2467 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu", 2468 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2469 msk->snd_nxt, msk->write_seq); 2470 2471 /* we still need to enqueue subflows or not really shutting down, 2472 * skip this 2473 */ 2474 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2475 mptcp_send_head(sk)) 2476 return; 2477 2478 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2479 2480 /* fallback socket will not get data_fin/ack, can move to the next 2481 * state now 2482 */ 2483 if (__mptcp_check_fallback(msk)) { 2484 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) { 2485 inet_sk_state_store(sk, TCP_CLOSE); 2486 mptcp_close_wake_up(sk); 2487 } else if (sk->sk_state == TCP_FIN_WAIT1) { 2488 inet_sk_state_store(sk, TCP_FIN_WAIT2); 2489 } 2490 } 2491 2492 __mptcp_flush_join_list(msk); 2493 mptcp_for_each_subflow(msk, subflow) { 2494 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2495 2496 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2497 } 2498 } 2499 2500 static void __mptcp_wr_shutdown(struct sock *sk) 2501 { 2502 struct mptcp_sock *msk = mptcp_sk(sk); 2503 2504 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d", 2505 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2506 !!mptcp_send_head(sk)); 2507 2508 /* will be ignored by fallback sockets */ 2509 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2510 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2511 2512 __mptcp_check_send_data_fin(sk); 2513 } 2514 2515 static void __mptcp_destroy_sock(struct sock *sk) 2516 { 2517 struct mptcp_subflow_context *subflow, *tmp; 2518 struct mptcp_sock *msk = mptcp_sk(sk); 2519 LIST_HEAD(conn_list); 2520 2521 pr_debug("msk=%p", msk); 2522 2523 might_sleep(); 2524 2525 /* dispose the ancillatory tcp socket, if any */ 2526 if (msk->subflow) { 2527 iput(SOCK_INODE(msk->subflow)); 2528 msk->subflow = NULL; 2529 } 2530 2531 /* be sure to always acquire the join list lock, to sync vs 2532 * mptcp_finish_join(). 2533 */ 2534 spin_lock_bh(&msk->join_list_lock); 2535 list_splice_tail_init(&msk->join_list, &msk->conn_list); 2536 spin_unlock_bh(&msk->join_list_lock); 2537 list_splice_init(&msk->conn_list, &conn_list); 2538 2539 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer); 2540 sk_stop_timer(sk, &sk->sk_timer); 2541 msk->pm.status = 0; 2542 2543 list_for_each_entry_safe(subflow, tmp, &conn_list, node) { 2544 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2545 __mptcp_close_ssk(sk, ssk, subflow); 2546 } 2547 2548 sk->sk_prot->destroy(sk); 2549 2550 WARN_ON_ONCE(msk->wmem_reserved); 2551 WARN_ON_ONCE(msk->rmem_released); 2552 sk_stream_kill_queues(sk); 2553 xfrm_sk_free_policy(sk); 2554 sk_refcnt_debug_release(sk); 2555 sock_put(sk); 2556 } 2557 2558 static void mptcp_close(struct sock *sk, long timeout) 2559 { 2560 struct mptcp_subflow_context *subflow; 2561 bool do_cancel_work = false; 2562 2563 lock_sock(sk); 2564 sk->sk_shutdown = SHUTDOWN_MASK; 2565 2566 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 2567 inet_sk_state_store(sk, TCP_CLOSE); 2568 goto cleanup; 2569 } 2570 2571 if (mptcp_close_state(sk)) 2572 __mptcp_wr_shutdown(sk); 2573 2574 sk_stream_wait_close(sk, timeout); 2575 2576 cleanup: 2577 /* orphan all the subflows */ 2578 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32; 2579 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) { 2580 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2581 bool slow = lock_sock_fast(ssk); 2582 2583 sock_orphan(ssk); 2584 unlock_sock_fast(ssk, slow); 2585 } 2586 sock_orphan(sk); 2587 2588 sock_hold(sk); 2589 pr_debug("msk=%p state=%d", sk, sk->sk_state); 2590 if (sk->sk_state == TCP_CLOSE) { 2591 __mptcp_destroy_sock(sk); 2592 do_cancel_work = true; 2593 } else { 2594 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN); 2595 } 2596 release_sock(sk); 2597 if (do_cancel_work) 2598 mptcp_cancel_work(sk); 2599 2600 if (mptcp_sk(sk)->token) 2601 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL); 2602 2603 sock_put(sk); 2604 } 2605 2606 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 2607 { 2608 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2609 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 2610 struct ipv6_pinfo *msk6 = inet6_sk(msk); 2611 2612 msk->sk_v6_daddr = ssk->sk_v6_daddr; 2613 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 2614 2615 if (msk6 && ssk6) { 2616 msk6->saddr = ssk6->saddr; 2617 msk6->flow_label = ssk6->flow_label; 2618 } 2619 #endif 2620 2621 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 2622 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 2623 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 2624 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 2625 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 2626 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 2627 } 2628 2629 static int mptcp_disconnect(struct sock *sk, int flags) 2630 { 2631 struct mptcp_subflow_context *subflow; 2632 struct mptcp_sock *msk = mptcp_sk(sk); 2633 2634 __mptcp_flush_join_list(msk); 2635 mptcp_for_each_subflow(msk, subflow) { 2636 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2637 2638 lock_sock(ssk); 2639 tcp_disconnect(ssk, flags); 2640 release_sock(ssk); 2641 } 2642 return 0; 2643 } 2644 2645 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2646 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 2647 { 2648 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo); 2649 2650 return (struct ipv6_pinfo *)(((u8 *)sk) + offset); 2651 } 2652 #endif 2653 2654 struct sock *mptcp_sk_clone(const struct sock *sk, 2655 const struct mptcp_options_received *mp_opt, 2656 struct request_sock *req) 2657 { 2658 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 2659 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 2660 struct mptcp_sock *msk; 2661 u64 ack_seq; 2662 2663 if (!nsk) 2664 return NULL; 2665 2666 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2667 if (nsk->sk_family == AF_INET6) 2668 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 2669 #endif 2670 2671 __mptcp_init_sock(nsk); 2672 2673 msk = mptcp_sk(nsk); 2674 msk->local_key = subflow_req->local_key; 2675 msk->token = subflow_req->token; 2676 msk->subflow = NULL; 2677 WRITE_ONCE(msk->fully_established, false); 2678 2679 msk->write_seq = subflow_req->idsn + 1; 2680 msk->snd_nxt = msk->write_seq; 2681 msk->snd_una = msk->write_seq; 2682 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd; 2683 2684 if (mp_opt->mp_capable) { 2685 msk->can_ack = true; 2686 msk->remote_key = mp_opt->sndr_key; 2687 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq); 2688 ack_seq++; 2689 WRITE_ONCE(msk->ack_seq, ack_seq); 2690 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq); 2691 } 2692 2693 sock_reset_flag(nsk, SOCK_RCU_FREE); 2694 /* will be fully established after successful MPC subflow creation */ 2695 inet_sk_state_store(nsk, TCP_SYN_RECV); 2696 2697 security_inet_csk_clone(nsk, req); 2698 bh_unlock_sock(nsk); 2699 2700 /* keep a single reference */ 2701 __sock_put(nsk); 2702 return nsk; 2703 } 2704 2705 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 2706 { 2707 const struct tcp_sock *tp = tcp_sk(ssk); 2708 2709 msk->rcvq_space.copied = 0; 2710 msk->rcvq_space.rtt_us = 0; 2711 2712 msk->rcvq_space.time = tp->tcp_mstamp; 2713 2714 /* initial rcv_space offering made to peer */ 2715 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 2716 TCP_INIT_CWND * tp->advmss); 2717 if (msk->rcvq_space.space == 0) 2718 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 2719 2720 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 2721 } 2722 2723 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err, 2724 bool kern) 2725 { 2726 struct mptcp_sock *msk = mptcp_sk(sk); 2727 struct socket *listener; 2728 struct sock *newsk; 2729 2730 listener = __mptcp_nmpc_socket(msk); 2731 if (WARN_ON_ONCE(!listener)) { 2732 *err = -EINVAL; 2733 return NULL; 2734 } 2735 2736 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk)); 2737 newsk = inet_csk_accept(listener->sk, flags, err, kern); 2738 if (!newsk) 2739 return NULL; 2740 2741 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk)); 2742 if (sk_is_mptcp(newsk)) { 2743 struct mptcp_subflow_context *subflow; 2744 struct sock *new_mptcp_sock; 2745 2746 subflow = mptcp_subflow_ctx(newsk); 2747 new_mptcp_sock = subflow->conn; 2748 2749 /* is_mptcp should be false if subflow->conn is missing, see 2750 * subflow_syn_recv_sock() 2751 */ 2752 if (WARN_ON_ONCE(!new_mptcp_sock)) { 2753 tcp_sk(newsk)->is_mptcp = 0; 2754 return newsk; 2755 } 2756 2757 /* acquire the 2nd reference for the owning socket */ 2758 sock_hold(new_mptcp_sock); 2759 newsk = new_mptcp_sock; 2760 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 2761 } else { 2762 MPTCP_INC_STATS(sock_net(sk), 2763 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK); 2764 } 2765 2766 return newsk; 2767 } 2768 2769 void mptcp_destroy_common(struct mptcp_sock *msk) 2770 { 2771 struct sock *sk = (struct sock *)msk; 2772 2773 __mptcp_clear_xmit(sk); 2774 2775 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */ 2776 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue); 2777 2778 skb_rbtree_purge(&msk->out_of_order_queue); 2779 mptcp_token_destroy(msk); 2780 mptcp_pm_free_anno_list(msk); 2781 } 2782 2783 static void mptcp_destroy(struct sock *sk) 2784 { 2785 struct mptcp_sock *msk = mptcp_sk(sk); 2786 2787 mptcp_destroy_common(msk); 2788 sk_sockets_allocated_dec(sk); 2789 } 2790 2791 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname, 2792 sockptr_t optval, unsigned int optlen) 2793 { 2794 struct sock *sk = (struct sock *)msk; 2795 struct socket *ssock; 2796 int ret; 2797 2798 switch (optname) { 2799 case SO_REUSEPORT: 2800 case SO_REUSEADDR: 2801 lock_sock(sk); 2802 ssock = __mptcp_nmpc_socket(msk); 2803 if (!ssock) { 2804 release_sock(sk); 2805 return -EINVAL; 2806 } 2807 2808 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen); 2809 if (ret == 0) { 2810 if (optname == SO_REUSEPORT) 2811 sk->sk_reuseport = ssock->sk->sk_reuseport; 2812 else if (optname == SO_REUSEADDR) 2813 sk->sk_reuse = ssock->sk->sk_reuse; 2814 } 2815 release_sock(sk); 2816 return ret; 2817 } 2818 2819 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen); 2820 } 2821 2822 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname, 2823 sockptr_t optval, unsigned int optlen) 2824 { 2825 struct sock *sk = (struct sock *)msk; 2826 int ret = -EOPNOTSUPP; 2827 struct socket *ssock; 2828 2829 switch (optname) { 2830 case IPV6_V6ONLY: 2831 lock_sock(sk); 2832 ssock = __mptcp_nmpc_socket(msk); 2833 if (!ssock) { 2834 release_sock(sk); 2835 return -EINVAL; 2836 } 2837 2838 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen); 2839 if (ret == 0) 2840 sk->sk_ipv6only = ssock->sk->sk_ipv6only; 2841 2842 release_sock(sk); 2843 break; 2844 } 2845 2846 return ret; 2847 } 2848 2849 static int mptcp_setsockopt(struct sock *sk, int level, int optname, 2850 sockptr_t optval, unsigned int optlen) 2851 { 2852 struct mptcp_sock *msk = mptcp_sk(sk); 2853 struct sock *ssk; 2854 2855 pr_debug("msk=%p", msk); 2856 2857 if (level == SOL_SOCKET) 2858 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen); 2859 2860 /* @@ the meaning of setsockopt() when the socket is connected and 2861 * there are multiple subflows is not yet defined. It is up to the 2862 * MPTCP-level socket to configure the subflows until the subflow 2863 * is in TCP fallback, when TCP socket options are passed through 2864 * to the one remaining subflow. 2865 */ 2866 lock_sock(sk); 2867 ssk = __mptcp_tcp_fallback(msk); 2868 release_sock(sk); 2869 if (ssk) 2870 return tcp_setsockopt(ssk, level, optname, optval, optlen); 2871 2872 if (level == SOL_IPV6) 2873 return mptcp_setsockopt_v6(msk, optname, optval, optlen); 2874 2875 return -EOPNOTSUPP; 2876 } 2877 2878 static int mptcp_getsockopt(struct sock *sk, int level, int optname, 2879 char __user *optval, int __user *option) 2880 { 2881 struct mptcp_sock *msk = mptcp_sk(sk); 2882 struct sock *ssk; 2883 2884 pr_debug("msk=%p", msk); 2885 2886 /* @@ the meaning of setsockopt() when the socket is connected and 2887 * there are multiple subflows is not yet defined. It is up to the 2888 * MPTCP-level socket to configure the subflows until the subflow 2889 * is in TCP fallback, when socket options are passed through 2890 * to the one remaining subflow. 2891 */ 2892 lock_sock(sk); 2893 ssk = __mptcp_tcp_fallback(msk); 2894 release_sock(sk); 2895 if (ssk) 2896 return tcp_getsockopt(ssk, level, optname, optval, option); 2897 2898 return -EOPNOTSUPP; 2899 } 2900 2901 void __mptcp_data_acked(struct sock *sk) 2902 { 2903 if (!sock_owned_by_user(sk)) 2904 __mptcp_clean_una(sk); 2905 else 2906 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags); 2907 2908 if (mptcp_pending_data_fin_ack(sk)) 2909 mptcp_schedule_work(sk); 2910 } 2911 2912 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 2913 { 2914 if (!mptcp_send_head(sk)) 2915 return; 2916 2917 if (!sock_owned_by_user(sk)) { 2918 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk)); 2919 2920 if (xmit_ssk == ssk) 2921 __mptcp_subflow_push_pending(sk, ssk); 2922 else if (xmit_ssk) 2923 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk)); 2924 } else { 2925 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags); 2926 } 2927 } 2928 2929 #define MPTCP_DEFERRED_ALL (TCPF_WRITE_TIMER_DEFERRED) 2930 2931 /* processes deferred events and flush wmem */ 2932 static void mptcp_release_cb(struct sock *sk) 2933 { 2934 unsigned long flags, nflags; 2935 2936 /* push_pending may touch wmem_reserved, do it before the later 2937 * cleanup 2938 */ 2939 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags)) 2940 __mptcp_clean_una(sk); 2941 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags)) { 2942 /* mptcp_push_pending() acquires the subflow socket lock 2943 * 2944 * 1) can't be invoked in atomic scope 2945 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 2946 * datapath acquires the msk socket spinlock while helding 2947 * the subflow socket lock 2948 */ 2949 2950 spin_unlock_bh(&sk->sk_lock.slock); 2951 mptcp_push_pending(sk, 0); 2952 spin_lock_bh(&sk->sk_lock.slock); 2953 } 2954 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags)) 2955 __mptcp_error_report(sk); 2956 2957 /* clear any wmem reservation and errors */ 2958 __mptcp_update_wmem(sk); 2959 __mptcp_update_rmem(sk); 2960 2961 do { 2962 flags = sk->sk_tsq_flags; 2963 if (!(flags & MPTCP_DEFERRED_ALL)) 2964 return; 2965 nflags = flags & ~MPTCP_DEFERRED_ALL; 2966 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 2967 2968 sock_release_ownership(sk); 2969 2970 if (flags & TCPF_WRITE_TIMER_DEFERRED) { 2971 mptcp_retransmit_handler(sk); 2972 __sock_put(sk); 2973 } 2974 } 2975 2976 void mptcp_subflow_process_delegated(struct sock *ssk) 2977 { 2978 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 2979 struct sock *sk = subflow->conn; 2980 2981 mptcp_data_lock(sk); 2982 if (!sock_owned_by_user(sk)) 2983 __mptcp_subflow_push_pending(sk, ssk); 2984 else 2985 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags); 2986 mptcp_data_unlock(sk); 2987 mptcp_subflow_delegated_done(subflow); 2988 } 2989 2990 static int mptcp_hash(struct sock *sk) 2991 { 2992 /* should never be called, 2993 * we hash the TCP subflows not the master socket 2994 */ 2995 WARN_ON_ONCE(1); 2996 return 0; 2997 } 2998 2999 static void mptcp_unhash(struct sock *sk) 3000 { 3001 /* called from sk_common_release(), but nothing to do here */ 3002 } 3003 3004 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3005 { 3006 struct mptcp_sock *msk = mptcp_sk(sk); 3007 struct socket *ssock; 3008 3009 ssock = __mptcp_nmpc_socket(msk); 3010 pr_debug("msk=%p, subflow=%p", msk, ssock); 3011 if (WARN_ON_ONCE(!ssock)) 3012 return -EINVAL; 3013 3014 return inet_csk_get_port(ssock->sk, snum); 3015 } 3016 3017 void mptcp_finish_connect(struct sock *ssk) 3018 { 3019 struct mptcp_subflow_context *subflow; 3020 struct mptcp_sock *msk; 3021 struct sock *sk; 3022 u64 ack_seq; 3023 3024 subflow = mptcp_subflow_ctx(ssk); 3025 sk = subflow->conn; 3026 msk = mptcp_sk(sk); 3027 3028 pr_debug("msk=%p, token=%u", sk, subflow->token); 3029 3030 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq); 3031 ack_seq++; 3032 subflow->map_seq = ack_seq; 3033 subflow->map_subflow_seq = 1; 3034 3035 /* the socket is not connected yet, no msk/subflow ops can access/race 3036 * accessing the field below 3037 */ 3038 WRITE_ONCE(msk->remote_key, subflow->remote_key); 3039 WRITE_ONCE(msk->local_key, subflow->local_key); 3040 WRITE_ONCE(msk->write_seq, subflow->idsn + 1); 3041 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3042 WRITE_ONCE(msk->ack_seq, ack_seq); 3043 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq); 3044 WRITE_ONCE(msk->can_ack, 1); 3045 WRITE_ONCE(msk->snd_una, msk->write_seq); 3046 3047 mptcp_pm_new_connection(msk, ssk, 0); 3048 3049 mptcp_rcv_space_init(msk, ssk); 3050 } 3051 3052 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3053 { 3054 write_lock_bh(&sk->sk_callback_lock); 3055 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3056 sk_set_socket(sk, parent); 3057 sk->sk_uid = SOCK_INODE(parent)->i_uid; 3058 write_unlock_bh(&sk->sk_callback_lock); 3059 } 3060 3061 bool mptcp_finish_join(struct sock *ssk) 3062 { 3063 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3064 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3065 struct sock *parent = (void *)msk; 3066 struct socket *parent_sock; 3067 bool ret; 3068 3069 pr_debug("msk=%p, subflow=%p", msk, subflow); 3070 3071 /* mptcp socket already closing? */ 3072 if (!mptcp_is_fully_established(parent)) 3073 return false; 3074 3075 if (!msk->pm.server_side) 3076 goto out; 3077 3078 if (!mptcp_pm_allow_new_subflow(msk)) 3079 return false; 3080 3081 /* active connections are already on conn_list, and we can't acquire 3082 * msk lock here. 3083 * use the join list lock as synchronization point and double-check 3084 * msk status to avoid racing with __mptcp_destroy_sock() 3085 */ 3086 spin_lock_bh(&msk->join_list_lock); 3087 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED; 3088 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) { 3089 list_add_tail(&subflow->node, &msk->join_list); 3090 sock_hold(ssk); 3091 } 3092 spin_unlock_bh(&msk->join_list_lock); 3093 if (!ret) 3094 return false; 3095 3096 /* attach to msk socket only after we are sure he will deal with us 3097 * at close time 3098 */ 3099 parent_sock = READ_ONCE(parent->sk_socket); 3100 if (parent_sock && !ssk->sk_socket) 3101 mptcp_sock_graft(ssk, parent_sock); 3102 subflow->map_seq = READ_ONCE(msk->ack_seq); 3103 out: 3104 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 3105 return true; 3106 } 3107 3108 static void mptcp_shutdown(struct sock *sk, int how) 3109 { 3110 pr_debug("sk=%p, how=%d", sk, how); 3111 3112 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3113 __mptcp_wr_shutdown(sk); 3114 } 3115 3116 static struct proto mptcp_prot = { 3117 .name = "MPTCP", 3118 .owner = THIS_MODULE, 3119 .init = mptcp_init_sock, 3120 .disconnect = mptcp_disconnect, 3121 .close = mptcp_close, 3122 .accept = mptcp_accept, 3123 .setsockopt = mptcp_setsockopt, 3124 .getsockopt = mptcp_getsockopt, 3125 .shutdown = mptcp_shutdown, 3126 .destroy = mptcp_destroy, 3127 .sendmsg = mptcp_sendmsg, 3128 .recvmsg = mptcp_recvmsg, 3129 .release_cb = mptcp_release_cb, 3130 .hash = mptcp_hash, 3131 .unhash = mptcp_unhash, 3132 .get_port = mptcp_get_port, 3133 .sockets_allocated = &mptcp_sockets_allocated, 3134 .memory_allocated = &tcp_memory_allocated, 3135 .memory_pressure = &tcp_memory_pressure, 3136 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3137 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3138 .sysctl_mem = sysctl_tcp_mem, 3139 .obj_size = sizeof(struct mptcp_sock), 3140 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3141 .no_autobind = true, 3142 }; 3143 3144 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3145 { 3146 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3147 struct socket *ssock; 3148 int err; 3149 3150 lock_sock(sock->sk); 3151 ssock = __mptcp_nmpc_socket(msk); 3152 if (!ssock) { 3153 err = -EINVAL; 3154 goto unlock; 3155 } 3156 3157 err = ssock->ops->bind(ssock, uaddr, addr_len); 3158 if (!err) 3159 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3160 3161 unlock: 3162 release_sock(sock->sk); 3163 return err; 3164 } 3165 3166 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk, 3167 struct mptcp_subflow_context *subflow) 3168 { 3169 subflow->request_mptcp = 0; 3170 __mptcp_do_fallback(msk); 3171 } 3172 3173 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr, 3174 int addr_len, int flags) 3175 { 3176 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3177 struct mptcp_subflow_context *subflow; 3178 struct socket *ssock; 3179 int err; 3180 3181 lock_sock(sock->sk); 3182 if (sock->state != SS_UNCONNECTED && msk->subflow) { 3183 /* pending connection or invalid state, let existing subflow 3184 * cope with that 3185 */ 3186 ssock = msk->subflow; 3187 goto do_connect; 3188 } 3189 3190 ssock = __mptcp_nmpc_socket(msk); 3191 if (!ssock) { 3192 err = -EINVAL; 3193 goto unlock; 3194 } 3195 3196 mptcp_token_destroy(msk); 3197 inet_sk_state_store(sock->sk, TCP_SYN_SENT); 3198 subflow = mptcp_subflow_ctx(ssock->sk); 3199 #ifdef CONFIG_TCP_MD5SIG 3200 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3201 * TCP option space. 3202 */ 3203 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info)) 3204 mptcp_subflow_early_fallback(msk, subflow); 3205 #endif 3206 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) 3207 mptcp_subflow_early_fallback(msk, subflow); 3208 3209 do_connect: 3210 err = ssock->ops->connect(ssock, uaddr, addr_len, flags); 3211 sock->state = ssock->state; 3212 3213 /* on successful connect, the msk state will be moved to established by 3214 * subflow_finish_connect() 3215 */ 3216 if (!err || err == -EINPROGRESS) 3217 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3218 else 3219 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk)); 3220 3221 unlock: 3222 release_sock(sock->sk); 3223 return err; 3224 } 3225 3226 static int mptcp_listen(struct socket *sock, int backlog) 3227 { 3228 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3229 struct socket *ssock; 3230 int err; 3231 3232 pr_debug("msk=%p", msk); 3233 3234 lock_sock(sock->sk); 3235 ssock = __mptcp_nmpc_socket(msk); 3236 if (!ssock) { 3237 err = -EINVAL; 3238 goto unlock; 3239 } 3240 3241 mptcp_token_destroy(msk); 3242 inet_sk_state_store(sock->sk, TCP_LISTEN); 3243 sock_set_flag(sock->sk, SOCK_RCU_FREE); 3244 3245 err = ssock->ops->listen(ssock, backlog); 3246 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk)); 3247 if (!err) 3248 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3249 3250 unlock: 3251 release_sock(sock->sk); 3252 return err; 3253 } 3254 3255 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3256 int flags, bool kern) 3257 { 3258 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3259 struct socket *ssock; 3260 int err; 3261 3262 pr_debug("msk=%p", msk); 3263 3264 lock_sock(sock->sk); 3265 if (sock->sk->sk_state != TCP_LISTEN) 3266 goto unlock_fail; 3267 3268 ssock = __mptcp_nmpc_socket(msk); 3269 if (!ssock) 3270 goto unlock_fail; 3271 3272 clear_bit(MPTCP_DATA_READY, &msk->flags); 3273 sock_hold(ssock->sk); 3274 release_sock(sock->sk); 3275 3276 err = ssock->ops->accept(sock, newsock, flags, kern); 3277 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) { 3278 struct mptcp_sock *msk = mptcp_sk(newsock->sk); 3279 struct mptcp_subflow_context *subflow; 3280 struct sock *newsk = newsock->sk; 3281 3282 lock_sock(newsk); 3283 3284 /* PM/worker can now acquire the first subflow socket 3285 * lock without racing with listener queue cleanup, 3286 * we can notify it, if needed. 3287 */ 3288 subflow = mptcp_subflow_ctx(msk->first); 3289 list_add(&subflow->node, &msk->conn_list); 3290 sock_hold(msk->first); 3291 if (mptcp_is_fully_established(newsk)) 3292 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL); 3293 3294 mptcp_copy_inaddrs(newsk, msk->first); 3295 mptcp_rcv_space_init(msk, msk->first); 3296 mptcp_propagate_sndbuf(newsk, msk->first); 3297 3298 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3299 * This is needed so NOSPACE flag can be set from tcp stack. 3300 */ 3301 __mptcp_flush_join_list(msk); 3302 mptcp_for_each_subflow(msk, subflow) { 3303 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3304 3305 if (!ssk->sk_socket) 3306 mptcp_sock_graft(ssk, newsock); 3307 } 3308 release_sock(newsk); 3309 } 3310 3311 if (inet_csk_listen_poll(ssock->sk)) 3312 set_bit(MPTCP_DATA_READY, &msk->flags); 3313 sock_put(ssock->sk); 3314 return err; 3315 3316 unlock_fail: 3317 release_sock(sock->sk); 3318 return -EINVAL; 3319 } 3320 3321 static __poll_t mptcp_check_readable(struct mptcp_sock *msk) 3322 { 3323 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 3324 0; 3325 } 3326 3327 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3328 { 3329 struct sock *sk = (struct sock *)msk; 3330 3331 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN)) 3332 return EPOLLOUT | EPOLLWRNORM; 3333 3334 if (sk_stream_is_writeable(sk)) 3335 return EPOLLOUT | EPOLLWRNORM; 3336 3337 mptcp_set_nospace(sk); 3338 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */ 3339 if (sk_stream_is_writeable(sk)) 3340 return EPOLLOUT | EPOLLWRNORM; 3341 3342 return 0; 3343 } 3344 3345 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3346 struct poll_table_struct *wait) 3347 { 3348 struct sock *sk = sock->sk; 3349 struct mptcp_sock *msk; 3350 __poll_t mask = 0; 3351 int state; 3352 3353 msk = mptcp_sk(sk); 3354 sock_poll_wait(file, sock, wait); 3355 3356 state = inet_sk_state_load(sk); 3357 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags); 3358 if (state == TCP_LISTEN) 3359 return mptcp_check_readable(msk); 3360 3361 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 3362 mask |= mptcp_check_readable(msk); 3363 mask |= mptcp_check_writeable(msk); 3364 } 3365 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 3366 mask |= EPOLLHUP; 3367 if (sk->sk_shutdown & RCV_SHUTDOWN) 3368 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 3369 3370 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 3371 smp_rmb(); 3372 if (sk->sk_err) 3373 mask |= EPOLLERR; 3374 3375 return mask; 3376 } 3377 3378 static const struct proto_ops mptcp_stream_ops = { 3379 .family = PF_INET, 3380 .owner = THIS_MODULE, 3381 .release = inet_release, 3382 .bind = mptcp_bind, 3383 .connect = mptcp_stream_connect, 3384 .socketpair = sock_no_socketpair, 3385 .accept = mptcp_stream_accept, 3386 .getname = inet_getname, 3387 .poll = mptcp_poll, 3388 .ioctl = inet_ioctl, 3389 .gettstamp = sock_gettstamp, 3390 .listen = mptcp_listen, 3391 .shutdown = inet_shutdown, 3392 .setsockopt = sock_common_setsockopt, 3393 .getsockopt = sock_common_getsockopt, 3394 .sendmsg = inet_sendmsg, 3395 .recvmsg = inet_recvmsg, 3396 .mmap = sock_no_mmap, 3397 .sendpage = inet_sendpage, 3398 }; 3399 3400 static struct inet_protosw mptcp_protosw = { 3401 .type = SOCK_STREAM, 3402 .protocol = IPPROTO_MPTCP, 3403 .prot = &mptcp_prot, 3404 .ops = &mptcp_stream_ops, 3405 .flags = INET_PROTOSW_ICSK, 3406 }; 3407 3408 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 3409 { 3410 struct mptcp_delegated_action *delegated; 3411 struct mptcp_subflow_context *subflow; 3412 int work_done = 0; 3413 3414 delegated = container_of(napi, struct mptcp_delegated_action, napi); 3415 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 3416 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3417 3418 bh_lock_sock_nested(ssk); 3419 if (!sock_owned_by_user(ssk) && 3420 mptcp_subflow_has_delegated_action(subflow)) 3421 mptcp_subflow_process_delegated(ssk); 3422 /* ... elsewhere tcp_release_cb_override already processed 3423 * the action or will do at next release_sock(). 3424 * In both case must dequeue the subflow here - on the same 3425 * CPU that scheduled it. 3426 */ 3427 bh_unlock_sock(ssk); 3428 sock_put(ssk); 3429 3430 if (++work_done == budget) 3431 return budget; 3432 } 3433 3434 /* always provide a 0 'work_done' argument, so that napi_complete_done 3435 * will not try accessing the NULL napi->dev ptr 3436 */ 3437 napi_complete_done(napi, 0); 3438 return work_done; 3439 } 3440 3441 void __init mptcp_proto_init(void) 3442 { 3443 struct mptcp_delegated_action *delegated; 3444 int cpu; 3445 3446 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 3447 3448 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 3449 panic("Failed to allocate MPTCP pcpu counter\n"); 3450 3451 init_dummy_netdev(&mptcp_napi_dev); 3452 for_each_possible_cpu(cpu) { 3453 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 3454 INIT_LIST_HEAD(&delegated->head); 3455 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll, 3456 NAPI_POLL_WEIGHT); 3457 napi_enable(&delegated->napi); 3458 } 3459 3460 mptcp_subflow_init(); 3461 mptcp_pm_init(); 3462 mptcp_token_init(); 3463 3464 if (proto_register(&mptcp_prot, 1) != 0) 3465 panic("Failed to register MPTCP proto.\n"); 3466 3467 inet_register_protosw(&mptcp_protosw); 3468 3469 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 3470 } 3471 3472 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3473 static const struct proto_ops mptcp_v6_stream_ops = { 3474 .family = PF_INET6, 3475 .owner = THIS_MODULE, 3476 .release = inet6_release, 3477 .bind = mptcp_bind, 3478 .connect = mptcp_stream_connect, 3479 .socketpair = sock_no_socketpair, 3480 .accept = mptcp_stream_accept, 3481 .getname = inet6_getname, 3482 .poll = mptcp_poll, 3483 .ioctl = inet6_ioctl, 3484 .gettstamp = sock_gettstamp, 3485 .listen = mptcp_listen, 3486 .shutdown = inet_shutdown, 3487 .setsockopt = sock_common_setsockopt, 3488 .getsockopt = sock_common_getsockopt, 3489 .sendmsg = inet6_sendmsg, 3490 .recvmsg = inet6_recvmsg, 3491 .mmap = sock_no_mmap, 3492 .sendpage = inet_sendpage, 3493 #ifdef CONFIG_COMPAT 3494 .compat_ioctl = inet6_compat_ioctl, 3495 #endif 3496 }; 3497 3498 static struct proto mptcp_v6_prot; 3499 3500 static void mptcp_v6_destroy(struct sock *sk) 3501 { 3502 mptcp_destroy(sk); 3503 inet6_destroy_sock(sk); 3504 } 3505 3506 static struct inet_protosw mptcp_v6_protosw = { 3507 .type = SOCK_STREAM, 3508 .protocol = IPPROTO_MPTCP, 3509 .prot = &mptcp_v6_prot, 3510 .ops = &mptcp_v6_stream_ops, 3511 .flags = INET_PROTOSW_ICSK, 3512 }; 3513 3514 int __init mptcp_proto_v6_init(void) 3515 { 3516 int err; 3517 3518 mptcp_v6_prot = mptcp_prot; 3519 strcpy(mptcp_v6_prot.name, "MPTCPv6"); 3520 mptcp_v6_prot.slab = NULL; 3521 mptcp_v6_prot.destroy = mptcp_v6_destroy; 3522 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 3523 3524 err = proto_register(&mptcp_v6_prot, 1); 3525 if (err) 3526 return err; 3527 3528 err = inet6_register_protosw(&mptcp_v6_protosw); 3529 if (err) 3530 proto_unregister(&mptcp_v6_prot); 3531 3532 return err; 3533 } 3534 #endif 3535