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