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 <asm/ioctls.h> 26 #include "protocol.h" 27 #include "mib.h" 28 29 #define CREATE_TRACE_POINTS 30 #include <trace/events/mptcp.h> 31 32 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 33 struct mptcp6_sock { 34 struct mptcp_sock msk; 35 struct ipv6_pinfo np; 36 }; 37 #endif 38 39 enum { 40 MPTCP_CMSG_TS = BIT(0), 41 MPTCP_CMSG_INQ = BIT(1), 42 }; 43 44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp; 45 46 static void __mptcp_destroy_sock(struct sock *sk); 47 static void mptcp_check_send_data_fin(struct sock *sk); 48 49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions); 50 static struct net_device mptcp_napi_dev; 51 52 /* Returns end sequence number of the receiver's advertised window */ 53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 54 { 55 return READ_ONCE(msk->wnd_end); 56 } 57 58 static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk) 59 { 60 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 61 if (sk->sk_prot == &tcpv6_prot) 62 return &inet6_stream_ops; 63 #endif 64 WARN_ON_ONCE(sk->sk_prot != &tcp_prot); 65 return &inet_stream_ops; 66 } 67 68 static int __mptcp_socket_create(struct mptcp_sock *msk) 69 { 70 struct mptcp_subflow_context *subflow; 71 struct sock *sk = (struct sock *)msk; 72 struct socket *ssock; 73 int err; 74 75 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock); 76 if (err) 77 return err; 78 79 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio; 80 WRITE_ONCE(msk->first, ssock->sk); 81 subflow = mptcp_subflow_ctx(ssock->sk); 82 list_add(&subflow->node, &msk->conn_list); 83 sock_hold(ssock->sk); 84 subflow->request_mptcp = 1; 85 subflow->subflow_id = msk->subflow_id++; 86 87 /* This is the first subflow, always with id 0 */ 88 WRITE_ONCE(subflow->local_id, 0); 89 mptcp_sock_graft(msk->first, sk->sk_socket); 90 iput(SOCK_INODE(ssock)); 91 92 return 0; 93 } 94 95 /* If the MPC handshake is not started, returns the first subflow, 96 * eventually allocating it. 97 */ 98 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk) 99 { 100 struct sock *sk = (struct sock *)msk; 101 int ret; 102 103 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 104 return ERR_PTR(-EINVAL); 105 106 if (!msk->first) { 107 ret = __mptcp_socket_create(msk); 108 if (ret) 109 return ERR_PTR(ret); 110 111 mptcp_sockopt_sync(msk, msk->first); 112 } 113 114 return msk->first; 115 } 116 117 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 118 { 119 sk_drops_add(sk, skb); 120 __kfree_skb(skb); 121 } 122 123 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size) 124 { 125 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc, 126 mptcp_sk(sk)->rmem_fwd_alloc + size); 127 } 128 129 static void mptcp_rmem_charge(struct sock *sk, int size) 130 { 131 mptcp_rmem_fwd_alloc_add(sk, -size); 132 } 133 134 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 135 struct sk_buff *from) 136 { 137 bool fragstolen; 138 int delta; 139 140 if (MPTCP_SKB_CB(from)->offset || 141 !skb_try_coalesce(to, from, &fragstolen, &delta)) 142 return false; 143 144 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n", 145 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 146 to->len, MPTCP_SKB_CB(from)->end_seq); 147 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 148 149 /* note the fwd memory can reach a negative value after accounting 150 * for the delta, but the later skb free will restore a non 151 * negative one 152 */ 153 atomic_add(delta, &sk->sk_rmem_alloc); 154 mptcp_rmem_charge(sk, delta); 155 kfree_skb_partial(from, fragstolen); 156 157 return true; 158 } 159 160 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 161 struct sk_buff *from) 162 { 163 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 164 return false; 165 166 return mptcp_try_coalesce((struct sock *)msk, to, from); 167 } 168 169 static void __mptcp_rmem_reclaim(struct sock *sk, int amount) 170 { 171 amount >>= PAGE_SHIFT; 172 mptcp_rmem_charge(sk, amount << PAGE_SHIFT); 173 __sk_mem_reduce_allocated(sk, amount); 174 } 175 176 static void mptcp_rmem_uncharge(struct sock *sk, int size) 177 { 178 struct mptcp_sock *msk = mptcp_sk(sk); 179 int reclaimable; 180 181 mptcp_rmem_fwd_alloc_add(sk, size); 182 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk); 183 184 /* see sk_mem_uncharge() for the rationale behind the following schema */ 185 if (unlikely(reclaimable >= PAGE_SIZE)) 186 __mptcp_rmem_reclaim(sk, reclaimable); 187 } 188 189 static void mptcp_rfree(struct sk_buff *skb) 190 { 191 unsigned int len = skb->truesize; 192 struct sock *sk = skb->sk; 193 194 atomic_sub(len, &sk->sk_rmem_alloc); 195 mptcp_rmem_uncharge(sk, len); 196 } 197 198 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk) 199 { 200 skb_orphan(skb); 201 skb->sk = sk; 202 skb->destructor = mptcp_rfree; 203 atomic_add(skb->truesize, &sk->sk_rmem_alloc); 204 mptcp_rmem_charge(sk, skb->truesize); 205 } 206 207 /* "inspired" by tcp_data_queue_ofo(), main differences: 208 * - use mptcp seqs 209 * - don't cope with sacks 210 */ 211 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 212 { 213 struct sock *sk = (struct sock *)msk; 214 struct rb_node **p, *parent; 215 u64 seq, end_seq, max_seq; 216 struct sk_buff *skb1; 217 218 seq = MPTCP_SKB_CB(skb)->map_seq; 219 end_seq = MPTCP_SKB_CB(skb)->end_seq; 220 max_seq = atomic64_read(&msk->rcv_wnd_sent); 221 222 pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq, 223 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 224 if (after64(end_seq, max_seq)) { 225 /* out of window */ 226 mptcp_drop(sk, skb); 227 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 228 (unsigned long long)end_seq - (unsigned long)max_seq, 229 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent)); 230 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 231 return; 232 } 233 234 p = &msk->out_of_order_queue.rb_node; 235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 236 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 237 rb_link_node(&skb->rbnode, NULL, p); 238 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 239 msk->ooo_last_skb = skb; 240 goto end; 241 } 242 243 /* with 2 subflows, adding at end of ooo queue is quite likely 244 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 245 */ 246 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 248 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 249 return; 250 } 251 252 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 253 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 254 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 255 parent = &msk->ooo_last_skb->rbnode; 256 p = &parent->rb_right; 257 goto insert; 258 } 259 260 /* Find place to insert this segment. Handle overlaps on the way. */ 261 parent = NULL; 262 while (*p) { 263 parent = *p; 264 skb1 = rb_to_skb(parent); 265 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 266 p = &parent->rb_left; 267 continue; 268 } 269 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 270 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 271 /* All the bits are present. Drop. */ 272 mptcp_drop(sk, skb); 273 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 274 return; 275 } 276 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 277 /* partial overlap: 278 * | skb | 279 * | skb1 | 280 * continue traversing 281 */ 282 } else { 283 /* skb's seq == skb1's seq and skb covers skb1. 284 * Replace skb1 with skb. 285 */ 286 rb_replace_node(&skb1->rbnode, &skb->rbnode, 287 &msk->out_of_order_queue); 288 mptcp_drop(sk, skb1); 289 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 290 goto merge_right; 291 } 292 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 293 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 294 return; 295 } 296 p = &parent->rb_right; 297 } 298 299 insert: 300 /* Insert segment into RB tree. */ 301 rb_link_node(&skb->rbnode, parent, p); 302 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 303 304 merge_right: 305 /* Remove other segments covered by skb. */ 306 while ((skb1 = skb_rb_next(skb)) != NULL) { 307 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 308 break; 309 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 310 mptcp_drop(sk, skb1); 311 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 312 } 313 /* If there is no skb after us, we are the last_skb ! */ 314 if (!skb1) 315 msk->ooo_last_skb = skb; 316 317 end: 318 skb_condense(skb); 319 mptcp_set_owner_r(skb, sk); 320 } 321 322 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size) 323 { 324 struct mptcp_sock *msk = mptcp_sk(sk); 325 int amt, amount; 326 327 if (size <= msk->rmem_fwd_alloc) 328 return true; 329 330 size -= msk->rmem_fwd_alloc; 331 amt = sk_mem_pages(size); 332 amount = amt << PAGE_SHIFT; 333 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV)) 334 return false; 335 336 mptcp_rmem_fwd_alloc_add(sk, amount); 337 return true; 338 } 339 340 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 341 struct sk_buff *skb, unsigned int offset, 342 size_t copy_len) 343 { 344 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 345 struct sock *sk = (struct sock *)msk; 346 struct sk_buff *tail; 347 bool has_rxtstamp; 348 349 __skb_unlink(skb, &ssk->sk_receive_queue); 350 351 skb_ext_reset(skb); 352 skb_orphan(skb); 353 354 /* try to fetch required memory from subflow */ 355 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize)) { 356 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 357 goto drop; 358 } 359 360 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 361 362 /* the skb map_seq accounts for the skb offset: 363 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 364 * value 365 */ 366 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 367 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 368 MPTCP_SKB_CB(skb)->offset = offset; 369 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 370 371 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 372 /* in sequence */ 373 msk->bytes_received += copy_len; 374 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 375 tail = skb_peek_tail(&sk->sk_receive_queue); 376 if (tail && mptcp_try_coalesce(sk, tail, skb)) 377 return true; 378 379 mptcp_set_owner_r(skb, sk); 380 __skb_queue_tail(&sk->sk_receive_queue, skb); 381 return true; 382 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 383 mptcp_data_queue_ofo(msk, skb); 384 return false; 385 } 386 387 /* old data, keep it simple and drop the whole pkt, sender 388 * will retransmit as needed, if needed. 389 */ 390 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 391 drop: 392 mptcp_drop(sk, skb); 393 return false; 394 } 395 396 static void mptcp_stop_rtx_timer(struct sock *sk) 397 { 398 struct inet_connection_sock *icsk = inet_csk(sk); 399 400 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 401 mptcp_sk(sk)->timer_ival = 0; 402 } 403 404 static void mptcp_close_wake_up(struct sock *sk) 405 { 406 if (sock_flag(sk, SOCK_DEAD)) 407 return; 408 409 sk->sk_state_change(sk); 410 if (sk->sk_shutdown == SHUTDOWN_MASK || 411 sk->sk_state == TCP_CLOSE) 412 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 413 else 414 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 415 } 416 417 static bool mptcp_pending_data_fin_ack(struct sock *sk) 418 { 419 struct mptcp_sock *msk = mptcp_sk(sk); 420 421 return ((1 << sk->sk_state) & 422 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 423 msk->write_seq == READ_ONCE(msk->snd_una); 424 } 425 426 static void mptcp_check_data_fin_ack(struct sock *sk) 427 { 428 struct mptcp_sock *msk = mptcp_sk(sk); 429 430 /* Look for an acknowledged DATA_FIN */ 431 if (mptcp_pending_data_fin_ack(sk)) { 432 WRITE_ONCE(msk->snd_data_fin_enable, 0); 433 434 switch (sk->sk_state) { 435 case TCP_FIN_WAIT1: 436 mptcp_set_state(sk, TCP_FIN_WAIT2); 437 break; 438 case TCP_CLOSING: 439 case TCP_LAST_ACK: 440 mptcp_set_state(sk, TCP_CLOSE); 441 break; 442 } 443 444 mptcp_close_wake_up(sk); 445 } 446 } 447 448 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 449 { 450 struct mptcp_sock *msk = mptcp_sk(sk); 451 452 if (READ_ONCE(msk->rcv_data_fin) && 453 ((1 << sk->sk_state) & 454 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 455 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 456 457 if (msk->ack_seq == rcv_data_fin_seq) { 458 if (seq) 459 *seq = rcv_data_fin_seq; 460 461 return true; 462 } 463 } 464 465 return false; 466 } 467 468 static void mptcp_set_datafin_timeout(struct sock *sk) 469 { 470 struct inet_connection_sock *icsk = inet_csk(sk); 471 u32 retransmits; 472 473 retransmits = min_t(u32, icsk->icsk_retransmits, 474 ilog2(TCP_RTO_MAX / TCP_RTO_MIN)); 475 476 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits; 477 } 478 479 static void __mptcp_set_timeout(struct sock *sk, long tout) 480 { 481 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 482 } 483 484 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 485 { 486 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 487 488 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 489 inet_csk(ssk)->icsk_timeout - jiffies : 0; 490 } 491 492 static void mptcp_set_timeout(struct sock *sk) 493 { 494 struct mptcp_subflow_context *subflow; 495 long tout = 0; 496 497 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 498 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 499 __mptcp_set_timeout(sk, tout); 500 } 501 502 static inline bool tcp_can_send_ack(const struct sock *ssk) 503 { 504 return !((1 << inet_sk_state_load(ssk)) & 505 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 506 } 507 508 void __mptcp_subflow_send_ack(struct sock *ssk) 509 { 510 if (tcp_can_send_ack(ssk)) 511 tcp_send_ack(ssk); 512 } 513 514 static void mptcp_subflow_send_ack(struct sock *ssk) 515 { 516 bool slow; 517 518 slow = lock_sock_fast(ssk); 519 __mptcp_subflow_send_ack(ssk); 520 unlock_sock_fast(ssk, slow); 521 } 522 523 static void mptcp_send_ack(struct mptcp_sock *msk) 524 { 525 struct mptcp_subflow_context *subflow; 526 527 mptcp_for_each_subflow(msk, subflow) 528 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow)); 529 } 530 531 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk) 532 { 533 bool slow; 534 535 slow = lock_sock_fast(ssk); 536 if (tcp_can_send_ack(ssk)) 537 tcp_cleanup_rbuf(ssk, 1); 538 unlock_sock_fast(ssk, slow); 539 } 540 541 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 542 { 543 const struct inet_connection_sock *icsk = inet_csk(ssk); 544 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 545 const struct tcp_sock *tp = tcp_sk(ssk); 546 547 return (ack_pending & ICSK_ACK_SCHED) && 548 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 549 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 550 (rx_empty && ack_pending & 551 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 552 } 553 554 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk) 555 { 556 int old_space = READ_ONCE(msk->old_wspace); 557 struct mptcp_subflow_context *subflow; 558 struct sock *sk = (struct sock *)msk; 559 int space = __mptcp_space(sk); 560 bool cleanup, rx_empty; 561 562 cleanup = (space > 0) && (space >= (old_space << 1)); 563 rx_empty = !__mptcp_rmem(sk); 564 565 mptcp_for_each_subflow(msk, subflow) { 566 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 567 568 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 569 mptcp_subflow_cleanup_rbuf(ssk); 570 } 571 } 572 573 static bool mptcp_check_data_fin(struct sock *sk) 574 { 575 struct mptcp_sock *msk = mptcp_sk(sk); 576 u64 rcv_data_fin_seq; 577 bool ret = false; 578 579 /* Need to ack a DATA_FIN received from a peer while this side 580 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 581 * msk->rcv_data_fin was set when parsing the incoming options 582 * at the subflow level and the msk lock was not held, so this 583 * is the first opportunity to act on the DATA_FIN and change 584 * the msk state. 585 * 586 * If we are caught up to the sequence number of the incoming 587 * DATA_FIN, send the DATA_ACK now and do state transition. If 588 * not caught up, do nothing and let the recv code send DATA_ACK 589 * when catching up. 590 */ 591 592 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 593 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 594 WRITE_ONCE(msk->rcv_data_fin, 0); 595 596 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 597 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 598 599 switch (sk->sk_state) { 600 case TCP_ESTABLISHED: 601 mptcp_set_state(sk, TCP_CLOSE_WAIT); 602 break; 603 case TCP_FIN_WAIT1: 604 mptcp_set_state(sk, TCP_CLOSING); 605 break; 606 case TCP_FIN_WAIT2: 607 mptcp_set_state(sk, TCP_CLOSE); 608 break; 609 default: 610 /* Other states not expected */ 611 WARN_ON_ONCE(1); 612 break; 613 } 614 615 ret = true; 616 if (!__mptcp_check_fallback(msk)) 617 mptcp_send_ack(msk); 618 mptcp_close_wake_up(sk); 619 } 620 return ret; 621 } 622 623 static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk) 624 { 625 if (READ_ONCE(msk->allow_infinite_fallback)) { 626 MPTCP_INC_STATS(sock_net(ssk), 627 MPTCP_MIB_DSSCORRUPTIONFALLBACK); 628 mptcp_do_fallback(ssk); 629 } else { 630 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET); 631 mptcp_subflow_reset(ssk); 632 } 633 } 634 635 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 636 struct sock *ssk, 637 unsigned int *bytes) 638 { 639 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 640 struct sock *sk = (struct sock *)msk; 641 unsigned int moved = 0; 642 bool more_data_avail; 643 struct tcp_sock *tp; 644 bool done = false; 645 int sk_rbuf; 646 647 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 648 649 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 650 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 651 652 if (unlikely(ssk_rbuf > sk_rbuf)) { 653 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf); 654 sk_rbuf = ssk_rbuf; 655 } 656 } 657 658 pr_debug("msk=%p ssk=%p\n", msk, ssk); 659 tp = tcp_sk(ssk); 660 do { 661 u32 map_remaining, offset; 662 u32 seq = tp->copied_seq; 663 struct sk_buff *skb; 664 bool fin; 665 666 /* try to move as much data as available */ 667 map_remaining = subflow->map_data_len - 668 mptcp_subflow_get_map_offset(subflow); 669 670 skb = skb_peek(&ssk->sk_receive_queue); 671 if (!skb) { 672 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(), 673 * a different CPU can have already processed the pending 674 * data, stop here or we can enter an infinite loop 675 */ 676 if (!moved) 677 done = true; 678 break; 679 } 680 681 if (__mptcp_check_fallback(msk)) { 682 /* Under fallback skbs have no MPTCP extension and TCP could 683 * collapse them between the dummy map creation and the 684 * current dequeue. Be sure to adjust the map size. 685 */ 686 map_remaining = skb->len; 687 subflow->map_data_len = skb->len; 688 } 689 690 offset = seq - TCP_SKB_CB(skb)->seq; 691 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 692 if (fin) { 693 done = true; 694 seq++; 695 } 696 697 if (offset < skb->len) { 698 size_t len = skb->len - offset; 699 700 if (tp->urg_data) 701 done = true; 702 703 if (__mptcp_move_skb(msk, ssk, skb, offset, len)) 704 moved += len; 705 seq += len; 706 707 if (unlikely(map_remaining < len)) { 708 DEBUG_NET_WARN_ON_ONCE(1); 709 mptcp_dss_corruption(msk, ssk); 710 } 711 } else { 712 if (unlikely(!fin)) { 713 DEBUG_NET_WARN_ON_ONCE(1); 714 mptcp_dss_corruption(msk, ssk); 715 } 716 717 sk_eat_skb(ssk, skb); 718 done = true; 719 } 720 721 WRITE_ONCE(tp->copied_seq, seq); 722 more_data_avail = mptcp_subflow_data_available(ssk); 723 724 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) { 725 done = true; 726 break; 727 } 728 } while (more_data_avail); 729 730 *bytes += moved; 731 return done; 732 } 733 734 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 735 { 736 struct sock *sk = (struct sock *)msk; 737 struct sk_buff *skb, *tail; 738 bool moved = false; 739 struct rb_node *p; 740 u64 end_seq; 741 742 p = rb_first(&msk->out_of_order_queue); 743 pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 744 while (p) { 745 skb = rb_to_skb(p); 746 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 747 break; 748 749 p = rb_next(p); 750 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 751 752 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 753 msk->ack_seq))) { 754 mptcp_drop(sk, skb); 755 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 756 continue; 757 } 758 759 end_seq = MPTCP_SKB_CB(skb)->end_seq; 760 tail = skb_peek_tail(&sk->sk_receive_queue); 761 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 762 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 763 764 /* skip overlapping data, if any */ 765 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n", 766 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 767 delta); 768 MPTCP_SKB_CB(skb)->offset += delta; 769 MPTCP_SKB_CB(skb)->map_seq += delta; 770 __skb_queue_tail(&sk->sk_receive_queue, skb); 771 } 772 msk->bytes_received += end_seq - msk->ack_seq; 773 msk->ack_seq = end_seq; 774 moved = true; 775 } 776 return moved; 777 } 778 779 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk) 780 { 781 int err = sock_error(ssk); 782 int ssk_state; 783 784 if (!err) 785 return false; 786 787 /* only propagate errors on fallen-back sockets or 788 * on MPC connect 789 */ 790 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk))) 791 return false; 792 793 /* We need to propagate only transition to CLOSE state. 794 * Orphaned socket will see such state change via 795 * subflow_sched_work_if_closed() and that path will properly 796 * destroy the msk as needed. 797 */ 798 ssk_state = inet_sk_state_load(ssk); 799 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD)) 800 mptcp_set_state(sk, ssk_state); 801 WRITE_ONCE(sk->sk_err, -err); 802 803 /* This barrier is coupled with smp_rmb() in mptcp_poll() */ 804 smp_wmb(); 805 sk_error_report(sk); 806 return true; 807 } 808 809 void __mptcp_error_report(struct sock *sk) 810 { 811 struct mptcp_subflow_context *subflow; 812 struct mptcp_sock *msk = mptcp_sk(sk); 813 814 mptcp_for_each_subflow(msk, subflow) 815 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow))) 816 break; 817 } 818 819 /* In most cases we will be able to lock the mptcp socket. If its already 820 * owned, we need to defer to the work queue to avoid ABBA deadlock. 821 */ 822 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 823 { 824 struct sock *sk = (struct sock *)msk; 825 unsigned int moved = 0; 826 827 __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 828 __mptcp_ofo_queue(msk); 829 if (unlikely(ssk->sk_err)) { 830 if (!sock_owned_by_user(sk)) 831 __mptcp_error_report(sk); 832 else 833 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags); 834 } 835 836 /* If the moves have caught up with the DATA_FIN sequence number 837 * it's time to ack the DATA_FIN and change socket state, but 838 * this is not a good place to change state. Let the workqueue 839 * do it. 840 */ 841 if (mptcp_pending_data_fin(sk, NULL)) 842 mptcp_schedule_work(sk); 843 return moved > 0; 844 } 845 846 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 847 { 848 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 849 struct mptcp_sock *msk = mptcp_sk(sk); 850 int sk_rbuf, ssk_rbuf; 851 852 /* The peer can send data while we are shutting down this 853 * subflow at msk destruction time, but we must avoid enqueuing 854 * more data to the msk receive queue 855 */ 856 if (unlikely(subflow->disposable)) 857 return; 858 859 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 860 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 861 if (unlikely(ssk_rbuf > sk_rbuf)) 862 sk_rbuf = ssk_rbuf; 863 864 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/ 865 if (__mptcp_rmem(sk) > sk_rbuf) 866 return; 867 868 /* Wake-up the reader only for in-sequence data */ 869 mptcp_data_lock(sk); 870 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk)) 871 sk->sk_data_ready(sk); 872 mptcp_data_unlock(sk); 873 } 874 875 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk) 876 { 877 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq); 878 WRITE_ONCE(msk->allow_infinite_fallback, false); 879 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 880 } 881 882 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk) 883 { 884 struct sock *sk = (struct sock *)msk; 885 886 if (sk->sk_state != TCP_ESTABLISHED) 887 return false; 888 889 /* attach to msk socket only after we are sure we will deal with it 890 * at close time 891 */ 892 if (sk->sk_socket && !ssk->sk_socket) 893 mptcp_sock_graft(ssk, sk->sk_socket); 894 895 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++; 896 mptcp_sockopt_sync_locked(msk, ssk); 897 mptcp_subflow_joined(msk, ssk); 898 mptcp_stop_tout_timer(sk); 899 __mptcp_propagate_sndbuf(sk, ssk); 900 return true; 901 } 902 903 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list) 904 { 905 struct mptcp_subflow_context *tmp, *subflow; 906 struct mptcp_sock *msk = mptcp_sk(sk); 907 908 list_for_each_entry_safe(subflow, tmp, join_list, node) { 909 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 910 bool slow = lock_sock_fast(ssk); 911 912 list_move_tail(&subflow->node, &msk->conn_list); 913 if (!__mptcp_finish_join(msk, ssk)) 914 mptcp_subflow_reset(ssk); 915 unlock_sock_fast(ssk, slow); 916 } 917 } 918 919 static bool mptcp_rtx_timer_pending(struct sock *sk) 920 { 921 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 922 } 923 924 static void mptcp_reset_rtx_timer(struct sock *sk) 925 { 926 struct inet_connection_sock *icsk = inet_csk(sk); 927 unsigned long tout; 928 929 /* prevent rescheduling on close */ 930 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 931 return; 932 933 tout = mptcp_sk(sk)->timer_ival; 934 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 935 } 936 937 bool mptcp_schedule_work(struct sock *sk) 938 { 939 if (inet_sk_state_load(sk) != TCP_CLOSE && 940 schedule_work(&mptcp_sk(sk)->work)) { 941 /* each subflow already holds a reference to the sk, and the 942 * workqueue is invoked by a subflow, so sk can't go away here. 943 */ 944 sock_hold(sk); 945 return true; 946 } 947 return false; 948 } 949 950 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk) 951 { 952 struct mptcp_subflow_context *subflow; 953 954 msk_owned_by_me(msk); 955 956 mptcp_for_each_subflow(msk, subflow) { 957 if (READ_ONCE(subflow->data_avail)) 958 return mptcp_subflow_tcp_sock(subflow); 959 } 960 961 return NULL; 962 } 963 964 static bool mptcp_skb_can_collapse_to(u64 write_seq, 965 const struct sk_buff *skb, 966 const struct mptcp_ext *mpext) 967 { 968 if (!tcp_skb_can_collapse_to(skb)) 969 return false; 970 971 /* can collapse only if MPTCP level sequence is in order and this 972 * mapping has not been xmitted yet 973 */ 974 return mpext && mpext->data_seq + mpext->data_len == write_seq && 975 !mpext->frozen; 976 } 977 978 /* we can append data to the given data frag if: 979 * - there is space available in the backing page_frag 980 * - the data frag tail matches the current page_frag free offset 981 * - the data frag end sequence number matches the current write seq 982 */ 983 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 984 const struct page_frag *pfrag, 985 const struct mptcp_data_frag *df) 986 { 987 return df && pfrag->page == df->page && 988 pfrag->size - pfrag->offset > 0 && 989 pfrag->offset == (df->offset + df->data_len) && 990 df->data_seq + df->data_len == msk->write_seq; 991 } 992 993 static void dfrag_uncharge(struct sock *sk, int len) 994 { 995 sk_mem_uncharge(sk, len); 996 sk_wmem_queued_add(sk, -len); 997 } 998 999 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 1000 { 1001 int len = dfrag->data_len + dfrag->overhead; 1002 1003 list_del(&dfrag->list); 1004 dfrag_uncharge(sk, len); 1005 put_page(dfrag->page); 1006 } 1007 1008 static void __mptcp_clean_una(struct sock *sk) 1009 { 1010 struct mptcp_sock *msk = mptcp_sk(sk); 1011 struct mptcp_data_frag *dtmp, *dfrag; 1012 u64 snd_una; 1013 1014 snd_una = msk->snd_una; 1015 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 1016 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 1017 break; 1018 1019 if (unlikely(dfrag == msk->first_pending)) { 1020 /* in recovery mode can see ack after the current snd head */ 1021 if (WARN_ON_ONCE(!msk->recovery)) 1022 break; 1023 1024 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1025 } 1026 1027 dfrag_clear(sk, dfrag); 1028 } 1029 1030 dfrag = mptcp_rtx_head(sk); 1031 if (dfrag && after64(snd_una, dfrag->data_seq)) { 1032 u64 delta = snd_una - dfrag->data_seq; 1033 1034 /* prevent wrap around in recovery mode */ 1035 if (unlikely(delta > dfrag->already_sent)) { 1036 if (WARN_ON_ONCE(!msk->recovery)) 1037 goto out; 1038 if (WARN_ON_ONCE(delta > dfrag->data_len)) 1039 goto out; 1040 dfrag->already_sent += delta - dfrag->already_sent; 1041 } 1042 1043 dfrag->data_seq += delta; 1044 dfrag->offset += delta; 1045 dfrag->data_len -= delta; 1046 dfrag->already_sent -= delta; 1047 1048 dfrag_uncharge(sk, delta); 1049 } 1050 1051 /* all retransmitted data acked, recovery completed */ 1052 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 1053 msk->recovery = false; 1054 1055 out: 1056 if (snd_una == READ_ONCE(msk->snd_nxt) && 1057 snd_una == READ_ONCE(msk->write_seq)) { 1058 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 1059 mptcp_stop_rtx_timer(sk); 1060 } else { 1061 mptcp_reset_rtx_timer(sk); 1062 } 1063 } 1064 1065 static void __mptcp_clean_una_wakeup(struct sock *sk) 1066 { 1067 lockdep_assert_held_once(&sk->sk_lock.slock); 1068 1069 __mptcp_clean_una(sk); 1070 mptcp_write_space(sk); 1071 } 1072 1073 static void mptcp_clean_una_wakeup(struct sock *sk) 1074 { 1075 mptcp_data_lock(sk); 1076 __mptcp_clean_una_wakeup(sk); 1077 mptcp_data_unlock(sk); 1078 } 1079 1080 static void mptcp_enter_memory_pressure(struct sock *sk) 1081 { 1082 struct mptcp_subflow_context *subflow; 1083 struct mptcp_sock *msk = mptcp_sk(sk); 1084 bool first = true; 1085 1086 mptcp_for_each_subflow(msk, subflow) { 1087 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1088 1089 if (first) 1090 tcp_enter_memory_pressure(ssk); 1091 sk_stream_moderate_sndbuf(ssk); 1092 1093 first = false; 1094 } 1095 __mptcp_sync_sndbuf(sk); 1096 } 1097 1098 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1099 * data 1100 */ 1101 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1102 { 1103 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1104 pfrag, sk->sk_allocation))) 1105 return true; 1106 1107 mptcp_enter_memory_pressure(sk); 1108 return false; 1109 } 1110 1111 static struct mptcp_data_frag * 1112 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1113 int orig_offset) 1114 { 1115 int offset = ALIGN(orig_offset, sizeof(long)); 1116 struct mptcp_data_frag *dfrag; 1117 1118 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1119 dfrag->data_len = 0; 1120 dfrag->data_seq = msk->write_seq; 1121 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1122 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1123 dfrag->already_sent = 0; 1124 dfrag->page = pfrag->page; 1125 1126 return dfrag; 1127 } 1128 1129 struct mptcp_sendmsg_info { 1130 int mss_now; 1131 int size_goal; 1132 u16 limit; 1133 u16 sent; 1134 unsigned int flags; 1135 bool data_lock_held; 1136 }; 1137 1138 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk, 1139 u64 data_seq, int avail_size) 1140 { 1141 u64 window_end = mptcp_wnd_end(msk); 1142 u64 mptcp_snd_wnd; 1143 1144 if (__mptcp_check_fallback(msk)) 1145 return avail_size; 1146 1147 mptcp_snd_wnd = window_end - data_seq; 1148 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size); 1149 1150 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) { 1151 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd); 1152 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED); 1153 } 1154 1155 return avail_size; 1156 } 1157 1158 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1159 { 1160 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1161 1162 if (!mpext) 1163 return false; 1164 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1165 return true; 1166 } 1167 1168 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1169 { 1170 struct sk_buff *skb; 1171 1172 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1173 if (likely(skb)) { 1174 if (likely(__mptcp_add_ext(skb, gfp))) { 1175 skb_reserve(skb, MAX_TCP_HEADER); 1176 skb->ip_summed = CHECKSUM_PARTIAL; 1177 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1178 return skb; 1179 } 1180 __kfree_skb(skb); 1181 } else { 1182 mptcp_enter_memory_pressure(sk); 1183 } 1184 return NULL; 1185 } 1186 1187 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1188 { 1189 struct sk_buff *skb; 1190 1191 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1192 if (!skb) 1193 return NULL; 1194 1195 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1196 tcp_skb_entail(ssk, skb); 1197 return skb; 1198 } 1199 tcp_skb_tsorted_anchor_cleanup(skb); 1200 kfree_skb(skb); 1201 return NULL; 1202 } 1203 1204 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1205 { 1206 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1207 1208 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1209 } 1210 1211 /* note: this always recompute the csum on the whole skb, even 1212 * if we just appended a single frag. More status info needed 1213 */ 1214 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1215 { 1216 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1217 __wsum csum = ~csum_unfold(mpext->csum); 1218 int offset = skb->len - added; 1219 1220 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1221 } 1222 1223 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1224 struct sock *ssk, 1225 struct mptcp_ext *mpext) 1226 { 1227 if (!mpext) 1228 return; 1229 1230 mpext->infinite_map = 1; 1231 mpext->data_len = 0; 1232 1233 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX); 1234 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1235 pr_fallback(msk); 1236 mptcp_do_fallback(ssk); 1237 } 1238 1239 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1)) 1240 1241 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1242 struct mptcp_data_frag *dfrag, 1243 struct mptcp_sendmsg_info *info) 1244 { 1245 u64 data_seq = dfrag->data_seq + info->sent; 1246 int offset = dfrag->offset + info->sent; 1247 struct mptcp_sock *msk = mptcp_sk(sk); 1248 bool zero_window_probe = false; 1249 struct mptcp_ext *mpext = NULL; 1250 bool can_coalesce = false; 1251 bool reuse_skb = true; 1252 struct sk_buff *skb; 1253 size_t copy; 1254 int i; 1255 1256 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n", 1257 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1258 1259 if (WARN_ON_ONCE(info->sent > info->limit || 1260 info->limit > dfrag->data_len)) 1261 return 0; 1262 1263 if (unlikely(!__tcp_can_send(ssk))) 1264 return -EAGAIN; 1265 1266 /* compute send limit */ 1267 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE)) 1268 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE; 1269 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1270 copy = info->size_goal; 1271 1272 skb = tcp_write_queue_tail(ssk); 1273 if (skb && copy > skb->len) { 1274 /* Limit the write to the size available in the 1275 * current skb, if any, so that we create at most a new skb. 1276 * Explicitly tells TCP internals to avoid collapsing on later 1277 * queue management operation, to avoid breaking the ext <-> 1278 * SSN association set here 1279 */ 1280 mpext = skb_ext_find(skb, SKB_EXT_MPTCP); 1281 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1282 TCP_SKB_CB(skb)->eor = 1; 1283 tcp_mark_push(tcp_sk(ssk), skb); 1284 goto alloc_skb; 1285 } 1286 1287 i = skb_shinfo(skb)->nr_frags; 1288 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1289 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) { 1290 tcp_mark_push(tcp_sk(ssk), skb); 1291 goto alloc_skb; 1292 } 1293 1294 copy -= skb->len; 1295 } else { 1296 alloc_skb: 1297 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1298 if (!skb) 1299 return -ENOMEM; 1300 1301 i = skb_shinfo(skb)->nr_frags; 1302 reuse_skb = false; 1303 mpext = skb_ext_find(skb, SKB_EXT_MPTCP); 1304 } 1305 1306 /* Zero window and all data acked? Probe. */ 1307 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy); 1308 if (copy == 0) { 1309 u64 snd_una = READ_ONCE(msk->snd_una); 1310 1311 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) { 1312 tcp_remove_empty_skb(ssk); 1313 return 0; 1314 } 1315 1316 zero_window_probe = true; 1317 data_seq = snd_una - 1; 1318 copy = 1; 1319 } 1320 1321 copy = min_t(size_t, copy, info->limit - info->sent); 1322 if (!sk_wmem_schedule(ssk, copy)) { 1323 tcp_remove_empty_skb(ssk); 1324 return -ENOMEM; 1325 } 1326 1327 if (can_coalesce) { 1328 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1329 } else { 1330 get_page(dfrag->page); 1331 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1332 } 1333 1334 skb->len += copy; 1335 skb->data_len += copy; 1336 skb->truesize += copy; 1337 sk_wmem_queued_add(ssk, copy); 1338 sk_mem_charge(ssk, copy); 1339 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1340 TCP_SKB_CB(skb)->end_seq += copy; 1341 tcp_skb_pcount_set(skb, 0); 1342 1343 /* on skb reuse we just need to update the DSS len */ 1344 if (reuse_skb) { 1345 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1346 mpext->data_len += copy; 1347 goto out; 1348 } 1349 1350 memset(mpext, 0, sizeof(*mpext)); 1351 mpext->data_seq = data_seq; 1352 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1353 mpext->data_len = copy; 1354 mpext->use_map = 1; 1355 mpext->dsn64 = 1; 1356 1357 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n", 1358 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1359 mpext->dsn64); 1360 1361 if (zero_window_probe) { 1362 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1363 mpext->frozen = 1; 1364 if (READ_ONCE(msk->csum_enabled)) 1365 mptcp_update_data_checksum(skb, copy); 1366 tcp_push_pending_frames(ssk); 1367 return 0; 1368 } 1369 out: 1370 if (READ_ONCE(msk->csum_enabled)) 1371 mptcp_update_data_checksum(skb, copy); 1372 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1373 mptcp_update_infinite_map(msk, ssk, mpext); 1374 trace_mptcp_sendmsg_frag(mpext); 1375 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1376 return copy; 1377 } 1378 1379 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1380 sizeof(struct tcphdr) - \ 1381 MAX_TCP_OPTION_SPACE - \ 1382 sizeof(struct ipv6hdr) - \ 1383 sizeof(struct frag_hdr)) 1384 1385 struct subflow_send_info { 1386 struct sock *ssk; 1387 u64 linger_time; 1388 }; 1389 1390 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1391 { 1392 if (!subflow->stale) 1393 return; 1394 1395 subflow->stale = 0; 1396 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1397 } 1398 1399 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1400 { 1401 if (unlikely(subflow->stale)) { 1402 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1403 1404 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1405 return false; 1406 1407 mptcp_subflow_set_active(subflow); 1408 } 1409 return __mptcp_subflow_active(subflow); 1410 } 1411 1412 #define SSK_MODE_ACTIVE 0 1413 #define SSK_MODE_BACKUP 1 1414 #define SSK_MODE_MAX 2 1415 1416 /* implement the mptcp packet scheduler; 1417 * returns the subflow that will transmit the next DSS 1418 * additionally updates the rtx timeout 1419 */ 1420 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1421 { 1422 struct subflow_send_info send_info[SSK_MODE_MAX]; 1423 struct mptcp_subflow_context *subflow; 1424 struct sock *sk = (struct sock *)msk; 1425 u32 pace, burst, wmem; 1426 int i, nr_active = 0; 1427 struct sock *ssk; 1428 u64 linger_time; 1429 long tout = 0; 1430 1431 /* pick the subflow with the lower wmem/wspace ratio */ 1432 for (i = 0; i < SSK_MODE_MAX; ++i) { 1433 send_info[i].ssk = NULL; 1434 send_info[i].linger_time = -1; 1435 } 1436 1437 mptcp_for_each_subflow(msk, subflow) { 1438 bool backup = subflow->backup || subflow->request_bkup; 1439 1440 trace_mptcp_subflow_get_send(subflow); 1441 ssk = mptcp_subflow_tcp_sock(subflow); 1442 if (!mptcp_subflow_active(subflow)) 1443 continue; 1444 1445 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1446 nr_active += !backup; 1447 pace = subflow->avg_pacing_rate; 1448 if (unlikely(!pace)) { 1449 /* init pacing rate from socket */ 1450 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1451 pace = subflow->avg_pacing_rate; 1452 if (!pace) 1453 continue; 1454 } 1455 1456 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1457 if (linger_time < send_info[backup].linger_time) { 1458 send_info[backup].ssk = ssk; 1459 send_info[backup].linger_time = linger_time; 1460 } 1461 } 1462 __mptcp_set_timeout(sk, tout); 1463 1464 /* pick the best backup if no other subflow is active */ 1465 if (!nr_active) 1466 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1467 1468 /* According to the blest algorithm, to avoid HoL blocking for the 1469 * faster flow, we need to: 1470 * - estimate the faster flow linger time 1471 * - use the above to estimate the amount of byte transferred 1472 * by the faster flow 1473 * - check that the amount of queued data is greter than the above, 1474 * otherwise do not use the picked, slower, subflow 1475 * We select the subflow with the shorter estimated time to flush 1476 * the queued mem, which basically ensure the above. We just need 1477 * to check that subflow has a non empty cwin. 1478 */ 1479 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1480 if (!ssk || !sk_stream_memory_free(ssk)) 1481 return NULL; 1482 1483 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt); 1484 wmem = READ_ONCE(ssk->sk_wmem_queued); 1485 if (!burst) 1486 return ssk; 1487 1488 subflow = mptcp_subflow_ctx(ssk); 1489 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1490 READ_ONCE(ssk->sk_pacing_rate) * burst, 1491 burst + wmem); 1492 msk->snd_burst = burst; 1493 return ssk; 1494 } 1495 1496 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1497 { 1498 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1499 release_sock(ssk); 1500 } 1501 1502 static void mptcp_update_post_push(struct mptcp_sock *msk, 1503 struct mptcp_data_frag *dfrag, 1504 u32 sent) 1505 { 1506 u64 snd_nxt_new = dfrag->data_seq; 1507 1508 dfrag->already_sent += sent; 1509 1510 msk->snd_burst -= sent; 1511 1512 snd_nxt_new += dfrag->already_sent; 1513 1514 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1515 * is recovering after a failover. In that event, this re-sends 1516 * old segments. 1517 * 1518 * Thus compute snd_nxt_new candidate based on 1519 * the dfrag->data_seq that was sent and the data 1520 * that has been handed to the subflow for transmission 1521 * and skip update in case it was old dfrag. 1522 */ 1523 if (likely(after64(snd_nxt_new, msk->snd_nxt))) { 1524 msk->bytes_sent += snd_nxt_new - msk->snd_nxt; 1525 msk->snd_nxt = snd_nxt_new; 1526 } 1527 } 1528 1529 void mptcp_check_and_set_pending(struct sock *sk) 1530 { 1531 if (mptcp_send_head(sk)) { 1532 mptcp_data_lock(sk); 1533 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING); 1534 mptcp_data_unlock(sk); 1535 } 1536 } 1537 1538 static int __subflow_push_pending(struct sock *sk, struct sock *ssk, 1539 struct mptcp_sendmsg_info *info) 1540 { 1541 struct mptcp_sock *msk = mptcp_sk(sk); 1542 struct mptcp_data_frag *dfrag; 1543 int len, copied = 0, err = 0; 1544 1545 while ((dfrag = mptcp_send_head(sk))) { 1546 info->sent = dfrag->already_sent; 1547 info->limit = dfrag->data_len; 1548 len = dfrag->data_len - dfrag->already_sent; 1549 while (len > 0) { 1550 int ret = 0; 1551 1552 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info); 1553 if (ret <= 0) { 1554 err = copied ? : ret; 1555 goto out; 1556 } 1557 1558 info->sent += ret; 1559 copied += ret; 1560 len -= ret; 1561 1562 mptcp_update_post_push(msk, dfrag, ret); 1563 } 1564 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1565 1566 if (msk->snd_burst <= 0 || 1567 !sk_stream_memory_free(ssk) || 1568 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) { 1569 err = copied; 1570 goto out; 1571 } 1572 mptcp_set_timeout(sk); 1573 } 1574 err = copied; 1575 1576 out: 1577 return err; 1578 } 1579 1580 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1581 { 1582 struct sock *prev_ssk = NULL, *ssk = NULL; 1583 struct mptcp_sock *msk = mptcp_sk(sk); 1584 struct mptcp_sendmsg_info info = { 1585 .flags = flags, 1586 }; 1587 bool do_check_data_fin = false; 1588 int push_count = 1; 1589 1590 while (mptcp_send_head(sk) && (push_count > 0)) { 1591 struct mptcp_subflow_context *subflow; 1592 int ret = 0; 1593 1594 if (mptcp_sched_get_send(msk)) 1595 break; 1596 1597 push_count = 0; 1598 1599 mptcp_for_each_subflow(msk, subflow) { 1600 if (READ_ONCE(subflow->scheduled)) { 1601 mptcp_subflow_set_scheduled(subflow, false); 1602 1603 prev_ssk = ssk; 1604 ssk = mptcp_subflow_tcp_sock(subflow); 1605 if (ssk != prev_ssk) { 1606 /* First check. If the ssk has changed since 1607 * the last round, release prev_ssk 1608 */ 1609 if (prev_ssk) 1610 mptcp_push_release(prev_ssk, &info); 1611 1612 /* Need to lock the new subflow only if different 1613 * from the previous one, otherwise we are still 1614 * helding the relevant lock 1615 */ 1616 lock_sock(ssk); 1617 } 1618 1619 push_count++; 1620 1621 ret = __subflow_push_pending(sk, ssk, &info); 1622 if (ret <= 0) { 1623 if (ret != -EAGAIN || 1624 (1 << ssk->sk_state) & 1625 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE)) 1626 push_count--; 1627 continue; 1628 } 1629 do_check_data_fin = true; 1630 } 1631 } 1632 } 1633 1634 /* at this point we held the socket lock for the last subflow we used */ 1635 if (ssk) 1636 mptcp_push_release(ssk, &info); 1637 1638 /* ensure the rtx timer is running */ 1639 if (!mptcp_rtx_timer_pending(sk)) 1640 mptcp_reset_rtx_timer(sk); 1641 if (do_check_data_fin) 1642 mptcp_check_send_data_fin(sk); 1643 } 1644 1645 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first) 1646 { 1647 struct mptcp_sock *msk = mptcp_sk(sk); 1648 struct mptcp_sendmsg_info info = { 1649 .data_lock_held = true, 1650 }; 1651 bool keep_pushing = true; 1652 struct sock *xmit_ssk; 1653 int copied = 0; 1654 1655 info.flags = 0; 1656 while (mptcp_send_head(sk) && keep_pushing) { 1657 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1658 int ret = 0; 1659 1660 /* check for a different subflow usage only after 1661 * spooling the first chunk of data 1662 */ 1663 if (first) { 1664 mptcp_subflow_set_scheduled(subflow, false); 1665 ret = __subflow_push_pending(sk, ssk, &info); 1666 first = false; 1667 if (ret <= 0) 1668 break; 1669 copied += ret; 1670 continue; 1671 } 1672 1673 if (mptcp_sched_get_send(msk)) 1674 goto out; 1675 1676 if (READ_ONCE(subflow->scheduled)) { 1677 mptcp_subflow_set_scheduled(subflow, false); 1678 ret = __subflow_push_pending(sk, ssk, &info); 1679 if (ret <= 0) 1680 keep_pushing = false; 1681 copied += ret; 1682 } 1683 1684 mptcp_for_each_subflow(msk, subflow) { 1685 if (READ_ONCE(subflow->scheduled)) { 1686 xmit_ssk = mptcp_subflow_tcp_sock(subflow); 1687 if (xmit_ssk != ssk) { 1688 mptcp_subflow_delegate(subflow, 1689 MPTCP_DELEGATE_SEND); 1690 keep_pushing = false; 1691 } 1692 } 1693 } 1694 } 1695 1696 out: 1697 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1698 * not going to flush it via release_sock() 1699 */ 1700 if (copied) { 1701 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1702 info.size_goal); 1703 if (!mptcp_rtx_timer_pending(sk)) 1704 mptcp_reset_rtx_timer(sk); 1705 1706 if (msk->snd_data_fin_enable && 1707 msk->snd_nxt + 1 == msk->write_seq) 1708 mptcp_schedule_work(sk); 1709 } 1710 } 1711 1712 static void mptcp_set_nospace(struct sock *sk) 1713 { 1714 /* enable autotune */ 1715 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1716 1717 /* will be cleared on avail space */ 1718 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags); 1719 } 1720 1721 static int mptcp_disconnect(struct sock *sk, int flags); 1722 1723 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1724 size_t len, int *copied_syn) 1725 { 1726 unsigned int saved_flags = msg->msg_flags; 1727 struct mptcp_sock *msk = mptcp_sk(sk); 1728 struct sock *ssk; 1729 int ret; 1730 1731 /* on flags based fastopen the mptcp is supposed to create the 1732 * first subflow right now. Otherwise we are in the defer_connect 1733 * path, and the first subflow must be already present. 1734 * Since the defer_connect flag is cleared after the first succsful 1735 * fastopen attempt, no need to check for additional subflow status. 1736 */ 1737 if (msg->msg_flags & MSG_FASTOPEN) { 1738 ssk = __mptcp_nmpc_sk(msk); 1739 if (IS_ERR(ssk)) 1740 return PTR_ERR(ssk); 1741 } 1742 if (!msk->first) 1743 return -EINVAL; 1744 1745 ssk = msk->first; 1746 1747 lock_sock(ssk); 1748 msg->msg_flags |= MSG_DONTWAIT; 1749 msk->fastopening = 1; 1750 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL); 1751 msk->fastopening = 0; 1752 msg->msg_flags = saved_flags; 1753 release_sock(ssk); 1754 1755 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */ 1756 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) { 1757 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1758 msg->msg_namelen, msg->msg_flags, 1); 1759 1760 /* Keep the same behaviour of plain TCP: zero the copied bytes in 1761 * case of any error, except timeout or signal 1762 */ 1763 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR) 1764 *copied_syn = 0; 1765 } else if (ret && ret != -EINPROGRESS) { 1766 /* The disconnect() op called by tcp_sendmsg_fastopen()/ 1767 * __inet_stream_connect() can fail, due to looking check, 1768 * see mptcp_disconnect(). 1769 * Attempt it again outside the problematic scope. 1770 */ 1771 if (!mptcp_disconnect(sk, 0)) 1772 sk->sk_socket->state = SS_UNCONNECTED; 1773 } 1774 inet_clear_bit(DEFER_CONNECT, sk); 1775 1776 return ret; 1777 } 1778 1779 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1780 { 1781 struct mptcp_sock *msk = mptcp_sk(sk); 1782 struct page_frag *pfrag; 1783 size_t copied = 0; 1784 int ret = 0; 1785 long timeo; 1786 1787 /* silently ignore everything else */ 1788 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN; 1789 1790 lock_sock(sk); 1791 1792 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) || 1793 msg->msg_flags & MSG_FASTOPEN)) { 1794 int copied_syn = 0; 1795 1796 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn); 1797 copied += copied_syn; 1798 if (ret == -EINPROGRESS && copied_syn > 0) 1799 goto out; 1800 else if (ret) 1801 goto do_error; 1802 } 1803 1804 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1805 1806 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1807 ret = sk_stream_wait_connect(sk, &timeo); 1808 if (ret) 1809 goto do_error; 1810 } 1811 1812 ret = -EPIPE; 1813 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))) 1814 goto do_error; 1815 1816 pfrag = sk_page_frag(sk); 1817 1818 while (msg_data_left(msg)) { 1819 int total_ts, frag_truesize = 0; 1820 struct mptcp_data_frag *dfrag; 1821 bool dfrag_collapsed; 1822 size_t psize, offset; 1823 1824 /* reuse tail pfrag, if possible, or carve a new one from the 1825 * page allocator 1826 */ 1827 dfrag = mptcp_pending_tail(sk); 1828 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1829 if (!dfrag_collapsed) { 1830 if (!sk_stream_memory_free(sk)) 1831 goto wait_for_memory; 1832 1833 if (!mptcp_page_frag_refill(sk, pfrag)) 1834 goto wait_for_memory; 1835 1836 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1837 frag_truesize = dfrag->overhead; 1838 } 1839 1840 /* we do not bound vs wspace, to allow a single packet. 1841 * memory accounting will prevent execessive memory usage 1842 * anyway 1843 */ 1844 offset = dfrag->offset + dfrag->data_len; 1845 psize = pfrag->size - offset; 1846 psize = min_t(size_t, psize, msg_data_left(msg)); 1847 total_ts = psize + frag_truesize; 1848 1849 if (!sk_wmem_schedule(sk, total_ts)) 1850 goto wait_for_memory; 1851 1852 if (copy_page_from_iter(dfrag->page, offset, psize, 1853 &msg->msg_iter) != psize) { 1854 ret = -EFAULT; 1855 goto do_error; 1856 } 1857 1858 /* data successfully copied into the write queue */ 1859 sk_forward_alloc_add(sk, -total_ts); 1860 copied += psize; 1861 dfrag->data_len += psize; 1862 frag_truesize += psize; 1863 pfrag->offset += frag_truesize; 1864 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1865 1866 /* charge data on mptcp pending queue to the msk socket 1867 * Note: we charge such data both to sk and ssk 1868 */ 1869 sk_wmem_queued_add(sk, frag_truesize); 1870 if (!dfrag_collapsed) { 1871 get_page(dfrag->page); 1872 list_add_tail(&dfrag->list, &msk->rtx_queue); 1873 if (!msk->first_pending) 1874 WRITE_ONCE(msk->first_pending, dfrag); 1875 } 1876 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk, 1877 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1878 !dfrag_collapsed); 1879 1880 continue; 1881 1882 wait_for_memory: 1883 mptcp_set_nospace(sk); 1884 __mptcp_push_pending(sk, msg->msg_flags); 1885 ret = sk_stream_wait_memory(sk, &timeo); 1886 if (ret) 1887 goto do_error; 1888 } 1889 1890 if (copied) 1891 __mptcp_push_pending(sk, msg->msg_flags); 1892 1893 out: 1894 release_sock(sk); 1895 return copied; 1896 1897 do_error: 1898 if (copied) 1899 goto out; 1900 1901 copied = sk_stream_error(sk, msg->msg_flags, ret); 1902 goto out; 1903 } 1904 1905 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk, 1906 struct msghdr *msg, 1907 size_t len, int flags, 1908 struct scm_timestamping_internal *tss, 1909 int *cmsg_flags) 1910 { 1911 struct sk_buff *skb, *tmp; 1912 int copied = 0; 1913 1914 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) { 1915 u32 offset = MPTCP_SKB_CB(skb)->offset; 1916 u32 data_len = skb->len - offset; 1917 u32 count = min_t(size_t, len - copied, data_len); 1918 int err; 1919 1920 if (!(flags & MSG_TRUNC)) { 1921 err = skb_copy_datagram_msg(skb, offset, msg, count); 1922 if (unlikely(err < 0)) { 1923 if (!copied) 1924 return err; 1925 break; 1926 } 1927 } 1928 1929 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 1930 tcp_update_recv_tstamps(skb, tss); 1931 *cmsg_flags |= MPTCP_CMSG_TS; 1932 } 1933 1934 copied += count; 1935 1936 if (count < data_len) { 1937 if (!(flags & MSG_PEEK)) { 1938 MPTCP_SKB_CB(skb)->offset += count; 1939 MPTCP_SKB_CB(skb)->map_seq += count; 1940 msk->bytes_consumed += count; 1941 } 1942 break; 1943 } 1944 1945 if (!(flags & MSG_PEEK)) { 1946 /* we will bulk release the skb memory later */ 1947 skb->destructor = NULL; 1948 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize); 1949 __skb_unlink(skb, &msk->receive_queue); 1950 __kfree_skb(skb); 1951 msk->bytes_consumed += count; 1952 } 1953 1954 if (copied >= len) 1955 break; 1956 } 1957 1958 return copied; 1959 } 1960 1961 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1962 * 1963 * Only difference: Use highest rtt estimate of the subflows in use. 1964 */ 1965 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1966 { 1967 struct mptcp_subflow_context *subflow; 1968 struct sock *sk = (struct sock *)msk; 1969 u8 scaling_ratio = U8_MAX; 1970 u32 time, advmss = 1; 1971 u64 rtt_us, mstamp; 1972 1973 msk_owned_by_me(msk); 1974 1975 if (copied <= 0) 1976 return; 1977 1978 if (!msk->rcvspace_init) 1979 mptcp_rcv_space_init(msk, msk->first); 1980 1981 msk->rcvq_space.copied += copied; 1982 1983 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1984 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1985 1986 rtt_us = msk->rcvq_space.rtt_us; 1987 if (rtt_us && time < (rtt_us >> 3)) 1988 return; 1989 1990 rtt_us = 0; 1991 mptcp_for_each_subflow(msk, subflow) { 1992 const struct tcp_sock *tp; 1993 u64 sf_rtt_us; 1994 u32 sf_advmss; 1995 1996 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1997 1998 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1999 sf_advmss = READ_ONCE(tp->advmss); 2000 2001 rtt_us = max(sf_rtt_us, rtt_us); 2002 advmss = max(sf_advmss, advmss); 2003 scaling_ratio = min(tp->scaling_ratio, scaling_ratio); 2004 } 2005 2006 msk->rcvq_space.rtt_us = rtt_us; 2007 msk->scaling_ratio = scaling_ratio; 2008 if (time < (rtt_us >> 3) || rtt_us == 0) 2009 return; 2010 2011 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 2012 goto new_measure; 2013 2014 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 2015 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 2016 u64 rcvwin, grow; 2017 int rcvbuf; 2018 2019 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 2020 2021 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 2022 2023 do_div(grow, msk->rcvq_space.space); 2024 rcvwin += (grow << 1); 2025 2026 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin), 2027 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2028 2029 if (rcvbuf > sk->sk_rcvbuf) { 2030 u32 window_clamp; 2031 2032 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf); 2033 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 2034 2035 /* Make subflows follow along. If we do not do this, we 2036 * get drops at subflow level if skbs can't be moved to 2037 * the mptcp rx queue fast enough (announced rcv_win can 2038 * exceed ssk->sk_rcvbuf). 2039 */ 2040 mptcp_for_each_subflow(msk, subflow) { 2041 struct sock *ssk; 2042 bool slow; 2043 2044 ssk = mptcp_subflow_tcp_sock(subflow); 2045 slow = lock_sock_fast(ssk); 2046 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 2047 WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp); 2048 if (tcp_can_send_ack(ssk)) 2049 tcp_cleanup_rbuf(ssk, 1); 2050 unlock_sock_fast(ssk, slow); 2051 } 2052 } 2053 } 2054 2055 msk->rcvq_space.space = msk->rcvq_space.copied; 2056 new_measure: 2057 msk->rcvq_space.copied = 0; 2058 msk->rcvq_space.time = mstamp; 2059 } 2060 2061 static void __mptcp_update_rmem(struct sock *sk) 2062 { 2063 struct mptcp_sock *msk = mptcp_sk(sk); 2064 2065 if (!msk->rmem_released) 2066 return; 2067 2068 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc); 2069 mptcp_rmem_uncharge(sk, msk->rmem_released); 2070 WRITE_ONCE(msk->rmem_released, 0); 2071 } 2072 2073 static void __mptcp_splice_receive_queue(struct sock *sk) 2074 { 2075 struct mptcp_sock *msk = mptcp_sk(sk); 2076 2077 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue); 2078 } 2079 2080 static bool __mptcp_move_skbs(struct mptcp_sock *msk) 2081 { 2082 struct sock *sk = (struct sock *)msk; 2083 unsigned int moved = 0; 2084 bool ret, done; 2085 2086 do { 2087 struct sock *ssk = mptcp_subflow_recv_lookup(msk); 2088 bool slowpath; 2089 2090 /* we can have data pending in the subflows only if the msk 2091 * receive buffer was full at subflow_data_ready() time, 2092 * that is an unlikely slow path. 2093 */ 2094 if (likely(!ssk)) 2095 break; 2096 2097 slowpath = lock_sock_fast(ssk); 2098 mptcp_data_lock(sk); 2099 __mptcp_update_rmem(sk); 2100 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 2101 mptcp_data_unlock(sk); 2102 2103 if (unlikely(ssk->sk_err)) 2104 __mptcp_error_report(sk); 2105 unlock_sock_fast(ssk, slowpath); 2106 } while (!done); 2107 2108 /* acquire the data lock only if some input data is pending */ 2109 ret = moved > 0; 2110 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) || 2111 !skb_queue_empty_lockless(&sk->sk_receive_queue)) { 2112 mptcp_data_lock(sk); 2113 __mptcp_update_rmem(sk); 2114 ret |= __mptcp_ofo_queue(msk); 2115 __mptcp_splice_receive_queue(sk); 2116 mptcp_data_unlock(sk); 2117 } 2118 if (ret) 2119 mptcp_check_data_fin((struct sock *)msk); 2120 return !skb_queue_empty(&msk->receive_queue); 2121 } 2122 2123 static unsigned int mptcp_inq_hint(const struct sock *sk) 2124 { 2125 const struct mptcp_sock *msk = mptcp_sk(sk); 2126 const struct sk_buff *skb; 2127 2128 skb = skb_peek(&msk->receive_queue); 2129 if (skb) { 2130 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 2131 2132 if (hint_val >= INT_MAX) 2133 return INT_MAX; 2134 2135 return (unsigned int)hint_val; 2136 } 2137 2138 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2139 return 1; 2140 2141 return 0; 2142 } 2143 2144 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2145 int flags, int *addr_len) 2146 { 2147 struct mptcp_sock *msk = mptcp_sk(sk); 2148 struct scm_timestamping_internal tss; 2149 int copied = 0, cmsg_flags = 0; 2150 int target; 2151 long timeo; 2152 2153 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2154 if (unlikely(flags & MSG_ERRQUEUE)) 2155 return inet_recv_error(sk, msg, len, addr_len); 2156 2157 lock_sock(sk); 2158 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2159 copied = -ENOTCONN; 2160 goto out_err; 2161 } 2162 2163 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2164 2165 len = min_t(size_t, len, INT_MAX); 2166 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2167 2168 if (unlikely(msk->recvmsg_inq)) 2169 cmsg_flags = MPTCP_CMSG_INQ; 2170 2171 while (copied < len) { 2172 int err, bytes_read; 2173 2174 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags); 2175 if (unlikely(bytes_read < 0)) { 2176 if (!copied) 2177 copied = bytes_read; 2178 goto out_err; 2179 } 2180 2181 copied += bytes_read; 2182 2183 /* be sure to advertise window change */ 2184 mptcp_cleanup_rbuf(msk); 2185 2186 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk)) 2187 continue; 2188 2189 /* only the master socket status is relevant here. The exit 2190 * conditions mirror closely tcp_recvmsg() 2191 */ 2192 if (copied >= target) 2193 break; 2194 2195 if (copied) { 2196 if (sk->sk_err || 2197 sk->sk_state == TCP_CLOSE || 2198 (sk->sk_shutdown & RCV_SHUTDOWN) || 2199 !timeo || 2200 signal_pending(current)) 2201 break; 2202 } else { 2203 if (sk->sk_err) { 2204 copied = sock_error(sk); 2205 break; 2206 } 2207 2208 if (sk->sk_shutdown & RCV_SHUTDOWN) { 2209 /* race breaker: the shutdown could be after the 2210 * previous receive queue check 2211 */ 2212 if (__mptcp_move_skbs(msk)) 2213 continue; 2214 break; 2215 } 2216 2217 if (sk->sk_state == TCP_CLOSE) { 2218 copied = -ENOTCONN; 2219 break; 2220 } 2221 2222 if (!timeo) { 2223 copied = -EAGAIN; 2224 break; 2225 } 2226 2227 if (signal_pending(current)) { 2228 copied = sock_intr_errno(timeo); 2229 break; 2230 } 2231 } 2232 2233 pr_debug("block timeout %ld\n", timeo); 2234 mptcp_rcv_space_adjust(msk, copied); 2235 err = sk_wait_data(sk, &timeo, NULL); 2236 if (err < 0) { 2237 err = copied ? : err; 2238 goto out_err; 2239 } 2240 } 2241 2242 mptcp_rcv_space_adjust(msk, copied); 2243 2244 out_err: 2245 if (cmsg_flags && copied >= 0) { 2246 if (cmsg_flags & MPTCP_CMSG_TS) 2247 tcp_recv_timestamp(msg, sk, &tss); 2248 2249 if (cmsg_flags & MPTCP_CMSG_INQ) { 2250 unsigned int inq = mptcp_inq_hint(sk); 2251 2252 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2253 } 2254 } 2255 2256 pr_debug("msk=%p rx queue empty=%d:%d copied=%d\n", 2257 msk, skb_queue_empty_lockless(&sk->sk_receive_queue), 2258 skb_queue_empty(&msk->receive_queue), copied); 2259 2260 release_sock(sk); 2261 return copied; 2262 } 2263 2264 static void mptcp_retransmit_timer(struct timer_list *t) 2265 { 2266 struct inet_connection_sock *icsk = from_timer(icsk, t, 2267 icsk_retransmit_timer); 2268 struct sock *sk = &icsk->icsk_inet.sk; 2269 struct mptcp_sock *msk = mptcp_sk(sk); 2270 2271 bh_lock_sock(sk); 2272 if (!sock_owned_by_user(sk)) { 2273 /* we need a process context to retransmit */ 2274 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2275 mptcp_schedule_work(sk); 2276 } else { 2277 /* delegate our work to tcp_release_cb() */ 2278 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2279 } 2280 bh_unlock_sock(sk); 2281 sock_put(sk); 2282 } 2283 2284 static void mptcp_tout_timer(struct timer_list *t) 2285 { 2286 struct sock *sk = from_timer(sk, t, sk_timer); 2287 2288 mptcp_schedule_work(sk); 2289 sock_put(sk); 2290 } 2291 2292 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2293 * level. 2294 * 2295 * A backup subflow is returned only if that is the only kind available. 2296 */ 2297 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2298 { 2299 struct sock *backup = NULL, *pick = NULL; 2300 struct mptcp_subflow_context *subflow; 2301 int min_stale_count = INT_MAX; 2302 2303 mptcp_for_each_subflow(msk, subflow) { 2304 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2305 2306 if (!__mptcp_subflow_active(subflow)) 2307 continue; 2308 2309 /* still data outstanding at TCP level? skip this */ 2310 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2311 mptcp_pm_subflow_chk_stale(msk, ssk); 2312 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2313 continue; 2314 } 2315 2316 if (subflow->backup || subflow->request_bkup) { 2317 if (!backup) 2318 backup = ssk; 2319 continue; 2320 } 2321 2322 if (!pick) 2323 pick = ssk; 2324 } 2325 2326 if (pick) 2327 return pick; 2328 2329 /* use backup only if there are no progresses anywhere */ 2330 return min_stale_count > 1 ? backup : NULL; 2331 } 2332 2333 bool __mptcp_retransmit_pending_data(struct sock *sk) 2334 { 2335 struct mptcp_data_frag *cur, *rtx_head; 2336 struct mptcp_sock *msk = mptcp_sk(sk); 2337 2338 if (__mptcp_check_fallback(msk)) 2339 return false; 2340 2341 /* the closing socket has some data untransmitted and/or unacked: 2342 * some data in the mptcp rtx queue has not really xmitted yet. 2343 * keep it simple and re-inject the whole mptcp level rtx queue 2344 */ 2345 mptcp_data_lock(sk); 2346 __mptcp_clean_una_wakeup(sk); 2347 rtx_head = mptcp_rtx_head(sk); 2348 if (!rtx_head) { 2349 mptcp_data_unlock(sk); 2350 return false; 2351 } 2352 2353 msk->recovery_snd_nxt = msk->snd_nxt; 2354 msk->recovery = true; 2355 mptcp_data_unlock(sk); 2356 2357 msk->first_pending = rtx_head; 2358 msk->snd_burst = 0; 2359 2360 /* be sure to clear the "sent status" on all re-injected fragments */ 2361 list_for_each_entry(cur, &msk->rtx_queue, list) { 2362 if (!cur->already_sent) 2363 break; 2364 cur->already_sent = 0; 2365 } 2366 2367 return true; 2368 } 2369 2370 /* flags for __mptcp_close_ssk() */ 2371 #define MPTCP_CF_PUSH BIT(1) 2372 #define MPTCP_CF_FASTCLOSE BIT(2) 2373 2374 /* be sure to send a reset only if the caller asked for it, also 2375 * clean completely the subflow status when the subflow reaches 2376 * TCP_CLOSE state 2377 */ 2378 static void __mptcp_subflow_disconnect(struct sock *ssk, 2379 struct mptcp_subflow_context *subflow, 2380 unsigned int flags) 2381 { 2382 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 2383 (flags & MPTCP_CF_FASTCLOSE)) { 2384 /* The MPTCP code never wait on the subflow sockets, TCP-level 2385 * disconnect should never fail 2386 */ 2387 WARN_ON_ONCE(tcp_disconnect(ssk, 0)); 2388 mptcp_subflow_ctx_reset(subflow); 2389 } else { 2390 tcp_shutdown(ssk, SEND_SHUTDOWN); 2391 } 2392 } 2393 2394 /* subflow sockets can be either outgoing (connect) or incoming 2395 * (accept). 2396 * 2397 * Outgoing subflows use in-kernel sockets. 2398 * Incoming subflows do not have their own 'struct socket' allocated, 2399 * so we need to use tcp_close() after detaching them from the mptcp 2400 * parent socket. 2401 */ 2402 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2403 struct mptcp_subflow_context *subflow, 2404 unsigned int flags) 2405 { 2406 struct mptcp_sock *msk = mptcp_sk(sk); 2407 bool dispose_it, need_push = false; 2408 2409 /* If the first subflow moved to a close state before accept, e.g. due 2410 * to an incoming reset or listener shutdown, the subflow socket is 2411 * already deleted by inet_child_forget() and the mptcp socket can't 2412 * survive too. 2413 */ 2414 if (msk->in_accept_queue && msk->first == ssk && 2415 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) { 2416 /* ensure later check in mptcp_worker() will dispose the msk */ 2417 mptcp_set_close_tout(sk, tcp_jiffies32 - (TCP_TIMEWAIT_LEN + 1)); 2418 sock_set_flag(sk, SOCK_DEAD); 2419 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2420 mptcp_subflow_drop_ctx(ssk); 2421 goto out_release; 2422 } 2423 2424 dispose_it = msk->free_first || ssk != msk->first; 2425 if (dispose_it) 2426 list_del(&subflow->node); 2427 2428 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2429 2430 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) { 2431 /* be sure to force the tcp_close path 2432 * to generate the egress reset 2433 */ 2434 ssk->sk_lingertime = 0; 2435 sock_set_flag(ssk, SOCK_LINGER); 2436 subflow->send_fastclose = 1; 2437 } 2438 2439 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2440 if (!dispose_it) { 2441 __mptcp_subflow_disconnect(ssk, subflow, flags); 2442 release_sock(ssk); 2443 2444 goto out; 2445 } 2446 2447 subflow->disposable = 1; 2448 2449 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2450 * the ssk has been already destroyed, we just need to release the 2451 * reference owned by msk; 2452 */ 2453 if (!inet_csk(ssk)->icsk_ulp_ops) { 2454 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD)); 2455 kfree_rcu(subflow, rcu); 2456 } else { 2457 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2458 __tcp_close(ssk, 0); 2459 2460 /* close acquired an extra ref */ 2461 __sock_put(ssk); 2462 } 2463 2464 out_release: 2465 __mptcp_subflow_error_report(sk, ssk); 2466 release_sock(ssk); 2467 2468 sock_put(ssk); 2469 2470 if (ssk == msk->first) 2471 WRITE_ONCE(msk->first, NULL); 2472 2473 out: 2474 __mptcp_sync_sndbuf(sk); 2475 if (need_push) 2476 __mptcp_push_pending(sk, 0); 2477 2478 /* Catch every 'all subflows closed' scenario, including peers silently 2479 * closing them, e.g. due to timeout. 2480 * For established sockets, allow an additional timeout before closing, 2481 * as the protocol can still create more subflows. 2482 */ 2483 if (list_is_singular(&msk->conn_list) && msk->first && 2484 inet_sk_state_load(msk->first) == TCP_CLOSE) { 2485 if (sk->sk_state != TCP_ESTABLISHED || 2486 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) { 2487 mptcp_set_state(sk, TCP_CLOSE); 2488 mptcp_close_wake_up(sk); 2489 } else { 2490 mptcp_start_tout_timer(sk); 2491 } 2492 } 2493 } 2494 2495 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2496 struct mptcp_subflow_context *subflow) 2497 { 2498 /* The first subflow can already be closed and still in the list */ 2499 if (subflow->close_event_done) 2500 return; 2501 2502 subflow->close_event_done = true; 2503 2504 if (sk->sk_state == TCP_ESTABLISHED) 2505 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2506 2507 /* subflow aborted before reaching the fully_established status 2508 * attempt the creation of the next subflow 2509 */ 2510 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow); 2511 2512 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2513 } 2514 2515 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2516 { 2517 return 0; 2518 } 2519 2520 static void __mptcp_close_subflow(struct sock *sk) 2521 { 2522 struct mptcp_subflow_context *subflow, *tmp; 2523 struct mptcp_sock *msk = mptcp_sk(sk); 2524 2525 might_sleep(); 2526 2527 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2528 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2529 int ssk_state = inet_sk_state_load(ssk); 2530 2531 if (ssk_state != TCP_CLOSE && 2532 (ssk_state != TCP_CLOSE_WAIT || 2533 inet_sk_state_load(sk) != TCP_ESTABLISHED)) 2534 continue; 2535 2536 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2537 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2538 continue; 2539 2540 mptcp_close_ssk(sk, ssk, subflow); 2541 } 2542 2543 } 2544 2545 static bool mptcp_close_tout_expired(const struct sock *sk) 2546 { 2547 if (!inet_csk(sk)->icsk_mtup.probe_timestamp || 2548 sk->sk_state == TCP_CLOSE) 2549 return false; 2550 2551 return time_after32(tcp_jiffies32, 2552 inet_csk(sk)->icsk_mtup.probe_timestamp + TCP_TIMEWAIT_LEN); 2553 } 2554 2555 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2556 { 2557 struct mptcp_subflow_context *subflow, *tmp; 2558 struct sock *sk = (struct sock *)msk; 2559 2560 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2561 return; 2562 2563 mptcp_token_destroy(msk); 2564 2565 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2566 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2567 bool slow; 2568 2569 slow = lock_sock_fast(tcp_sk); 2570 if (tcp_sk->sk_state != TCP_CLOSE) { 2571 tcp_send_active_reset(tcp_sk, GFP_ATOMIC); 2572 tcp_set_state(tcp_sk, TCP_CLOSE); 2573 } 2574 unlock_sock_fast(tcp_sk, slow); 2575 } 2576 2577 /* Mirror the tcp_reset() error propagation */ 2578 switch (sk->sk_state) { 2579 case TCP_SYN_SENT: 2580 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 2581 break; 2582 case TCP_CLOSE_WAIT: 2583 WRITE_ONCE(sk->sk_err, EPIPE); 2584 break; 2585 case TCP_CLOSE: 2586 return; 2587 default: 2588 WRITE_ONCE(sk->sk_err, ECONNRESET); 2589 } 2590 2591 mptcp_set_state(sk, TCP_CLOSE); 2592 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2593 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2594 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2595 2596 /* the calling mptcp_worker will properly destroy the socket */ 2597 if (sock_flag(sk, SOCK_DEAD)) 2598 return; 2599 2600 sk->sk_state_change(sk); 2601 sk_error_report(sk); 2602 } 2603 2604 static void __mptcp_retrans(struct sock *sk) 2605 { 2606 struct mptcp_sock *msk = mptcp_sk(sk); 2607 struct mptcp_subflow_context *subflow; 2608 struct mptcp_sendmsg_info info = {}; 2609 struct mptcp_data_frag *dfrag; 2610 struct sock *ssk; 2611 int ret, err; 2612 u16 len = 0; 2613 2614 mptcp_clean_una_wakeup(sk); 2615 2616 /* first check ssk: need to kick "stale" logic */ 2617 err = mptcp_sched_get_retrans(msk); 2618 dfrag = mptcp_rtx_head(sk); 2619 if (!dfrag) { 2620 if (mptcp_data_fin_enabled(msk)) { 2621 struct inet_connection_sock *icsk = inet_csk(sk); 2622 2623 icsk->icsk_retransmits++; 2624 mptcp_set_datafin_timeout(sk); 2625 mptcp_send_ack(msk); 2626 2627 goto reset_timer; 2628 } 2629 2630 if (!mptcp_send_head(sk)) 2631 return; 2632 2633 goto reset_timer; 2634 } 2635 2636 if (err) 2637 goto reset_timer; 2638 2639 mptcp_for_each_subflow(msk, subflow) { 2640 if (READ_ONCE(subflow->scheduled)) { 2641 u16 copied = 0; 2642 2643 mptcp_subflow_set_scheduled(subflow, false); 2644 2645 ssk = mptcp_subflow_tcp_sock(subflow); 2646 2647 lock_sock(ssk); 2648 2649 /* limit retransmission to the bytes already sent on some subflows */ 2650 info.sent = 0; 2651 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : 2652 dfrag->already_sent; 2653 while (info.sent < info.limit) { 2654 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2655 if (ret <= 0) 2656 break; 2657 2658 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2659 copied += ret; 2660 info.sent += ret; 2661 } 2662 if (copied) { 2663 len = max(copied, len); 2664 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2665 info.size_goal); 2666 WRITE_ONCE(msk->allow_infinite_fallback, false); 2667 } 2668 2669 release_sock(ssk); 2670 } 2671 } 2672 2673 msk->bytes_retrans += len; 2674 dfrag->already_sent = max(dfrag->already_sent, len); 2675 2676 reset_timer: 2677 mptcp_check_and_set_pending(sk); 2678 2679 if (!mptcp_rtx_timer_pending(sk)) 2680 mptcp_reset_rtx_timer(sk); 2681 } 2682 2683 /* schedule the timeout timer for the relevant event: either close timeout 2684 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one 2685 */ 2686 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout) 2687 { 2688 struct sock *sk = (struct sock *)msk; 2689 unsigned long timeout, close_timeout; 2690 2691 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp) 2692 return; 2693 2694 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp - 2695 tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN; 2696 2697 /* the close timeout takes precedence on the fail one, and here at least one of 2698 * them is active 2699 */ 2700 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout; 2701 2702 sk_reset_timer(sk, &sk->sk_timer, timeout); 2703 } 2704 2705 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2706 { 2707 struct sock *ssk = msk->first; 2708 bool slow; 2709 2710 if (!ssk) 2711 return; 2712 2713 pr_debug("MP_FAIL doesn't respond, reset the subflow\n"); 2714 2715 slow = lock_sock_fast(ssk); 2716 mptcp_subflow_reset(ssk); 2717 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0); 2718 unlock_sock_fast(ssk, slow); 2719 } 2720 2721 static void mptcp_do_fastclose(struct sock *sk) 2722 { 2723 struct mptcp_subflow_context *subflow, *tmp; 2724 struct mptcp_sock *msk = mptcp_sk(sk); 2725 2726 mptcp_set_state(sk, TCP_CLOSE); 2727 mptcp_for_each_subflow_safe(msk, subflow, tmp) 2728 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), 2729 subflow, MPTCP_CF_FASTCLOSE); 2730 } 2731 2732 static void mptcp_worker(struct work_struct *work) 2733 { 2734 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2735 struct sock *sk = (struct sock *)msk; 2736 unsigned long fail_tout; 2737 int state; 2738 2739 lock_sock(sk); 2740 state = sk->sk_state; 2741 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN))) 2742 goto unlock; 2743 2744 mptcp_check_fastclose(msk); 2745 2746 mptcp_pm_nl_work(msk); 2747 2748 mptcp_check_send_data_fin(sk); 2749 mptcp_check_data_fin_ack(sk); 2750 mptcp_check_data_fin(sk); 2751 2752 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2753 __mptcp_close_subflow(sk); 2754 2755 if (mptcp_close_tout_expired(sk)) { 2756 mptcp_do_fastclose(sk); 2757 mptcp_close_wake_up(sk); 2758 } 2759 2760 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) { 2761 __mptcp_destroy_sock(sk); 2762 goto unlock; 2763 } 2764 2765 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2766 __mptcp_retrans(sk); 2767 2768 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0; 2769 if (fail_tout && time_after(jiffies, fail_tout)) 2770 mptcp_mp_fail_no_response(msk); 2771 2772 unlock: 2773 release_sock(sk); 2774 sock_put(sk); 2775 } 2776 2777 static void __mptcp_init_sock(struct sock *sk) 2778 { 2779 struct mptcp_sock *msk = mptcp_sk(sk); 2780 2781 INIT_LIST_HEAD(&msk->conn_list); 2782 INIT_LIST_HEAD(&msk->join_list); 2783 INIT_LIST_HEAD(&msk->rtx_queue); 2784 INIT_WORK(&msk->work, mptcp_worker); 2785 __skb_queue_head_init(&msk->receive_queue); 2786 msk->out_of_order_queue = RB_ROOT; 2787 msk->first_pending = NULL; 2788 msk->rmem_fwd_alloc = 0; 2789 WRITE_ONCE(msk->rmem_released, 0); 2790 msk->timer_ival = TCP_RTO_MIN; 2791 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO; 2792 2793 WRITE_ONCE(msk->first, NULL); 2794 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2795 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2796 WRITE_ONCE(msk->allow_infinite_fallback, true); 2797 msk->recovery = false; 2798 msk->subflow_id = 1; 2799 2800 mptcp_pm_data_init(msk); 2801 2802 /* re-use the csk retrans timer for MPTCP-level retrans */ 2803 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2804 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0); 2805 } 2806 2807 static void mptcp_ca_reset(struct sock *sk) 2808 { 2809 struct inet_connection_sock *icsk = inet_csk(sk); 2810 2811 tcp_assign_congestion_control(sk); 2812 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name); 2813 2814 /* no need to keep a reference to the ops, the name will suffice */ 2815 tcp_cleanup_congestion_control(sk); 2816 icsk->icsk_ca_ops = NULL; 2817 } 2818 2819 static int mptcp_init_sock(struct sock *sk) 2820 { 2821 struct net *net = sock_net(sk); 2822 int ret; 2823 2824 __mptcp_init_sock(sk); 2825 2826 if (!mptcp_is_enabled(net)) 2827 return -ENOPROTOOPT; 2828 2829 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2830 return -ENOMEM; 2831 2832 rcu_read_lock(); 2833 ret = mptcp_init_sched(mptcp_sk(sk), 2834 mptcp_sched_find(mptcp_get_scheduler(net))); 2835 rcu_read_unlock(); 2836 if (ret) 2837 return ret; 2838 2839 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); 2840 2841 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2842 * propagate the correct value 2843 */ 2844 mptcp_ca_reset(sk); 2845 2846 sk_sockets_allocated_inc(sk); 2847 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]); 2848 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]); 2849 2850 return 0; 2851 } 2852 2853 static void __mptcp_clear_xmit(struct sock *sk) 2854 { 2855 struct mptcp_sock *msk = mptcp_sk(sk); 2856 struct mptcp_data_frag *dtmp, *dfrag; 2857 2858 WRITE_ONCE(msk->first_pending, NULL); 2859 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2860 dfrag_clear(sk, dfrag); 2861 } 2862 2863 void mptcp_cancel_work(struct sock *sk) 2864 { 2865 struct mptcp_sock *msk = mptcp_sk(sk); 2866 2867 if (cancel_work_sync(&msk->work)) 2868 __sock_put(sk); 2869 } 2870 2871 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2872 { 2873 lock_sock(ssk); 2874 2875 switch (ssk->sk_state) { 2876 case TCP_LISTEN: 2877 if (!(how & RCV_SHUTDOWN)) 2878 break; 2879 fallthrough; 2880 case TCP_SYN_SENT: 2881 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK)); 2882 break; 2883 default: 2884 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2885 pr_debug("Fallback\n"); 2886 ssk->sk_shutdown |= how; 2887 tcp_shutdown(ssk, how); 2888 2889 /* simulate the data_fin ack reception to let the state 2890 * machine move forward 2891 */ 2892 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt); 2893 mptcp_schedule_work(sk); 2894 } else { 2895 pr_debug("Sending DATA_FIN on subflow %p\n", ssk); 2896 tcp_send_ack(ssk); 2897 if (!mptcp_rtx_timer_pending(sk)) 2898 mptcp_reset_rtx_timer(sk); 2899 } 2900 break; 2901 } 2902 2903 release_sock(ssk); 2904 } 2905 2906 void mptcp_set_state(struct sock *sk, int state) 2907 { 2908 int oldstate = sk->sk_state; 2909 2910 switch (state) { 2911 case TCP_ESTABLISHED: 2912 if (oldstate != TCP_ESTABLISHED) 2913 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2914 break; 2915 case TCP_CLOSE_WAIT: 2916 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state: 2917 * MPTCP "accepted" sockets will be created later on. So no 2918 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT. 2919 */ 2920 break; 2921 default: 2922 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 2923 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2924 } 2925 2926 inet_sk_state_store(sk, state); 2927 } 2928 2929 static const unsigned char new_state[16] = { 2930 /* current state: new state: action: */ 2931 [0 /* (Invalid) */] = TCP_CLOSE, 2932 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2933 [TCP_SYN_SENT] = TCP_CLOSE, 2934 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2935 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2936 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2937 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2938 [TCP_CLOSE] = TCP_CLOSE, 2939 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2940 [TCP_LAST_ACK] = TCP_LAST_ACK, 2941 [TCP_LISTEN] = TCP_CLOSE, 2942 [TCP_CLOSING] = TCP_CLOSING, 2943 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2944 }; 2945 2946 static int mptcp_close_state(struct sock *sk) 2947 { 2948 int next = (int)new_state[sk->sk_state]; 2949 int ns = next & TCP_STATE_MASK; 2950 2951 mptcp_set_state(sk, ns); 2952 2953 return next & TCP_ACTION_FIN; 2954 } 2955 2956 static void mptcp_check_send_data_fin(struct sock *sk) 2957 { 2958 struct mptcp_subflow_context *subflow; 2959 struct mptcp_sock *msk = mptcp_sk(sk); 2960 2961 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n", 2962 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2963 msk->snd_nxt, msk->write_seq); 2964 2965 /* we still need to enqueue subflows or not really shutting down, 2966 * skip this 2967 */ 2968 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2969 mptcp_send_head(sk)) 2970 return; 2971 2972 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2973 2974 mptcp_for_each_subflow(msk, subflow) { 2975 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2976 2977 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2978 } 2979 } 2980 2981 static void __mptcp_wr_shutdown(struct sock *sk) 2982 { 2983 struct mptcp_sock *msk = mptcp_sk(sk); 2984 2985 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n", 2986 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2987 !!mptcp_send_head(sk)); 2988 2989 /* will be ignored by fallback sockets */ 2990 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2991 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2992 2993 mptcp_check_send_data_fin(sk); 2994 } 2995 2996 static void __mptcp_destroy_sock(struct sock *sk) 2997 { 2998 struct mptcp_sock *msk = mptcp_sk(sk); 2999 3000 pr_debug("msk=%p\n", msk); 3001 3002 might_sleep(); 3003 3004 mptcp_stop_rtx_timer(sk); 3005 sk_stop_timer(sk, &sk->sk_timer); 3006 msk->pm.status = 0; 3007 mptcp_release_sched(msk); 3008 3009 sk->sk_prot->destroy(sk); 3010 3011 WARN_ON_ONCE(msk->rmem_fwd_alloc); 3012 WARN_ON_ONCE(msk->rmem_released); 3013 sk_stream_kill_queues(sk); 3014 xfrm_sk_free_policy(sk); 3015 3016 sock_put(sk); 3017 } 3018 3019 void __mptcp_unaccepted_force_close(struct sock *sk) 3020 { 3021 sock_set_flag(sk, SOCK_DEAD); 3022 mptcp_do_fastclose(sk); 3023 __mptcp_destroy_sock(sk); 3024 } 3025 3026 static __poll_t mptcp_check_readable(struct sock *sk) 3027 { 3028 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0; 3029 } 3030 3031 static void mptcp_check_listen_stop(struct sock *sk) 3032 { 3033 struct sock *ssk; 3034 3035 if (inet_sk_state_load(sk) != TCP_LISTEN) 3036 return; 3037 3038 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 3039 ssk = mptcp_sk(sk)->first; 3040 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN)) 3041 return; 3042 3043 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 3044 tcp_set_state(ssk, TCP_CLOSE); 3045 mptcp_subflow_queue_clean(sk, ssk); 3046 inet_csk_listen_stop(ssk); 3047 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED); 3048 release_sock(ssk); 3049 } 3050 3051 bool __mptcp_close(struct sock *sk, long timeout) 3052 { 3053 struct mptcp_subflow_context *subflow; 3054 struct mptcp_sock *msk = mptcp_sk(sk); 3055 bool do_cancel_work = false; 3056 int subflows_alive = 0; 3057 3058 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3059 3060 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 3061 mptcp_check_listen_stop(sk); 3062 mptcp_set_state(sk, TCP_CLOSE); 3063 goto cleanup; 3064 } 3065 3066 if (mptcp_data_avail(msk) || timeout < 0) { 3067 /* If the msk has read data, or the caller explicitly ask it, 3068 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose 3069 */ 3070 mptcp_do_fastclose(sk); 3071 timeout = 0; 3072 } else if (mptcp_close_state(sk)) { 3073 __mptcp_wr_shutdown(sk); 3074 } 3075 3076 sk_stream_wait_close(sk, timeout); 3077 3078 cleanup: 3079 /* orphan all the subflows */ 3080 mptcp_for_each_subflow(msk, subflow) { 3081 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3082 bool slow = lock_sock_fast_nested(ssk); 3083 3084 subflows_alive += ssk->sk_state != TCP_CLOSE; 3085 3086 /* since the close timeout takes precedence on the fail one, 3087 * cancel the latter 3088 */ 3089 if (ssk == msk->first) 3090 subflow->fail_tout = 0; 3091 3092 /* detach from the parent socket, but allow data_ready to 3093 * push incoming data into the mptcp stack, to properly ack it 3094 */ 3095 ssk->sk_socket = NULL; 3096 ssk->sk_wq = NULL; 3097 unlock_sock_fast(ssk, slow); 3098 } 3099 sock_orphan(sk); 3100 3101 /* all the subflows are closed, only timeout can change the msk 3102 * state, let's not keep resources busy for no reasons 3103 */ 3104 if (subflows_alive == 0) 3105 mptcp_set_state(sk, TCP_CLOSE); 3106 3107 sock_hold(sk); 3108 pr_debug("msk=%p state=%d\n", sk, sk->sk_state); 3109 if (msk->token) 3110 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL); 3111 3112 if (sk->sk_state == TCP_CLOSE) { 3113 __mptcp_destroy_sock(sk); 3114 do_cancel_work = true; 3115 } else { 3116 mptcp_start_tout_timer(sk); 3117 } 3118 3119 return do_cancel_work; 3120 } 3121 3122 static void mptcp_close(struct sock *sk, long timeout) 3123 { 3124 bool do_cancel_work; 3125 3126 lock_sock(sk); 3127 3128 do_cancel_work = __mptcp_close(sk, timeout); 3129 release_sock(sk); 3130 if (do_cancel_work) 3131 mptcp_cancel_work(sk); 3132 3133 sock_put(sk); 3134 } 3135 3136 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 3137 { 3138 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3139 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 3140 struct ipv6_pinfo *msk6 = inet6_sk(msk); 3141 3142 msk->sk_v6_daddr = ssk->sk_v6_daddr; 3143 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 3144 3145 if (msk6 && ssk6) { 3146 msk6->saddr = ssk6->saddr; 3147 msk6->flow_label = ssk6->flow_label; 3148 } 3149 #endif 3150 3151 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 3152 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 3153 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 3154 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 3155 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 3156 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 3157 } 3158 3159 static int mptcp_disconnect(struct sock *sk, int flags) 3160 { 3161 struct mptcp_sock *msk = mptcp_sk(sk); 3162 3163 /* We are on the fastopen error path. We can't call straight into the 3164 * subflows cleanup code due to lock nesting (we are already under 3165 * msk->firstsocket lock). 3166 */ 3167 if (msk->fastopening) 3168 return -EBUSY; 3169 3170 mptcp_check_listen_stop(sk); 3171 mptcp_set_state(sk, TCP_CLOSE); 3172 3173 mptcp_stop_rtx_timer(sk); 3174 mptcp_stop_tout_timer(sk); 3175 3176 if (msk->token) 3177 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL); 3178 3179 /* msk->subflow is still intact, the following will not free the first 3180 * subflow 3181 */ 3182 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE); 3183 WRITE_ONCE(msk->flags, 0); 3184 msk->cb_flags = 0; 3185 msk->recovery = false; 3186 msk->can_ack = false; 3187 msk->fully_established = false; 3188 msk->rcv_data_fin = false; 3189 msk->snd_data_fin_enable = false; 3190 msk->rcv_fastclose = false; 3191 msk->use_64bit_ack = false; 3192 msk->bytes_consumed = 0; 3193 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 3194 mptcp_pm_data_reset(msk); 3195 mptcp_ca_reset(sk); 3196 msk->bytes_acked = 0; 3197 msk->bytes_received = 0; 3198 msk->bytes_sent = 0; 3199 msk->bytes_retrans = 0; 3200 msk->rcvspace_init = 0; 3201 3202 WRITE_ONCE(sk->sk_shutdown, 0); 3203 sk_error_report(sk); 3204 return 0; 3205 } 3206 3207 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3208 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 3209 { 3210 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo); 3211 3212 return (struct ipv6_pinfo *)(((u8 *)sk) + offset); 3213 } 3214 3215 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk) 3216 { 3217 const struct ipv6_pinfo *np = inet6_sk(sk); 3218 struct ipv6_txoptions *opt; 3219 struct ipv6_pinfo *newnp; 3220 3221 newnp = inet6_sk(newsk); 3222 3223 rcu_read_lock(); 3224 opt = rcu_dereference(np->opt); 3225 if (opt) { 3226 opt = ipv6_dup_options(newsk, opt); 3227 if (!opt) 3228 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__); 3229 } 3230 RCU_INIT_POINTER(newnp->opt, opt); 3231 rcu_read_unlock(); 3232 } 3233 #endif 3234 3235 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk) 3236 { 3237 struct ip_options_rcu *inet_opt, *newopt = NULL; 3238 const struct inet_sock *inet = inet_sk(sk); 3239 struct inet_sock *newinet; 3240 3241 newinet = inet_sk(newsk); 3242 3243 rcu_read_lock(); 3244 inet_opt = rcu_dereference(inet->inet_opt); 3245 if (inet_opt) { 3246 newopt = sock_kmalloc(newsk, sizeof(*inet_opt) + 3247 inet_opt->opt.optlen, GFP_ATOMIC); 3248 if (newopt) 3249 memcpy(newopt, inet_opt, sizeof(*inet_opt) + 3250 inet_opt->opt.optlen); 3251 else 3252 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__); 3253 } 3254 RCU_INIT_POINTER(newinet->inet_opt, newopt); 3255 rcu_read_unlock(); 3256 } 3257 3258 struct sock *mptcp_sk_clone_init(const struct sock *sk, 3259 const struct mptcp_options_received *mp_opt, 3260 struct sock *ssk, 3261 struct request_sock *req) 3262 { 3263 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 3264 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 3265 struct mptcp_subflow_context *subflow; 3266 struct mptcp_sock *msk; 3267 3268 if (!nsk) 3269 return NULL; 3270 3271 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3272 if (nsk->sk_family == AF_INET6) 3273 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 3274 #endif 3275 3276 __mptcp_init_sock(nsk); 3277 3278 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3279 if (nsk->sk_family == AF_INET6) 3280 mptcp_copy_ip6_options(nsk, sk); 3281 else 3282 #endif 3283 mptcp_copy_ip_options(nsk, sk); 3284 3285 msk = mptcp_sk(nsk); 3286 msk->local_key = subflow_req->local_key; 3287 msk->token = subflow_req->token; 3288 msk->in_accept_queue = 1; 3289 WRITE_ONCE(msk->fully_established, false); 3290 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 3291 WRITE_ONCE(msk->csum_enabled, true); 3292 3293 msk->write_seq = subflow_req->idsn + 1; 3294 msk->snd_nxt = msk->write_seq; 3295 msk->snd_una = msk->write_seq; 3296 msk->wnd_end = msk->snd_nxt + tcp_sk(ssk)->snd_wnd; 3297 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 3298 mptcp_init_sched(msk, mptcp_sk(sk)->sched); 3299 3300 /* passive msk is created after the first/MPC subflow */ 3301 msk->subflow_id = 2; 3302 3303 sock_reset_flag(nsk, SOCK_RCU_FREE); 3304 security_inet_csk_clone(nsk, req); 3305 3306 /* this can't race with mptcp_close(), as the msk is 3307 * not yet exposted to user-space 3308 */ 3309 mptcp_set_state(nsk, TCP_ESTABLISHED); 3310 3311 /* The msk maintain a ref to each subflow in the connections list */ 3312 WRITE_ONCE(msk->first, ssk); 3313 subflow = mptcp_subflow_ctx(ssk); 3314 list_add(&subflow->node, &msk->conn_list); 3315 sock_hold(ssk); 3316 3317 /* new mpc subflow takes ownership of the newly 3318 * created mptcp socket 3319 */ 3320 mptcp_token_accept(subflow_req, msk); 3321 3322 /* set msk addresses early to ensure mptcp_pm_get_local_id() 3323 * uses the correct data 3324 */ 3325 mptcp_copy_inaddrs(nsk, ssk); 3326 __mptcp_propagate_sndbuf(nsk, ssk); 3327 3328 mptcp_rcv_space_init(msk, ssk); 3329 3330 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 3331 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 3332 bh_unlock_sock(nsk); 3333 3334 /* note: the newly allocated socket refcount is 2 now */ 3335 return nsk; 3336 } 3337 3338 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3339 { 3340 const struct tcp_sock *tp = tcp_sk(ssk); 3341 3342 msk->rcvspace_init = 1; 3343 msk->rcvq_space.copied = 0; 3344 msk->rcvq_space.rtt_us = 0; 3345 3346 msk->rcvq_space.time = tp->tcp_mstamp; 3347 3348 /* initial rcv_space offering made to peer */ 3349 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3350 TCP_INIT_CWND * tp->advmss); 3351 if (msk->rcvq_space.space == 0) 3352 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3353 } 3354 3355 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags) 3356 { 3357 struct mptcp_subflow_context *subflow, *tmp; 3358 struct sock *sk = (struct sock *)msk; 3359 3360 __mptcp_clear_xmit(sk); 3361 3362 /* join list will be eventually flushed (with rst) at sock lock release time */ 3363 mptcp_for_each_subflow_safe(msk, subflow, tmp) 3364 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags); 3365 3366 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */ 3367 mptcp_data_lock(sk); 3368 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue); 3369 __skb_queue_purge(&sk->sk_receive_queue); 3370 skb_rbtree_purge(&msk->out_of_order_queue); 3371 mptcp_data_unlock(sk); 3372 3373 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3374 * inet_sock_destruct() will dispose it 3375 */ 3376 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc); 3377 WRITE_ONCE(msk->rmem_fwd_alloc, 0); 3378 mptcp_token_destroy(msk); 3379 mptcp_pm_free_anno_list(msk); 3380 mptcp_free_local_addr_list(msk); 3381 } 3382 3383 static void mptcp_destroy(struct sock *sk) 3384 { 3385 struct mptcp_sock *msk = mptcp_sk(sk); 3386 3387 /* allow the following to close even the initial subflow */ 3388 msk->free_first = 1; 3389 mptcp_destroy_common(msk, 0); 3390 sk_sockets_allocated_dec(sk); 3391 } 3392 3393 void __mptcp_data_acked(struct sock *sk) 3394 { 3395 if (!sock_owned_by_user(sk)) 3396 __mptcp_clean_una(sk); 3397 else 3398 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3399 3400 if (mptcp_pending_data_fin_ack(sk)) 3401 mptcp_schedule_work(sk); 3402 } 3403 3404 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3405 { 3406 if (!mptcp_send_head(sk)) 3407 return; 3408 3409 if (!sock_owned_by_user(sk)) 3410 __mptcp_subflow_push_pending(sk, ssk, false); 3411 else 3412 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3413 } 3414 3415 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3416 BIT(MPTCP_RETRANSMIT) | \ 3417 BIT(MPTCP_FLUSH_JOIN_LIST)) 3418 3419 /* processes deferred events and flush wmem */ 3420 static void mptcp_release_cb(struct sock *sk) 3421 __must_hold(&sk->sk_lock.slock) 3422 { 3423 struct mptcp_sock *msk = mptcp_sk(sk); 3424 3425 for (;;) { 3426 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED); 3427 struct list_head join_list; 3428 3429 if (!flags) 3430 break; 3431 3432 INIT_LIST_HEAD(&join_list); 3433 list_splice_init(&msk->join_list, &join_list); 3434 3435 /* the following actions acquire the subflow socket lock 3436 * 3437 * 1) can't be invoked in atomic scope 3438 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3439 * datapath acquires the msk socket spinlock while helding 3440 * the subflow socket lock 3441 */ 3442 msk->cb_flags &= ~flags; 3443 spin_unlock_bh(&sk->sk_lock.slock); 3444 3445 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3446 __mptcp_flush_join_list(sk, &join_list); 3447 if (flags & BIT(MPTCP_PUSH_PENDING)) 3448 __mptcp_push_pending(sk, 0); 3449 if (flags & BIT(MPTCP_RETRANSMIT)) 3450 __mptcp_retrans(sk); 3451 3452 cond_resched(); 3453 spin_lock_bh(&sk->sk_lock.slock); 3454 } 3455 3456 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3457 __mptcp_clean_una_wakeup(sk); 3458 if (unlikely(msk->cb_flags)) { 3459 /* be sure to sync the msk state before taking actions 3460 * depending on sk_state (MPTCP_ERROR_REPORT) 3461 * On sk release avoid actions depending on the first subflow 3462 */ 3463 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first) 3464 __mptcp_sync_state(sk, msk->pending_state); 3465 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3466 __mptcp_error_report(sk); 3467 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags)) 3468 __mptcp_sync_sndbuf(sk); 3469 } 3470 3471 __mptcp_update_rmem(sk); 3472 } 3473 3474 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3475 * TCP can't schedule delack timer before the subflow is fully established. 3476 * MPTCP uses the delack timer to do 3rd ack retransmissions 3477 */ 3478 static void schedule_3rdack_retransmission(struct sock *ssk) 3479 { 3480 struct inet_connection_sock *icsk = inet_csk(ssk); 3481 struct tcp_sock *tp = tcp_sk(ssk); 3482 unsigned long timeout; 3483 3484 if (mptcp_subflow_ctx(ssk)->fully_established) 3485 return; 3486 3487 /* reschedule with a timeout above RTT, as we must look only for drop */ 3488 if (tp->srtt_us) 3489 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3490 else 3491 timeout = TCP_TIMEOUT_INIT; 3492 timeout += jiffies; 3493 3494 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3495 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3496 icsk->icsk_ack.timeout = timeout; 3497 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3498 } 3499 3500 void mptcp_subflow_process_delegated(struct sock *ssk, long status) 3501 { 3502 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3503 struct sock *sk = subflow->conn; 3504 3505 if (status & BIT(MPTCP_DELEGATE_SEND)) { 3506 mptcp_data_lock(sk); 3507 if (!sock_owned_by_user(sk)) 3508 __mptcp_subflow_push_pending(sk, ssk, true); 3509 else 3510 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3511 mptcp_data_unlock(sk); 3512 } 3513 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) { 3514 mptcp_data_lock(sk); 3515 if (!sock_owned_by_user(sk)) 3516 __mptcp_sync_sndbuf(sk); 3517 else 3518 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags); 3519 mptcp_data_unlock(sk); 3520 } 3521 if (status & BIT(MPTCP_DELEGATE_ACK)) 3522 schedule_3rdack_retransmission(ssk); 3523 } 3524 3525 static int mptcp_hash(struct sock *sk) 3526 { 3527 /* should never be called, 3528 * we hash the TCP subflows not the master socket 3529 */ 3530 WARN_ON_ONCE(1); 3531 return 0; 3532 } 3533 3534 static void mptcp_unhash(struct sock *sk) 3535 { 3536 /* called from sk_common_release(), but nothing to do here */ 3537 } 3538 3539 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3540 { 3541 struct mptcp_sock *msk = mptcp_sk(sk); 3542 3543 pr_debug("msk=%p, ssk=%p\n", msk, msk->first); 3544 if (WARN_ON_ONCE(!msk->first)) 3545 return -EINVAL; 3546 3547 return inet_csk_get_port(msk->first, snum); 3548 } 3549 3550 void mptcp_finish_connect(struct sock *ssk) 3551 { 3552 struct mptcp_subflow_context *subflow; 3553 struct mptcp_sock *msk; 3554 struct sock *sk; 3555 3556 subflow = mptcp_subflow_ctx(ssk); 3557 sk = subflow->conn; 3558 msk = mptcp_sk(sk); 3559 3560 pr_debug("msk=%p, token=%u\n", sk, subflow->token); 3561 3562 subflow->map_seq = subflow->iasn; 3563 subflow->map_subflow_seq = 1; 3564 3565 /* the socket is not connected yet, no msk/subflow ops can access/race 3566 * accessing the field below 3567 */ 3568 WRITE_ONCE(msk->local_key, subflow->local_key); 3569 3570 mptcp_pm_new_connection(msk, ssk, 0); 3571 } 3572 3573 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3574 { 3575 write_lock_bh(&sk->sk_callback_lock); 3576 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3577 sk_set_socket(sk, parent); 3578 sk->sk_uid = SOCK_INODE(parent)->i_uid; 3579 write_unlock_bh(&sk->sk_callback_lock); 3580 } 3581 3582 bool mptcp_finish_join(struct sock *ssk) 3583 { 3584 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3585 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3586 struct sock *parent = (void *)msk; 3587 bool ret = true; 3588 3589 pr_debug("msk=%p, subflow=%p\n", msk, subflow); 3590 3591 /* mptcp socket already closing? */ 3592 if (!mptcp_is_fully_established(parent)) { 3593 subflow->reset_reason = MPTCP_RST_EMPTCP; 3594 return false; 3595 } 3596 3597 /* active subflow, already present inside the conn_list */ 3598 if (!list_empty(&subflow->node)) { 3599 mptcp_subflow_joined(msk, ssk); 3600 mptcp_propagate_sndbuf(parent, ssk); 3601 return true; 3602 } 3603 3604 if (!mptcp_pm_allow_new_subflow(msk)) 3605 goto err_prohibited; 3606 3607 /* If we can't acquire msk socket lock here, let the release callback 3608 * handle it 3609 */ 3610 mptcp_data_lock(parent); 3611 if (!sock_owned_by_user(parent)) { 3612 ret = __mptcp_finish_join(msk, ssk); 3613 if (ret) { 3614 sock_hold(ssk); 3615 list_add_tail(&subflow->node, &msk->conn_list); 3616 } 3617 } else { 3618 sock_hold(ssk); 3619 list_add_tail(&subflow->node, &msk->join_list); 3620 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3621 } 3622 mptcp_data_unlock(parent); 3623 3624 if (!ret) { 3625 err_prohibited: 3626 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3627 return false; 3628 } 3629 3630 return true; 3631 } 3632 3633 static void mptcp_shutdown(struct sock *sk, int how) 3634 { 3635 pr_debug("sk=%p, how=%d\n", sk, how); 3636 3637 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3638 __mptcp_wr_shutdown(sk); 3639 } 3640 3641 static int mptcp_forward_alloc_get(const struct sock *sk) 3642 { 3643 return READ_ONCE(sk->sk_forward_alloc) + 3644 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc); 3645 } 3646 3647 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3648 { 3649 const struct sock *sk = (void *)msk; 3650 u64 delta; 3651 3652 if (sk->sk_state == TCP_LISTEN) 3653 return -EINVAL; 3654 3655 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3656 return 0; 3657 3658 delta = msk->write_seq - v; 3659 if (__mptcp_check_fallback(msk) && msk->first) { 3660 struct tcp_sock *tp = tcp_sk(msk->first); 3661 3662 /* the first subflow is disconnected after close - see 3663 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3664 * so ignore that status, too. 3665 */ 3666 if (!((1 << msk->first->sk_state) & 3667 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3668 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3669 } 3670 if (delta > INT_MAX) 3671 delta = INT_MAX; 3672 3673 return (int)delta; 3674 } 3675 3676 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg) 3677 { 3678 struct mptcp_sock *msk = mptcp_sk(sk); 3679 bool slow; 3680 3681 switch (cmd) { 3682 case SIOCINQ: 3683 if (sk->sk_state == TCP_LISTEN) 3684 return -EINVAL; 3685 3686 lock_sock(sk); 3687 __mptcp_move_skbs(msk); 3688 *karg = mptcp_inq_hint(sk); 3689 release_sock(sk); 3690 break; 3691 case SIOCOUTQ: 3692 slow = lock_sock_fast(sk); 3693 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3694 unlock_sock_fast(sk, slow); 3695 break; 3696 case SIOCOUTQNSD: 3697 slow = lock_sock_fast(sk); 3698 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt); 3699 unlock_sock_fast(sk, slow); 3700 break; 3701 default: 3702 return -ENOIOCTLCMD; 3703 } 3704 3705 return 0; 3706 } 3707 3708 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk, 3709 struct mptcp_subflow_context *subflow) 3710 { 3711 subflow->request_mptcp = 0; 3712 __mptcp_do_fallback(msk); 3713 } 3714 3715 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) 3716 { 3717 struct mptcp_subflow_context *subflow; 3718 struct mptcp_sock *msk = mptcp_sk(sk); 3719 int err = -EINVAL; 3720 struct sock *ssk; 3721 3722 ssk = __mptcp_nmpc_sk(msk); 3723 if (IS_ERR(ssk)) 3724 return PTR_ERR(ssk); 3725 3726 mptcp_set_state(sk, TCP_SYN_SENT); 3727 subflow = mptcp_subflow_ctx(ssk); 3728 #ifdef CONFIG_TCP_MD5SIG 3729 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3730 * TCP option space. 3731 */ 3732 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info)) 3733 mptcp_subflow_early_fallback(msk, subflow); 3734 #endif 3735 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) { 3736 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT); 3737 mptcp_subflow_early_fallback(msk, subflow); 3738 } 3739 3740 WRITE_ONCE(msk->write_seq, subflow->idsn); 3741 WRITE_ONCE(msk->snd_nxt, subflow->idsn); 3742 WRITE_ONCE(msk->snd_una, subflow->idsn); 3743 if (likely(!__mptcp_check_fallback(msk))) 3744 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE); 3745 3746 /* if reaching here via the fastopen/sendmsg path, the caller already 3747 * acquired the subflow socket lock, too. 3748 */ 3749 if (!msk->fastopening) 3750 lock_sock(ssk); 3751 3752 /* the following mirrors closely a very small chunk of code from 3753 * __inet_stream_connect() 3754 */ 3755 if (ssk->sk_state != TCP_CLOSE) 3756 goto out; 3757 3758 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) { 3759 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len); 3760 if (err) 3761 goto out; 3762 } 3763 3764 err = ssk->sk_prot->connect(ssk, uaddr, addr_len); 3765 if (err < 0) 3766 goto out; 3767 3768 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk)); 3769 3770 out: 3771 if (!msk->fastopening) 3772 release_sock(ssk); 3773 3774 /* on successful connect, the msk state will be moved to established by 3775 * subflow_finish_connect() 3776 */ 3777 if (unlikely(err)) { 3778 /* avoid leaving a dangling token in an unconnected socket */ 3779 mptcp_token_destroy(msk); 3780 mptcp_set_state(sk, TCP_CLOSE); 3781 return err; 3782 } 3783 3784 mptcp_copy_inaddrs(sk, ssk); 3785 return 0; 3786 } 3787 3788 static struct proto mptcp_prot = { 3789 .name = "MPTCP", 3790 .owner = THIS_MODULE, 3791 .init = mptcp_init_sock, 3792 .connect = mptcp_connect, 3793 .disconnect = mptcp_disconnect, 3794 .close = mptcp_close, 3795 .setsockopt = mptcp_setsockopt, 3796 .getsockopt = mptcp_getsockopt, 3797 .shutdown = mptcp_shutdown, 3798 .destroy = mptcp_destroy, 3799 .sendmsg = mptcp_sendmsg, 3800 .ioctl = mptcp_ioctl, 3801 .recvmsg = mptcp_recvmsg, 3802 .release_cb = mptcp_release_cb, 3803 .hash = mptcp_hash, 3804 .unhash = mptcp_unhash, 3805 .get_port = mptcp_get_port, 3806 .forward_alloc_get = mptcp_forward_alloc_get, 3807 .sockets_allocated = &mptcp_sockets_allocated, 3808 3809 .memory_allocated = &tcp_memory_allocated, 3810 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc, 3811 3812 .memory_pressure = &tcp_memory_pressure, 3813 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3814 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3815 .sysctl_mem = sysctl_tcp_mem, 3816 .obj_size = sizeof(struct mptcp_sock), 3817 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3818 .no_autobind = true, 3819 }; 3820 3821 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3822 { 3823 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3824 struct sock *ssk, *sk = sock->sk; 3825 int err = -EINVAL; 3826 3827 lock_sock(sk); 3828 ssk = __mptcp_nmpc_sk(msk); 3829 if (IS_ERR(ssk)) { 3830 err = PTR_ERR(ssk); 3831 goto unlock; 3832 } 3833 3834 if (sk->sk_family == AF_INET) 3835 err = inet_bind_sk(ssk, uaddr, addr_len); 3836 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3837 else if (sk->sk_family == AF_INET6) 3838 err = inet6_bind_sk(ssk, uaddr, addr_len); 3839 #endif 3840 if (!err) 3841 mptcp_copy_inaddrs(sk, ssk); 3842 3843 unlock: 3844 release_sock(sk); 3845 return err; 3846 } 3847 3848 static int mptcp_listen(struct socket *sock, int backlog) 3849 { 3850 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3851 struct sock *sk = sock->sk; 3852 struct sock *ssk; 3853 int err; 3854 3855 pr_debug("msk=%p\n", msk); 3856 3857 lock_sock(sk); 3858 3859 err = -EINVAL; 3860 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM) 3861 goto unlock; 3862 3863 ssk = __mptcp_nmpc_sk(msk); 3864 if (IS_ERR(ssk)) { 3865 err = PTR_ERR(ssk); 3866 goto unlock; 3867 } 3868 3869 mptcp_set_state(sk, TCP_LISTEN); 3870 sock_set_flag(sk, SOCK_RCU_FREE); 3871 3872 lock_sock(ssk); 3873 err = __inet_listen_sk(ssk, backlog); 3874 release_sock(ssk); 3875 mptcp_set_state(sk, inet_sk_state_load(ssk)); 3876 3877 if (!err) { 3878 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 3879 mptcp_copy_inaddrs(sk, ssk); 3880 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED); 3881 } 3882 3883 unlock: 3884 release_sock(sk); 3885 return err; 3886 } 3887 3888 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3889 int flags, bool kern) 3890 { 3891 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3892 struct sock *ssk, *newsk; 3893 int err; 3894 3895 pr_debug("msk=%p\n", msk); 3896 3897 /* Buggy applications can call accept on socket states other then LISTEN 3898 * but no need to allocate the first subflow just to error out. 3899 */ 3900 ssk = READ_ONCE(msk->first); 3901 if (!ssk) 3902 return -EINVAL; 3903 3904 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk)); 3905 newsk = inet_csk_accept(ssk, flags, &err, kern); 3906 if (!newsk) 3907 return err; 3908 3909 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk)); 3910 if (sk_is_mptcp(newsk)) { 3911 struct mptcp_subflow_context *subflow; 3912 struct sock *new_mptcp_sock; 3913 3914 subflow = mptcp_subflow_ctx(newsk); 3915 new_mptcp_sock = subflow->conn; 3916 3917 /* is_mptcp should be false if subflow->conn is missing, see 3918 * subflow_syn_recv_sock() 3919 */ 3920 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3921 tcp_sk(newsk)->is_mptcp = 0; 3922 goto tcpfallback; 3923 } 3924 3925 newsk = new_mptcp_sock; 3926 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3927 3928 newsk->sk_kern_sock = kern; 3929 lock_sock(newsk); 3930 __inet_accept(sock, newsock, newsk); 3931 3932 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags); 3933 msk = mptcp_sk(newsk); 3934 msk->in_accept_queue = 0; 3935 3936 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3937 * This is needed so NOSPACE flag can be set from tcp stack. 3938 */ 3939 mptcp_for_each_subflow(msk, subflow) { 3940 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3941 3942 if (!ssk->sk_socket) 3943 mptcp_sock_graft(ssk, newsock); 3944 } 3945 3946 /* Do late cleanup for the first subflow as necessary. Also 3947 * deal with bad peers not doing a complete shutdown. 3948 */ 3949 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) { 3950 __mptcp_close_ssk(newsk, msk->first, 3951 mptcp_subflow_ctx(msk->first), 0); 3952 if (unlikely(list_is_singular(&msk->conn_list))) 3953 mptcp_set_state(newsk, TCP_CLOSE); 3954 } 3955 } else { 3956 tcpfallback: 3957 newsk->sk_kern_sock = kern; 3958 lock_sock(newsk); 3959 __inet_accept(sock, newsock, newsk); 3960 /* we are being invoked after accepting a non-mp-capable 3961 * flow: sk is a tcp_sk, not an mptcp one. 3962 * 3963 * Hand the socket over to tcp so all further socket ops 3964 * bypass mptcp. 3965 */ 3966 WRITE_ONCE(newsock->sk->sk_socket->ops, 3967 mptcp_fallback_tcp_ops(newsock->sk)); 3968 } 3969 release_sock(newsk); 3970 3971 return 0; 3972 } 3973 3974 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3975 { 3976 struct sock *sk = (struct sock *)msk; 3977 3978 if (sk_stream_is_writeable(sk)) 3979 return EPOLLOUT | EPOLLWRNORM; 3980 3981 mptcp_set_nospace(sk); 3982 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */ 3983 if (sk_stream_is_writeable(sk)) 3984 return EPOLLOUT | EPOLLWRNORM; 3985 3986 return 0; 3987 } 3988 3989 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3990 struct poll_table_struct *wait) 3991 { 3992 struct sock *sk = sock->sk; 3993 struct mptcp_sock *msk; 3994 __poll_t mask = 0; 3995 u8 shutdown; 3996 int state; 3997 3998 msk = mptcp_sk(sk); 3999 sock_poll_wait(file, sock, wait); 4000 4001 state = inet_sk_state_load(sk); 4002 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags); 4003 if (state == TCP_LISTEN) { 4004 struct sock *ssk = READ_ONCE(msk->first); 4005 4006 if (WARN_ON_ONCE(!ssk)) 4007 return 0; 4008 4009 return inet_csk_listen_poll(ssk); 4010 } 4011 4012 shutdown = READ_ONCE(sk->sk_shutdown); 4013 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 4014 mask |= EPOLLHUP; 4015 if (shutdown & RCV_SHUTDOWN) 4016 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 4017 4018 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 4019 mask |= mptcp_check_readable(sk); 4020 if (shutdown & SEND_SHUTDOWN) 4021 mask |= EPOLLOUT | EPOLLWRNORM; 4022 else 4023 mask |= mptcp_check_writeable(msk); 4024 } else if (state == TCP_SYN_SENT && 4025 inet_test_bit(DEFER_CONNECT, sk)) { 4026 /* cf tcp_poll() note about TFO */ 4027 mask |= EPOLLOUT | EPOLLWRNORM; 4028 } 4029 4030 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */ 4031 smp_rmb(); 4032 if (READ_ONCE(sk->sk_err)) 4033 mask |= EPOLLERR; 4034 4035 return mask; 4036 } 4037 4038 static const struct proto_ops mptcp_stream_ops = { 4039 .family = PF_INET, 4040 .owner = THIS_MODULE, 4041 .release = inet_release, 4042 .bind = mptcp_bind, 4043 .connect = inet_stream_connect, 4044 .socketpair = sock_no_socketpair, 4045 .accept = mptcp_stream_accept, 4046 .getname = inet_getname, 4047 .poll = mptcp_poll, 4048 .ioctl = inet_ioctl, 4049 .gettstamp = sock_gettstamp, 4050 .listen = mptcp_listen, 4051 .shutdown = inet_shutdown, 4052 .setsockopt = sock_common_setsockopt, 4053 .getsockopt = sock_common_getsockopt, 4054 .sendmsg = inet_sendmsg, 4055 .recvmsg = inet_recvmsg, 4056 .mmap = sock_no_mmap, 4057 .set_rcvlowat = mptcp_set_rcvlowat, 4058 }; 4059 4060 static struct inet_protosw mptcp_protosw = { 4061 .type = SOCK_STREAM, 4062 .protocol = IPPROTO_MPTCP, 4063 .prot = &mptcp_prot, 4064 .ops = &mptcp_stream_ops, 4065 .flags = INET_PROTOSW_ICSK, 4066 }; 4067 4068 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 4069 { 4070 struct mptcp_delegated_action *delegated; 4071 struct mptcp_subflow_context *subflow; 4072 int work_done = 0; 4073 4074 delegated = container_of(napi, struct mptcp_delegated_action, napi); 4075 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 4076 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4077 4078 bh_lock_sock_nested(ssk); 4079 if (!sock_owned_by_user(ssk)) { 4080 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0)); 4081 } else { 4082 /* tcp_release_cb_override already processed 4083 * the action or will do at next release_sock(). 4084 * In both case must dequeue the subflow here - on the same 4085 * CPU that scheduled it. 4086 */ 4087 smp_wmb(); 4088 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status); 4089 } 4090 bh_unlock_sock(ssk); 4091 sock_put(ssk); 4092 4093 if (++work_done == budget) 4094 return budget; 4095 } 4096 4097 /* always provide a 0 'work_done' argument, so that napi_complete_done 4098 * will not try accessing the NULL napi->dev ptr 4099 */ 4100 napi_complete_done(napi, 0); 4101 return work_done; 4102 } 4103 4104 void __init mptcp_proto_init(void) 4105 { 4106 struct mptcp_delegated_action *delegated; 4107 int cpu; 4108 4109 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 4110 4111 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 4112 panic("Failed to allocate MPTCP pcpu counter\n"); 4113 4114 init_dummy_netdev(&mptcp_napi_dev); 4115 for_each_possible_cpu(cpu) { 4116 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 4117 INIT_LIST_HEAD(&delegated->head); 4118 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi, 4119 mptcp_napi_poll); 4120 napi_enable(&delegated->napi); 4121 } 4122 4123 mptcp_subflow_init(); 4124 mptcp_pm_init(); 4125 mptcp_sched_init(); 4126 mptcp_token_init(); 4127 4128 if (proto_register(&mptcp_prot, 1) != 0) 4129 panic("Failed to register MPTCP proto.\n"); 4130 4131 inet_register_protosw(&mptcp_protosw); 4132 4133 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 4134 } 4135 4136 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 4137 static const struct proto_ops mptcp_v6_stream_ops = { 4138 .family = PF_INET6, 4139 .owner = THIS_MODULE, 4140 .release = inet6_release, 4141 .bind = mptcp_bind, 4142 .connect = inet_stream_connect, 4143 .socketpair = sock_no_socketpair, 4144 .accept = mptcp_stream_accept, 4145 .getname = inet6_getname, 4146 .poll = mptcp_poll, 4147 .ioctl = inet6_ioctl, 4148 .gettstamp = sock_gettstamp, 4149 .listen = mptcp_listen, 4150 .shutdown = inet_shutdown, 4151 .setsockopt = sock_common_setsockopt, 4152 .getsockopt = sock_common_getsockopt, 4153 .sendmsg = inet6_sendmsg, 4154 .recvmsg = inet6_recvmsg, 4155 .mmap = sock_no_mmap, 4156 #ifdef CONFIG_COMPAT 4157 .compat_ioctl = inet6_compat_ioctl, 4158 #endif 4159 .set_rcvlowat = mptcp_set_rcvlowat, 4160 }; 4161 4162 static struct proto mptcp_v6_prot; 4163 4164 static struct inet_protosw mptcp_v6_protosw = { 4165 .type = SOCK_STREAM, 4166 .protocol = IPPROTO_MPTCP, 4167 .prot = &mptcp_v6_prot, 4168 .ops = &mptcp_v6_stream_ops, 4169 .flags = INET_PROTOSW_ICSK, 4170 }; 4171 4172 int __init mptcp_proto_v6_init(void) 4173 { 4174 int err; 4175 4176 mptcp_v6_prot = mptcp_prot; 4177 strcpy(mptcp_v6_prot.name, "MPTCPv6"); 4178 mptcp_v6_prot.slab = NULL; 4179 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 4180 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np); 4181 4182 err = proto_register(&mptcp_v6_prot, 1); 4183 if (err) 4184 return err; 4185 4186 err = inet6_register_protosw(&mptcp_v6_protosw); 4187 if (err) 4188 proto_unregister(&mptcp_v6_prot); 4189 4190 return err; 4191 } 4192 #endif 4193