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