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 struct mptcp_skb_cb { 40 u64 map_seq; 41 u64 end_seq; 42 u32 offset; 43 u8 has_rxtstamp:1; 44 }; 45 46 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0])) 47 48 enum { 49 MPTCP_CMSG_TS = BIT(0), 50 MPTCP_CMSG_INQ = BIT(1), 51 }; 52 53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp; 54 55 static void __mptcp_destroy_sock(struct sock *sk); 56 static void __mptcp_check_send_data_fin(struct sock *sk); 57 58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions); 59 static struct net_device mptcp_napi_dev; 60 61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not 62 * completed yet or has failed, return the subflow socket. 63 * Otherwise return NULL. 64 */ 65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk) 66 { 67 if (!msk->subflow || READ_ONCE(msk->can_ack)) 68 return NULL; 69 70 return msk->subflow; 71 } 72 73 /* Returns end sequence number of the receiver's advertised window */ 74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 75 { 76 return READ_ONCE(msk->wnd_end); 77 } 78 79 static bool mptcp_is_tcpsk(struct sock *sk) 80 { 81 struct socket *sock = sk->sk_socket; 82 83 if (unlikely(sk->sk_prot == &tcp_prot)) { 84 /* we are being invoked after mptcp_accept() has 85 * accepted a non-mp-capable flow: sk is a tcp_sk, 86 * not an mptcp one. 87 * 88 * Hand the socket over to tcp so all further socket ops 89 * bypass mptcp. 90 */ 91 sock->ops = &inet_stream_ops; 92 return true; 93 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 94 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) { 95 sock->ops = &inet6_stream_ops; 96 return true; 97 #endif 98 } 99 100 return false; 101 } 102 103 static int __mptcp_socket_create(struct mptcp_sock *msk) 104 { 105 struct mptcp_subflow_context *subflow; 106 struct sock *sk = (struct sock *)msk; 107 struct socket *ssock; 108 int err; 109 110 err = mptcp_subflow_create_socket(sk, &ssock); 111 if (err) 112 return err; 113 114 msk->first = ssock->sk; 115 msk->subflow = ssock; 116 subflow = mptcp_subflow_ctx(ssock->sk); 117 list_add(&subflow->node, &msk->conn_list); 118 sock_hold(ssock->sk); 119 subflow->request_mptcp = 1; 120 121 /* This is the first subflow, always with id 0 */ 122 subflow->local_id_valid = 1; 123 mptcp_sock_graft(msk->first, sk->sk_socket); 124 125 return 0; 126 } 127 128 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 129 { 130 sk_drops_add(sk, skb); 131 __kfree_skb(skb); 132 } 133 134 static void mptcp_rmem_charge(struct sock *sk, int size) 135 { 136 mptcp_sk(sk)->rmem_fwd_alloc -= size; 137 } 138 139 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 140 struct sk_buff *from) 141 { 142 bool fragstolen; 143 int delta; 144 145 if (MPTCP_SKB_CB(from)->offset || 146 !skb_try_coalesce(to, from, &fragstolen, &delta)) 147 return false; 148 149 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx", 150 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 151 to->len, MPTCP_SKB_CB(from)->end_seq); 152 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 153 kfree_skb_partial(from, fragstolen); 154 atomic_add(delta, &sk->sk_rmem_alloc); 155 mptcp_rmem_charge(sk, delta); 156 return true; 157 } 158 159 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 160 struct sk_buff *from) 161 { 162 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 163 return false; 164 165 return mptcp_try_coalesce((struct sock *)msk, to, from); 166 } 167 168 static void __mptcp_rmem_reclaim(struct sock *sk, int amount) 169 { 170 amount >>= SK_MEM_QUANTUM_SHIFT; 171 mptcp_sk(sk)->rmem_fwd_alloc -= amount << SK_MEM_QUANTUM_SHIFT; 172 __sk_mem_reduce_allocated(sk, amount); 173 } 174 175 static void mptcp_rmem_uncharge(struct sock *sk, int size) 176 { 177 struct mptcp_sock *msk = mptcp_sk(sk); 178 int reclaimable; 179 180 msk->rmem_fwd_alloc += size; 181 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk); 182 183 /* see sk_mem_uncharge() for the rationale behind the following schema */ 184 if (unlikely(reclaimable >= SK_RECLAIM_THRESHOLD)) 185 __mptcp_rmem_reclaim(sk, SK_RECLAIM_CHUNK); 186 } 187 188 static void mptcp_rfree(struct sk_buff *skb) 189 { 190 unsigned int len = skb->truesize; 191 struct sock *sk = skb->sk; 192 193 atomic_sub(len, &sk->sk_rmem_alloc); 194 mptcp_rmem_uncharge(sk, len); 195 } 196 197 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk) 198 { 199 skb_orphan(skb); 200 skb->sk = sk; 201 skb->destructor = mptcp_rfree; 202 atomic_add(skb->truesize, &sk->sk_rmem_alloc); 203 mptcp_rmem_charge(sk, skb->truesize); 204 } 205 206 /* "inspired" by tcp_data_queue_ofo(), main differences: 207 * - use mptcp seqs 208 * - don't cope with sacks 209 */ 210 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 211 { 212 struct sock *sk = (struct sock *)msk; 213 struct rb_node **p, *parent; 214 u64 seq, end_seq, max_seq; 215 struct sk_buff *skb1; 216 217 seq = MPTCP_SKB_CB(skb)->map_seq; 218 end_seq = MPTCP_SKB_CB(skb)->end_seq; 219 max_seq = READ_ONCE(msk->rcv_wnd_sent); 220 221 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq, 222 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 223 if (after64(end_seq, max_seq)) { 224 /* out of window */ 225 mptcp_drop(sk, skb); 226 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 227 (unsigned long long)end_seq - (unsigned long)max_seq, 228 (unsigned long long)msk->rcv_wnd_sent); 229 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 230 return; 231 } 232 233 p = &msk->out_of_order_queue.rb_node; 234 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 235 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 236 rb_link_node(&skb->rbnode, NULL, p); 237 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 238 msk->ooo_last_skb = skb; 239 goto end; 240 } 241 242 /* with 2 subflows, adding at end of ooo queue is quite likely 243 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 244 */ 245 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 248 return; 249 } 250 251 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 252 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 254 parent = &msk->ooo_last_skb->rbnode; 255 p = &parent->rb_right; 256 goto insert; 257 } 258 259 /* Find place to insert this segment. Handle overlaps on the way. */ 260 parent = NULL; 261 while (*p) { 262 parent = *p; 263 skb1 = rb_to_skb(parent); 264 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 265 p = &parent->rb_left; 266 continue; 267 } 268 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 269 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 270 /* All the bits are present. Drop. */ 271 mptcp_drop(sk, skb); 272 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 273 return; 274 } 275 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 276 /* partial overlap: 277 * | skb | 278 * | skb1 | 279 * continue traversing 280 */ 281 } else { 282 /* skb's seq == skb1's seq and skb covers skb1. 283 * Replace skb1 with skb. 284 */ 285 rb_replace_node(&skb1->rbnode, &skb->rbnode, 286 &msk->out_of_order_queue); 287 mptcp_drop(sk, skb1); 288 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 289 goto merge_right; 290 } 291 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 292 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 293 return; 294 } 295 p = &parent->rb_right; 296 } 297 298 insert: 299 /* Insert segment into RB tree. */ 300 rb_link_node(&skb->rbnode, parent, p); 301 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 302 303 merge_right: 304 /* Remove other segments covered by skb. */ 305 while ((skb1 = skb_rb_next(skb)) != NULL) { 306 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 307 break; 308 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 309 mptcp_drop(sk, skb1); 310 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 311 } 312 /* If there is no skb after us, we are the last_skb ! */ 313 if (!skb1) 314 msk->ooo_last_skb = skb; 315 316 end: 317 skb_condense(skb); 318 mptcp_set_owner_r(skb, sk); 319 } 320 321 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size) 322 { 323 struct mptcp_sock *msk = mptcp_sk(sk); 324 int amt, amount; 325 326 if (size < msk->rmem_fwd_alloc) 327 return true; 328 329 amt = sk_mem_pages(size); 330 amount = amt << SK_MEM_QUANTUM_SHIFT; 331 msk->rmem_fwd_alloc += amount; 332 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV)) { 333 if (ssk->sk_forward_alloc < amount) { 334 msk->rmem_fwd_alloc -= amount; 335 return false; 336 } 337 338 ssk->sk_forward_alloc -= amount; 339 } 340 return true; 341 } 342 343 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 344 struct sk_buff *skb, unsigned int offset, 345 size_t copy_len) 346 { 347 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 348 struct sock *sk = (struct sock *)msk; 349 struct sk_buff *tail; 350 bool has_rxtstamp; 351 352 __skb_unlink(skb, &ssk->sk_receive_queue); 353 354 skb_ext_reset(skb); 355 skb_orphan(skb); 356 357 /* try to fetch required memory from subflow */ 358 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize)) 359 goto drop; 360 361 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 362 363 /* the skb map_seq accounts for the skb offset: 364 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 365 * value 366 */ 367 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 368 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 369 MPTCP_SKB_CB(skb)->offset = offset; 370 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 371 372 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 373 /* in sequence */ 374 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 375 tail = skb_peek_tail(&sk->sk_receive_queue); 376 if (tail && mptcp_try_coalesce(sk, tail, skb)) 377 return true; 378 379 mptcp_set_owner_r(skb, sk); 380 __skb_queue_tail(&sk->sk_receive_queue, skb); 381 return true; 382 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 383 mptcp_data_queue_ofo(msk, skb); 384 return false; 385 } 386 387 /* old data, keep it simple and drop the whole pkt, sender 388 * will retransmit as needed, if needed. 389 */ 390 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 391 drop: 392 mptcp_drop(sk, skb); 393 return false; 394 } 395 396 static void mptcp_stop_timer(struct sock *sk) 397 { 398 struct inet_connection_sock *icsk = inet_csk(sk); 399 400 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 401 mptcp_sk(sk)->timer_ival = 0; 402 } 403 404 static void mptcp_close_wake_up(struct sock *sk) 405 { 406 if (sock_flag(sk, SOCK_DEAD)) 407 return; 408 409 sk->sk_state_change(sk); 410 if (sk->sk_shutdown == SHUTDOWN_MASK || 411 sk->sk_state == TCP_CLOSE) 412 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 413 else 414 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 415 } 416 417 static bool mptcp_pending_data_fin_ack(struct sock *sk) 418 { 419 struct mptcp_sock *msk = mptcp_sk(sk); 420 421 return !__mptcp_check_fallback(msk) && 422 ((1 << sk->sk_state) & 423 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 424 msk->write_seq == READ_ONCE(msk->snd_una); 425 } 426 427 static void mptcp_check_data_fin_ack(struct sock *sk) 428 { 429 struct mptcp_sock *msk = mptcp_sk(sk); 430 431 /* Look for an acknowledged DATA_FIN */ 432 if (mptcp_pending_data_fin_ack(sk)) { 433 WRITE_ONCE(msk->snd_data_fin_enable, 0); 434 435 switch (sk->sk_state) { 436 case TCP_FIN_WAIT1: 437 inet_sk_state_store(sk, TCP_FIN_WAIT2); 438 break; 439 case TCP_CLOSING: 440 case TCP_LAST_ACK: 441 inet_sk_state_store(sk, TCP_CLOSE); 442 break; 443 } 444 445 mptcp_close_wake_up(sk); 446 } 447 } 448 449 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 450 { 451 struct mptcp_sock *msk = mptcp_sk(sk); 452 453 if (READ_ONCE(msk->rcv_data_fin) && 454 ((1 << sk->sk_state) & 455 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 456 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 457 458 if (msk->ack_seq == rcv_data_fin_seq) { 459 if (seq) 460 *seq = rcv_data_fin_seq; 461 462 return true; 463 } 464 } 465 466 return false; 467 } 468 469 static void mptcp_set_datafin_timeout(const struct sock *sk) 470 { 471 struct inet_connection_sock *icsk = inet_csk(sk); 472 u32 retransmits; 473 474 retransmits = min_t(u32, icsk->icsk_retransmits, 475 ilog2(TCP_RTO_MAX / TCP_RTO_MIN)); 476 477 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits; 478 } 479 480 static void __mptcp_set_timeout(struct sock *sk, long tout) 481 { 482 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 483 } 484 485 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 486 { 487 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 488 489 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 490 inet_csk(ssk)->icsk_timeout - jiffies : 0; 491 } 492 493 static void mptcp_set_timeout(struct sock *sk) 494 { 495 struct mptcp_subflow_context *subflow; 496 long tout = 0; 497 498 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 499 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 500 __mptcp_set_timeout(sk, tout); 501 } 502 503 static bool tcp_can_send_ack(const struct sock *ssk) 504 { 505 return !((1 << inet_sk_state_load(ssk)) & 506 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 507 } 508 509 void mptcp_subflow_send_ack(struct sock *ssk) 510 { 511 bool slow; 512 513 slow = lock_sock_fast(ssk); 514 if (tcp_can_send_ack(ssk)) 515 tcp_send_ack(ssk); 516 unlock_sock_fast(ssk, slow); 517 } 518 519 static void mptcp_send_ack(struct mptcp_sock *msk) 520 { 521 struct mptcp_subflow_context *subflow; 522 523 mptcp_for_each_subflow(msk, subflow) 524 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow)); 525 } 526 527 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk) 528 { 529 bool slow; 530 531 slow = lock_sock_fast(ssk); 532 if (tcp_can_send_ack(ssk)) 533 tcp_cleanup_rbuf(ssk, 1); 534 unlock_sock_fast(ssk, slow); 535 } 536 537 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 538 { 539 const struct inet_connection_sock *icsk = inet_csk(ssk); 540 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 541 const struct tcp_sock *tp = tcp_sk(ssk); 542 543 return (ack_pending & ICSK_ACK_SCHED) && 544 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 545 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 546 (rx_empty && ack_pending & 547 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 548 } 549 550 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk) 551 { 552 int old_space = READ_ONCE(msk->old_wspace); 553 struct mptcp_subflow_context *subflow; 554 struct sock *sk = (struct sock *)msk; 555 int space = __mptcp_space(sk); 556 bool cleanup, rx_empty; 557 558 cleanup = (space > 0) && (space >= (old_space << 1)); 559 rx_empty = !__mptcp_rmem(sk); 560 561 mptcp_for_each_subflow(msk, subflow) { 562 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 563 564 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 565 mptcp_subflow_cleanup_rbuf(ssk); 566 } 567 } 568 569 static bool mptcp_check_data_fin(struct sock *sk) 570 { 571 struct mptcp_sock *msk = mptcp_sk(sk); 572 u64 rcv_data_fin_seq; 573 bool ret = false; 574 575 if (__mptcp_check_fallback(msk)) 576 return ret; 577 578 /* Need to ack a DATA_FIN received from a peer while this side 579 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 580 * msk->rcv_data_fin was set when parsing the incoming options 581 * at the subflow level and the msk lock was not held, so this 582 * is the first opportunity to act on the DATA_FIN and change 583 * the msk state. 584 * 585 * If we are caught up to the sequence number of the incoming 586 * DATA_FIN, send the DATA_ACK now and do state transition. If 587 * not caught up, do nothing and let the recv code send DATA_ACK 588 * when catching up. 589 */ 590 591 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 592 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 593 WRITE_ONCE(msk->rcv_data_fin, 0); 594 595 sk->sk_shutdown |= RCV_SHUTDOWN; 596 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 597 598 switch (sk->sk_state) { 599 case TCP_ESTABLISHED: 600 inet_sk_state_store(sk, TCP_CLOSE_WAIT); 601 break; 602 case TCP_FIN_WAIT1: 603 inet_sk_state_store(sk, TCP_CLOSING); 604 break; 605 case TCP_FIN_WAIT2: 606 inet_sk_state_store(sk, TCP_CLOSE); 607 break; 608 default: 609 /* Other states not expected */ 610 WARN_ON_ONCE(1); 611 break; 612 } 613 614 ret = true; 615 mptcp_send_ack(msk); 616 mptcp_close_wake_up(sk); 617 } 618 return ret; 619 } 620 621 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 622 struct sock *ssk, 623 unsigned int *bytes) 624 { 625 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 626 struct sock *sk = (struct sock *)msk; 627 unsigned int moved = 0; 628 bool more_data_avail; 629 struct tcp_sock *tp; 630 bool done = false; 631 int sk_rbuf; 632 633 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 634 635 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 636 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 637 638 if (unlikely(ssk_rbuf > sk_rbuf)) { 639 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf); 640 sk_rbuf = ssk_rbuf; 641 } 642 } 643 644 pr_debug("msk=%p ssk=%p", msk, ssk); 645 tp = tcp_sk(ssk); 646 do { 647 u32 map_remaining, offset; 648 u32 seq = tp->copied_seq; 649 struct sk_buff *skb; 650 bool fin; 651 652 /* try to move as much data as available */ 653 map_remaining = subflow->map_data_len - 654 mptcp_subflow_get_map_offset(subflow); 655 656 skb = skb_peek(&ssk->sk_receive_queue); 657 if (!skb) { 658 /* if no data is found, a racing workqueue/recvmsg 659 * already processed the new data, stop here or we 660 * can enter an infinite loop 661 */ 662 if (!moved) 663 done = true; 664 break; 665 } 666 667 if (__mptcp_check_fallback(msk)) { 668 /* if we are running under the workqueue, TCP could have 669 * collapsed skbs between dummy map creation and now 670 * be sure to adjust the size 671 */ 672 map_remaining = skb->len; 673 subflow->map_data_len = skb->len; 674 } 675 676 offset = seq - TCP_SKB_CB(skb)->seq; 677 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 678 if (fin) { 679 done = true; 680 seq++; 681 } 682 683 if (offset < skb->len) { 684 size_t len = skb->len - offset; 685 686 if (tp->urg_data) 687 done = true; 688 689 if (__mptcp_move_skb(msk, ssk, skb, offset, len)) 690 moved += len; 691 seq += len; 692 693 if (WARN_ON_ONCE(map_remaining < len)) 694 break; 695 } else { 696 WARN_ON_ONCE(!fin); 697 sk_eat_skb(ssk, skb); 698 done = true; 699 } 700 701 WRITE_ONCE(tp->copied_seq, seq); 702 more_data_avail = mptcp_subflow_data_available(ssk); 703 704 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) { 705 done = true; 706 break; 707 } 708 } while (more_data_avail); 709 710 *bytes += moved; 711 return done; 712 } 713 714 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 715 { 716 struct sock *sk = (struct sock *)msk; 717 struct sk_buff *skb, *tail; 718 bool moved = false; 719 struct rb_node *p; 720 u64 end_seq; 721 722 p = rb_first(&msk->out_of_order_queue); 723 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 724 while (p) { 725 skb = rb_to_skb(p); 726 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 727 break; 728 729 p = rb_next(p); 730 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 731 732 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 733 msk->ack_seq))) { 734 mptcp_drop(sk, skb); 735 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 736 continue; 737 } 738 739 end_seq = MPTCP_SKB_CB(skb)->end_seq; 740 tail = skb_peek_tail(&sk->sk_receive_queue); 741 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 742 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 743 744 /* skip overlapping data, if any */ 745 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d", 746 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 747 delta); 748 MPTCP_SKB_CB(skb)->offset += delta; 749 MPTCP_SKB_CB(skb)->map_seq += delta; 750 __skb_queue_tail(&sk->sk_receive_queue, skb); 751 } 752 msk->ack_seq = end_seq; 753 moved = true; 754 } 755 return moved; 756 } 757 758 /* In most cases we will be able to lock the mptcp socket. If its already 759 * owned, we need to defer to the work queue to avoid ABBA deadlock. 760 */ 761 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 762 { 763 struct sock *sk = (struct sock *)msk; 764 unsigned int moved = 0; 765 766 __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 767 __mptcp_ofo_queue(msk); 768 if (unlikely(ssk->sk_err)) { 769 if (!sock_owned_by_user(sk)) 770 __mptcp_error_report(sk); 771 else 772 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags); 773 } 774 775 /* If the moves have caught up with the DATA_FIN sequence number 776 * it's time to ack the DATA_FIN and change socket state, but 777 * this is not a good place to change state. Let the workqueue 778 * do it. 779 */ 780 if (mptcp_pending_data_fin(sk, NULL)) 781 mptcp_schedule_work(sk); 782 return moved > 0; 783 } 784 785 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 786 { 787 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 788 struct mptcp_sock *msk = mptcp_sk(sk); 789 int sk_rbuf, ssk_rbuf; 790 791 /* The peer can send data while we are shutting down this 792 * subflow at msk destruction time, but we must avoid enqueuing 793 * more data to the msk receive queue 794 */ 795 if (unlikely(subflow->disposable)) 796 return; 797 798 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf); 799 sk_rbuf = READ_ONCE(sk->sk_rcvbuf); 800 if (unlikely(ssk_rbuf > sk_rbuf)) 801 sk_rbuf = ssk_rbuf; 802 803 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/ 804 if (__mptcp_rmem(sk) > sk_rbuf) { 805 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 806 return; 807 } 808 809 /* Wake-up the reader only for in-sequence data */ 810 mptcp_data_lock(sk); 811 if (move_skbs_to_msk(msk, ssk)) 812 sk->sk_data_ready(sk); 813 814 mptcp_data_unlock(sk); 815 } 816 817 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk) 818 { 819 struct sock *sk = (struct sock *)msk; 820 821 if (sk->sk_state != TCP_ESTABLISHED) 822 return false; 823 824 /* attach to msk socket only after we are sure we will deal with it 825 * at close time 826 */ 827 if (sk->sk_socket && !ssk->sk_socket) 828 mptcp_sock_graft(ssk, sk->sk_socket); 829 830 mptcp_propagate_sndbuf((struct sock *)msk, ssk); 831 mptcp_sockopt_sync_locked(msk, ssk); 832 return true; 833 } 834 835 static void __mptcp_flush_join_list(struct sock *sk) 836 { 837 struct mptcp_subflow_context *tmp, *subflow; 838 struct mptcp_sock *msk = mptcp_sk(sk); 839 840 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) { 841 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 842 bool slow = lock_sock_fast(ssk); 843 844 list_move_tail(&subflow->node, &msk->conn_list); 845 if (!__mptcp_finish_join(msk, ssk)) 846 mptcp_subflow_reset(ssk); 847 unlock_sock_fast(ssk, slow); 848 } 849 } 850 851 static bool mptcp_timer_pending(struct sock *sk) 852 { 853 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 854 } 855 856 static void mptcp_reset_timer(struct sock *sk) 857 { 858 struct inet_connection_sock *icsk = inet_csk(sk); 859 unsigned long tout; 860 861 /* prevent rescheduling on close */ 862 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 863 return; 864 865 tout = mptcp_sk(sk)->timer_ival; 866 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 867 } 868 869 bool mptcp_schedule_work(struct sock *sk) 870 { 871 if (inet_sk_state_load(sk) != TCP_CLOSE && 872 schedule_work(&mptcp_sk(sk)->work)) { 873 /* each subflow already holds a reference to the sk, and the 874 * workqueue is invoked by a subflow, so sk can't go away here. 875 */ 876 sock_hold(sk); 877 return true; 878 } 879 return false; 880 } 881 882 void mptcp_subflow_eof(struct sock *sk) 883 { 884 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags)) 885 mptcp_schedule_work(sk); 886 } 887 888 static void mptcp_check_for_eof(struct mptcp_sock *msk) 889 { 890 struct mptcp_subflow_context *subflow; 891 struct sock *sk = (struct sock *)msk; 892 int receivers = 0; 893 894 mptcp_for_each_subflow(msk, subflow) 895 receivers += !subflow->rx_eof; 896 if (receivers) 897 return; 898 899 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) { 900 /* hopefully temporary hack: propagate shutdown status 901 * to msk, when all subflows agree on it 902 */ 903 sk->sk_shutdown |= RCV_SHUTDOWN; 904 905 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 906 sk->sk_data_ready(sk); 907 } 908 909 switch (sk->sk_state) { 910 case TCP_ESTABLISHED: 911 inet_sk_state_store(sk, TCP_CLOSE_WAIT); 912 break; 913 case TCP_FIN_WAIT1: 914 inet_sk_state_store(sk, TCP_CLOSING); 915 break; 916 case TCP_FIN_WAIT2: 917 inet_sk_state_store(sk, TCP_CLOSE); 918 break; 919 default: 920 return; 921 } 922 mptcp_close_wake_up(sk); 923 } 924 925 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk) 926 { 927 struct mptcp_subflow_context *subflow; 928 struct sock *sk = (struct sock *)msk; 929 930 sock_owned_by_me(sk); 931 932 mptcp_for_each_subflow(msk, subflow) { 933 if (READ_ONCE(subflow->data_avail)) 934 return mptcp_subflow_tcp_sock(subflow); 935 } 936 937 return NULL; 938 } 939 940 static bool mptcp_skb_can_collapse_to(u64 write_seq, 941 const struct sk_buff *skb, 942 const struct mptcp_ext *mpext) 943 { 944 if (!tcp_skb_can_collapse_to(skb)) 945 return false; 946 947 /* can collapse only if MPTCP level sequence is in order and this 948 * mapping has not been xmitted yet 949 */ 950 return mpext && mpext->data_seq + mpext->data_len == write_seq && 951 !mpext->frozen; 952 } 953 954 /* we can append data to the given data frag if: 955 * - there is space available in the backing page_frag 956 * - the data frag tail matches the current page_frag free offset 957 * - the data frag end sequence number matches the current write seq 958 */ 959 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 960 const struct page_frag *pfrag, 961 const struct mptcp_data_frag *df) 962 { 963 return df && pfrag->page == df->page && 964 pfrag->size - pfrag->offset > 0 && 965 pfrag->offset == (df->offset + df->data_len) && 966 df->data_seq + df->data_len == msk->write_seq; 967 } 968 969 static void __mptcp_mem_reclaim_partial(struct sock *sk) 970 { 971 int reclaimable = mptcp_sk(sk)->rmem_fwd_alloc - sk_unused_reserved_mem(sk); 972 973 lockdep_assert_held_once(&sk->sk_lock.slock); 974 975 if (reclaimable > SK_MEM_QUANTUM) 976 __mptcp_rmem_reclaim(sk, reclaimable - 1); 977 978 sk_mem_reclaim_partial(sk); 979 } 980 981 static void mptcp_mem_reclaim_partial(struct sock *sk) 982 { 983 mptcp_data_lock(sk); 984 __mptcp_mem_reclaim_partial(sk); 985 mptcp_data_unlock(sk); 986 } 987 988 static void dfrag_uncharge(struct sock *sk, int len) 989 { 990 sk_mem_uncharge(sk, len); 991 sk_wmem_queued_add(sk, -len); 992 } 993 994 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 995 { 996 int len = dfrag->data_len + dfrag->overhead; 997 998 list_del(&dfrag->list); 999 dfrag_uncharge(sk, len); 1000 put_page(dfrag->page); 1001 } 1002 1003 static void __mptcp_clean_una(struct sock *sk) 1004 { 1005 struct mptcp_sock *msk = mptcp_sk(sk); 1006 struct mptcp_data_frag *dtmp, *dfrag; 1007 bool cleaned = false; 1008 u64 snd_una; 1009 1010 /* on fallback we just need to ignore snd_una, as this is really 1011 * plain TCP 1012 */ 1013 if (__mptcp_check_fallback(msk)) 1014 msk->snd_una = READ_ONCE(msk->snd_nxt); 1015 1016 snd_una = msk->snd_una; 1017 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 1018 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 1019 break; 1020 1021 if (unlikely(dfrag == msk->first_pending)) { 1022 /* in recovery mode can see ack after the current snd head */ 1023 if (WARN_ON_ONCE(!msk->recovery)) 1024 break; 1025 1026 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1027 } 1028 1029 dfrag_clear(sk, dfrag); 1030 cleaned = true; 1031 } 1032 1033 dfrag = mptcp_rtx_head(sk); 1034 if (dfrag && after64(snd_una, dfrag->data_seq)) { 1035 u64 delta = snd_una - dfrag->data_seq; 1036 1037 /* prevent wrap around in recovery mode */ 1038 if (unlikely(delta > dfrag->already_sent)) { 1039 if (WARN_ON_ONCE(!msk->recovery)) 1040 goto out; 1041 if (WARN_ON_ONCE(delta > dfrag->data_len)) 1042 goto out; 1043 dfrag->already_sent += delta - dfrag->already_sent; 1044 } 1045 1046 dfrag->data_seq += delta; 1047 dfrag->offset += delta; 1048 dfrag->data_len -= delta; 1049 dfrag->already_sent -= delta; 1050 1051 dfrag_uncharge(sk, delta); 1052 cleaned = true; 1053 } 1054 1055 /* all retransmitted data acked, recovery completed */ 1056 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 1057 msk->recovery = false; 1058 1059 out: 1060 if (cleaned && tcp_under_memory_pressure(sk)) 1061 __mptcp_mem_reclaim_partial(sk); 1062 1063 if (snd_una == READ_ONCE(msk->snd_nxt) && 1064 snd_una == READ_ONCE(msk->write_seq)) { 1065 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 1066 mptcp_stop_timer(sk); 1067 } else { 1068 mptcp_reset_timer(sk); 1069 } 1070 } 1071 1072 static void __mptcp_clean_una_wakeup(struct sock *sk) 1073 { 1074 lockdep_assert_held_once(&sk->sk_lock.slock); 1075 1076 __mptcp_clean_una(sk); 1077 mptcp_write_space(sk); 1078 } 1079 1080 static void mptcp_clean_una_wakeup(struct sock *sk) 1081 { 1082 mptcp_data_lock(sk); 1083 __mptcp_clean_una_wakeup(sk); 1084 mptcp_data_unlock(sk); 1085 } 1086 1087 static void mptcp_enter_memory_pressure(struct sock *sk) 1088 { 1089 struct mptcp_subflow_context *subflow; 1090 struct mptcp_sock *msk = mptcp_sk(sk); 1091 bool first = true; 1092 1093 sk_stream_moderate_sndbuf(sk); 1094 mptcp_for_each_subflow(msk, subflow) { 1095 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1096 1097 if (first) 1098 tcp_enter_memory_pressure(ssk); 1099 sk_stream_moderate_sndbuf(ssk); 1100 first = false; 1101 } 1102 } 1103 1104 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1105 * data 1106 */ 1107 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1108 { 1109 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1110 pfrag, sk->sk_allocation))) 1111 return true; 1112 1113 mptcp_enter_memory_pressure(sk); 1114 return false; 1115 } 1116 1117 static struct mptcp_data_frag * 1118 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1119 int orig_offset) 1120 { 1121 int offset = ALIGN(orig_offset, sizeof(long)); 1122 struct mptcp_data_frag *dfrag; 1123 1124 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1125 dfrag->data_len = 0; 1126 dfrag->data_seq = msk->write_seq; 1127 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1128 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1129 dfrag->already_sent = 0; 1130 dfrag->page = pfrag->page; 1131 1132 return dfrag; 1133 } 1134 1135 struct mptcp_sendmsg_info { 1136 int mss_now; 1137 int size_goal; 1138 u16 limit; 1139 u16 sent; 1140 unsigned int flags; 1141 bool data_lock_held; 1142 }; 1143 1144 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq, 1145 int avail_size) 1146 { 1147 u64 window_end = mptcp_wnd_end(msk); 1148 1149 if (__mptcp_check_fallback(msk)) 1150 return avail_size; 1151 1152 if (!before64(data_seq + avail_size, window_end)) { 1153 u64 allowed_size = window_end - data_seq; 1154 1155 return min_t(unsigned int, allowed_size, avail_size); 1156 } 1157 1158 return avail_size; 1159 } 1160 1161 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1162 { 1163 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1164 1165 if (!mpext) 1166 return false; 1167 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1168 return true; 1169 } 1170 1171 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1172 { 1173 struct sk_buff *skb; 1174 1175 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1176 if (likely(skb)) { 1177 if (likely(__mptcp_add_ext(skb, gfp))) { 1178 skb_reserve(skb, MAX_TCP_HEADER); 1179 skb->ip_summed = CHECKSUM_PARTIAL; 1180 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1181 return skb; 1182 } 1183 __kfree_skb(skb); 1184 } else { 1185 mptcp_enter_memory_pressure(sk); 1186 } 1187 return NULL; 1188 } 1189 1190 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1191 { 1192 struct sk_buff *skb; 1193 1194 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1195 if (!skb) 1196 return NULL; 1197 1198 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1199 tcp_skb_entail(ssk, skb); 1200 return skb; 1201 } 1202 tcp_skb_tsorted_anchor_cleanup(skb); 1203 kfree_skb(skb); 1204 return NULL; 1205 } 1206 1207 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1208 { 1209 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1210 1211 if (unlikely(tcp_under_memory_pressure(sk))) { 1212 if (data_lock_held) 1213 __mptcp_mem_reclaim_partial(sk); 1214 else 1215 mptcp_mem_reclaim_partial(sk); 1216 } 1217 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1218 } 1219 1220 /* note: this always recompute the csum on the whole skb, even 1221 * if we just appended a single frag. More status info needed 1222 */ 1223 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1224 { 1225 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1226 __wsum csum = ~csum_unfold(mpext->csum); 1227 int offset = skb->len - added; 1228 1229 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1230 } 1231 1232 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1233 struct sock *ssk, 1234 struct mptcp_ext *mpext) 1235 { 1236 if (!mpext) 1237 return; 1238 1239 mpext->infinite_map = 1; 1240 mpext->data_len = 0; 1241 1242 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX); 1243 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1244 pr_fallback(msk); 1245 __mptcp_do_fallback(msk); 1246 } 1247 1248 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1249 struct mptcp_data_frag *dfrag, 1250 struct mptcp_sendmsg_info *info) 1251 { 1252 u64 data_seq = dfrag->data_seq + info->sent; 1253 int offset = dfrag->offset + info->sent; 1254 struct mptcp_sock *msk = mptcp_sk(sk); 1255 bool zero_window_probe = false; 1256 struct mptcp_ext *mpext = NULL; 1257 bool can_coalesce = false; 1258 bool reuse_skb = true; 1259 struct sk_buff *skb; 1260 size_t copy; 1261 int i; 1262 1263 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u", 1264 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1265 1266 if (WARN_ON_ONCE(info->sent > info->limit || 1267 info->limit > dfrag->data_len)) 1268 return 0; 1269 1270 /* compute send limit */ 1271 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1272 copy = info->size_goal; 1273 1274 skb = tcp_write_queue_tail(ssk); 1275 if (skb && copy > skb->len) { 1276 /* Limit the write to the size available in the 1277 * current skb, if any, so that we create at most a new skb. 1278 * Explicitly tells TCP internals to avoid collapsing on later 1279 * queue management operation, to avoid breaking the ext <-> 1280 * SSN association set here 1281 */ 1282 mpext = skb_ext_find(skb, SKB_EXT_MPTCP); 1283 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1284 TCP_SKB_CB(skb)->eor = 1; 1285 goto alloc_skb; 1286 } 1287 1288 i = skb_shinfo(skb)->nr_frags; 1289 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1290 if (!can_coalesce && i >= sysctl_max_skb_frags) { 1291 tcp_mark_push(tcp_sk(ssk), skb); 1292 goto alloc_skb; 1293 } 1294 1295 copy -= skb->len; 1296 } else { 1297 alloc_skb: 1298 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1299 if (!skb) 1300 return -ENOMEM; 1301 1302 i = skb_shinfo(skb)->nr_frags; 1303 reuse_skb = false; 1304 mpext = skb_ext_find(skb, SKB_EXT_MPTCP); 1305 } 1306 1307 /* Zero window and all data acked? Probe. */ 1308 copy = mptcp_check_allowed_size(msk, data_seq, copy); 1309 if (copy == 0) { 1310 u64 snd_una = READ_ONCE(msk->snd_una); 1311 1312 if (snd_una != msk->snd_nxt) { 1313 tcp_remove_empty_skb(ssk); 1314 return 0; 1315 } 1316 1317 zero_window_probe = true; 1318 data_seq = snd_una - 1; 1319 copy = 1; 1320 1321 /* all mptcp-level data is acked, no skbs should be present into the 1322 * ssk write queue 1323 */ 1324 WARN_ON_ONCE(reuse_skb); 1325 } 1326 1327 copy = min_t(size_t, copy, info->limit - info->sent); 1328 if (!sk_wmem_schedule(ssk, copy)) { 1329 tcp_remove_empty_skb(ssk); 1330 return -ENOMEM; 1331 } 1332 1333 if (can_coalesce) { 1334 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1335 } else { 1336 get_page(dfrag->page); 1337 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1338 } 1339 1340 skb->len += copy; 1341 skb->data_len += copy; 1342 skb->truesize += copy; 1343 sk_wmem_queued_add(ssk, copy); 1344 sk_mem_charge(ssk, copy); 1345 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1346 TCP_SKB_CB(skb)->end_seq += copy; 1347 tcp_skb_pcount_set(skb, 0); 1348 1349 /* on skb reuse we just need to update the DSS len */ 1350 if (reuse_skb) { 1351 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1352 mpext->data_len += copy; 1353 WARN_ON_ONCE(zero_window_probe); 1354 goto out; 1355 } 1356 1357 memset(mpext, 0, sizeof(*mpext)); 1358 mpext->data_seq = data_seq; 1359 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1360 mpext->data_len = copy; 1361 mpext->use_map = 1; 1362 mpext->dsn64 = 1; 1363 1364 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d", 1365 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1366 mpext->dsn64); 1367 1368 if (zero_window_probe) { 1369 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1370 mpext->frozen = 1; 1371 if (READ_ONCE(msk->csum_enabled)) 1372 mptcp_update_data_checksum(skb, copy); 1373 tcp_push_pending_frames(ssk); 1374 return 0; 1375 } 1376 out: 1377 if (READ_ONCE(msk->csum_enabled)) 1378 mptcp_update_data_checksum(skb, copy); 1379 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1380 mptcp_update_infinite_map(msk, ssk, mpext); 1381 trace_mptcp_sendmsg_frag(mpext); 1382 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1383 return copy; 1384 } 1385 1386 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1387 sizeof(struct tcphdr) - \ 1388 MAX_TCP_OPTION_SPACE - \ 1389 sizeof(struct ipv6hdr) - \ 1390 sizeof(struct frag_hdr)) 1391 1392 struct subflow_send_info { 1393 struct sock *ssk; 1394 u64 linger_time; 1395 }; 1396 1397 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1398 { 1399 if (!subflow->stale) 1400 return; 1401 1402 subflow->stale = 0; 1403 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1404 } 1405 1406 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1407 { 1408 if (unlikely(subflow->stale)) { 1409 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1410 1411 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1412 return false; 1413 1414 mptcp_subflow_set_active(subflow); 1415 } 1416 return __mptcp_subflow_active(subflow); 1417 } 1418 1419 #define SSK_MODE_ACTIVE 0 1420 #define SSK_MODE_BACKUP 1 1421 #define SSK_MODE_MAX 2 1422 1423 /* implement the mptcp packet scheduler; 1424 * returns the subflow that will transmit the next DSS 1425 * additionally updates the rtx timeout 1426 */ 1427 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1428 { 1429 struct subflow_send_info send_info[SSK_MODE_MAX]; 1430 struct mptcp_subflow_context *subflow; 1431 struct sock *sk = (struct sock *)msk; 1432 u32 pace, burst, wmem; 1433 int i, nr_active = 0; 1434 struct sock *ssk; 1435 u64 linger_time; 1436 long tout = 0; 1437 1438 sock_owned_by_me(sk); 1439 1440 if (__mptcp_check_fallback(msk)) { 1441 if (!msk->first) 1442 return NULL; 1443 return sk_stream_memory_free(msk->first) ? msk->first : NULL; 1444 } 1445 1446 /* re-use last subflow, if the burst allow that */ 1447 if (msk->last_snd && msk->snd_burst > 0 && 1448 sk_stream_memory_free(msk->last_snd) && 1449 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) { 1450 mptcp_set_timeout(sk); 1451 return msk->last_snd; 1452 } 1453 1454 /* pick the subflow with the lower wmem/wspace ratio */ 1455 for (i = 0; i < SSK_MODE_MAX; ++i) { 1456 send_info[i].ssk = NULL; 1457 send_info[i].linger_time = -1; 1458 } 1459 1460 mptcp_for_each_subflow(msk, subflow) { 1461 trace_mptcp_subflow_get_send(subflow); 1462 ssk = mptcp_subflow_tcp_sock(subflow); 1463 if (!mptcp_subflow_active(subflow)) 1464 continue; 1465 1466 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1467 nr_active += !subflow->backup; 1468 pace = subflow->avg_pacing_rate; 1469 if (unlikely(!pace)) { 1470 /* init pacing rate from socket */ 1471 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1472 pace = subflow->avg_pacing_rate; 1473 if (!pace) 1474 continue; 1475 } 1476 1477 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1478 if (linger_time < send_info[subflow->backup].linger_time) { 1479 send_info[subflow->backup].ssk = ssk; 1480 send_info[subflow->backup].linger_time = linger_time; 1481 } 1482 } 1483 __mptcp_set_timeout(sk, tout); 1484 1485 /* pick the best backup if no other subflow is active */ 1486 if (!nr_active) 1487 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1488 1489 /* According to the blest algorithm, to avoid HoL blocking for the 1490 * faster flow, we need to: 1491 * - estimate the faster flow linger time 1492 * - use the above to estimate the amount of byte transferred 1493 * by the faster flow 1494 * - check that the amount of queued data is greter than the above, 1495 * otherwise do not use the picked, slower, subflow 1496 * We select the subflow with the shorter estimated time to flush 1497 * the queued mem, which basically ensure the above. We just need 1498 * to check that subflow has a non empty cwin. 1499 */ 1500 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1501 if (!ssk || !sk_stream_memory_free(ssk) || !tcp_sk(ssk)->snd_wnd) 1502 return NULL; 1503 1504 burst = min_t(int, MPTCP_SEND_BURST_SIZE, tcp_sk(ssk)->snd_wnd); 1505 wmem = READ_ONCE(ssk->sk_wmem_queued); 1506 subflow = mptcp_subflow_ctx(ssk); 1507 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1508 READ_ONCE(ssk->sk_pacing_rate) * burst, 1509 burst + wmem); 1510 msk->last_snd = ssk; 1511 msk->snd_burst = burst; 1512 return ssk; 1513 } 1514 1515 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1516 { 1517 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1518 release_sock(ssk); 1519 } 1520 1521 static void mptcp_update_post_push(struct mptcp_sock *msk, 1522 struct mptcp_data_frag *dfrag, 1523 u32 sent) 1524 { 1525 u64 snd_nxt_new = dfrag->data_seq; 1526 1527 dfrag->already_sent += sent; 1528 1529 msk->snd_burst -= sent; 1530 1531 snd_nxt_new += dfrag->already_sent; 1532 1533 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1534 * is recovering after a failover. In that event, this re-sends 1535 * old segments. 1536 * 1537 * Thus compute snd_nxt_new candidate based on 1538 * the dfrag->data_seq that was sent and the data 1539 * that has been handed to the subflow for transmission 1540 * and skip update in case it was old dfrag. 1541 */ 1542 if (likely(after64(snd_nxt_new, msk->snd_nxt))) 1543 msk->snd_nxt = snd_nxt_new; 1544 } 1545 1546 void mptcp_check_and_set_pending(struct sock *sk) 1547 { 1548 if (mptcp_send_head(sk)) 1549 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING); 1550 } 1551 1552 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1553 { 1554 struct sock *prev_ssk = NULL, *ssk = NULL; 1555 struct mptcp_sock *msk = mptcp_sk(sk); 1556 struct mptcp_sendmsg_info info = { 1557 .flags = flags, 1558 }; 1559 struct mptcp_data_frag *dfrag; 1560 int len, copied = 0; 1561 1562 while ((dfrag = mptcp_send_head(sk))) { 1563 info.sent = dfrag->already_sent; 1564 info.limit = dfrag->data_len; 1565 len = dfrag->data_len - dfrag->already_sent; 1566 while (len > 0) { 1567 int ret = 0; 1568 1569 prev_ssk = ssk; 1570 ssk = mptcp_subflow_get_send(msk); 1571 1572 /* First check. If the ssk has changed since 1573 * the last round, release prev_ssk 1574 */ 1575 if (ssk != prev_ssk && prev_ssk) 1576 mptcp_push_release(prev_ssk, &info); 1577 if (!ssk) 1578 goto out; 1579 1580 /* Need to lock the new subflow only if different 1581 * from the previous one, otherwise we are still 1582 * helding the relevant lock 1583 */ 1584 if (ssk != prev_ssk) 1585 lock_sock(ssk); 1586 1587 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 1588 if (ret <= 0) { 1589 mptcp_push_release(ssk, &info); 1590 goto out; 1591 } 1592 1593 info.sent += ret; 1594 copied += ret; 1595 len -= ret; 1596 1597 mptcp_update_post_push(msk, dfrag, ret); 1598 } 1599 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1600 } 1601 1602 /* at this point we held the socket lock for the last subflow we used */ 1603 if (ssk) 1604 mptcp_push_release(ssk, &info); 1605 1606 out: 1607 /* ensure the rtx timer is running */ 1608 mptcp_data_lock(sk); 1609 if (!mptcp_timer_pending(sk)) 1610 mptcp_reset_timer(sk); 1611 mptcp_data_unlock(sk); 1612 if (copied) 1613 __mptcp_check_send_data_fin(sk); 1614 } 1615 1616 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk) 1617 { 1618 struct mptcp_sock *msk = mptcp_sk(sk); 1619 struct mptcp_sendmsg_info info = { 1620 .data_lock_held = true, 1621 }; 1622 struct mptcp_data_frag *dfrag; 1623 struct sock *xmit_ssk; 1624 int len, copied = 0; 1625 bool first = true; 1626 1627 info.flags = 0; 1628 while ((dfrag = mptcp_send_head(sk))) { 1629 info.sent = dfrag->already_sent; 1630 info.limit = dfrag->data_len; 1631 len = dfrag->data_len - dfrag->already_sent; 1632 while (len > 0) { 1633 int ret = 0; 1634 1635 /* the caller already invoked the packet scheduler, 1636 * check for a different subflow usage only after 1637 * spooling the first chunk of data 1638 */ 1639 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk)); 1640 if (!xmit_ssk) 1641 goto out; 1642 if (xmit_ssk != ssk) { 1643 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), 1644 MPTCP_DELEGATE_SEND); 1645 goto out; 1646 } 1647 1648 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 1649 if (ret <= 0) 1650 goto out; 1651 1652 info.sent += ret; 1653 copied += ret; 1654 len -= ret; 1655 first = false; 1656 1657 mptcp_update_post_push(msk, dfrag, ret); 1658 } 1659 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1660 } 1661 1662 out: 1663 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1664 * not going to flush it via release_sock() 1665 */ 1666 if (copied) { 1667 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1668 info.size_goal); 1669 if (!mptcp_timer_pending(sk)) 1670 mptcp_reset_timer(sk); 1671 1672 if (msk->snd_data_fin_enable && 1673 msk->snd_nxt + 1 == msk->write_seq) 1674 mptcp_schedule_work(sk); 1675 } 1676 } 1677 1678 static void mptcp_set_nospace(struct sock *sk) 1679 { 1680 /* enable autotune */ 1681 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1682 1683 /* will be cleared on avail space */ 1684 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags); 1685 } 1686 1687 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1688 { 1689 struct mptcp_sock *msk = mptcp_sk(sk); 1690 struct page_frag *pfrag; 1691 size_t copied = 0; 1692 int ret = 0; 1693 long timeo; 1694 1695 /* we don't support FASTOPEN yet */ 1696 if (msg->msg_flags & MSG_FASTOPEN) 1697 return -EOPNOTSUPP; 1698 1699 /* silently ignore everything else */ 1700 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL; 1701 1702 lock_sock(sk); 1703 1704 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1705 1706 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1707 ret = sk_stream_wait_connect(sk, &timeo); 1708 if (ret) 1709 goto out; 1710 } 1711 1712 pfrag = sk_page_frag(sk); 1713 1714 while (msg_data_left(msg)) { 1715 int total_ts, frag_truesize = 0; 1716 struct mptcp_data_frag *dfrag; 1717 bool dfrag_collapsed; 1718 size_t psize, offset; 1719 1720 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) { 1721 ret = -EPIPE; 1722 goto out; 1723 } 1724 1725 /* reuse tail pfrag, if possible, or carve a new one from the 1726 * page allocator 1727 */ 1728 dfrag = mptcp_pending_tail(sk); 1729 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1730 if (!dfrag_collapsed) { 1731 if (!sk_stream_memory_free(sk)) 1732 goto wait_for_memory; 1733 1734 if (!mptcp_page_frag_refill(sk, pfrag)) 1735 goto wait_for_memory; 1736 1737 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1738 frag_truesize = dfrag->overhead; 1739 } 1740 1741 /* we do not bound vs wspace, to allow a single packet. 1742 * memory accounting will prevent execessive memory usage 1743 * anyway 1744 */ 1745 offset = dfrag->offset + dfrag->data_len; 1746 psize = pfrag->size - offset; 1747 psize = min_t(size_t, psize, msg_data_left(msg)); 1748 total_ts = psize + frag_truesize; 1749 1750 if (!sk_wmem_schedule(sk, total_ts)) 1751 goto wait_for_memory; 1752 1753 if (copy_page_from_iter(dfrag->page, offset, psize, 1754 &msg->msg_iter) != psize) { 1755 ret = -EFAULT; 1756 goto out; 1757 } 1758 1759 /* data successfully copied into the write queue */ 1760 sk->sk_forward_alloc -= total_ts; 1761 copied += psize; 1762 dfrag->data_len += psize; 1763 frag_truesize += psize; 1764 pfrag->offset += frag_truesize; 1765 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1766 1767 /* charge data on mptcp pending queue to the msk socket 1768 * Note: we charge such data both to sk and ssk 1769 */ 1770 sk_wmem_queued_add(sk, frag_truesize); 1771 if (!dfrag_collapsed) { 1772 get_page(dfrag->page); 1773 list_add_tail(&dfrag->list, &msk->rtx_queue); 1774 if (!msk->first_pending) 1775 WRITE_ONCE(msk->first_pending, dfrag); 1776 } 1777 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk, 1778 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1779 !dfrag_collapsed); 1780 1781 continue; 1782 1783 wait_for_memory: 1784 mptcp_set_nospace(sk); 1785 __mptcp_push_pending(sk, msg->msg_flags); 1786 ret = sk_stream_wait_memory(sk, &timeo); 1787 if (ret) 1788 goto out; 1789 } 1790 1791 if (copied) 1792 __mptcp_push_pending(sk, msg->msg_flags); 1793 1794 out: 1795 release_sock(sk); 1796 return copied ? : ret; 1797 } 1798 1799 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk, 1800 struct msghdr *msg, 1801 size_t len, int flags, 1802 struct scm_timestamping_internal *tss, 1803 int *cmsg_flags) 1804 { 1805 struct sk_buff *skb, *tmp; 1806 int copied = 0; 1807 1808 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) { 1809 u32 offset = MPTCP_SKB_CB(skb)->offset; 1810 u32 data_len = skb->len - offset; 1811 u32 count = min_t(size_t, len - copied, data_len); 1812 int err; 1813 1814 if (!(flags & MSG_TRUNC)) { 1815 err = skb_copy_datagram_msg(skb, offset, msg, count); 1816 if (unlikely(err < 0)) { 1817 if (!copied) 1818 return err; 1819 break; 1820 } 1821 } 1822 1823 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 1824 tcp_update_recv_tstamps(skb, tss); 1825 *cmsg_flags |= MPTCP_CMSG_TS; 1826 } 1827 1828 copied += count; 1829 1830 if (count < data_len) { 1831 if (!(flags & MSG_PEEK)) { 1832 MPTCP_SKB_CB(skb)->offset += count; 1833 MPTCP_SKB_CB(skb)->map_seq += count; 1834 } 1835 break; 1836 } 1837 1838 if (!(flags & MSG_PEEK)) { 1839 /* we will bulk release the skb memory later */ 1840 skb->destructor = NULL; 1841 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize); 1842 __skb_unlink(skb, &msk->receive_queue); 1843 __kfree_skb(skb); 1844 } 1845 1846 if (copied >= len) 1847 break; 1848 } 1849 1850 return copied; 1851 } 1852 1853 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1854 * 1855 * Only difference: Use highest rtt estimate of the subflows in use. 1856 */ 1857 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1858 { 1859 struct mptcp_subflow_context *subflow; 1860 struct sock *sk = (struct sock *)msk; 1861 u32 time, advmss = 1; 1862 u64 rtt_us, mstamp; 1863 1864 sock_owned_by_me(sk); 1865 1866 if (copied <= 0) 1867 return; 1868 1869 msk->rcvq_space.copied += copied; 1870 1871 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1872 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1873 1874 rtt_us = msk->rcvq_space.rtt_us; 1875 if (rtt_us && time < (rtt_us >> 3)) 1876 return; 1877 1878 rtt_us = 0; 1879 mptcp_for_each_subflow(msk, subflow) { 1880 const struct tcp_sock *tp; 1881 u64 sf_rtt_us; 1882 u32 sf_advmss; 1883 1884 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1885 1886 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1887 sf_advmss = READ_ONCE(tp->advmss); 1888 1889 rtt_us = max(sf_rtt_us, rtt_us); 1890 advmss = max(sf_advmss, advmss); 1891 } 1892 1893 msk->rcvq_space.rtt_us = rtt_us; 1894 if (time < (rtt_us >> 3) || rtt_us == 0) 1895 return; 1896 1897 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 1898 goto new_measure; 1899 1900 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && 1901 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 1902 int rcvmem, rcvbuf; 1903 u64 rcvwin, grow; 1904 1905 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 1906 1907 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 1908 1909 do_div(grow, msk->rcvq_space.space); 1910 rcvwin += (grow << 1); 1911 1912 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER); 1913 while (tcp_win_from_space(sk, rcvmem) < advmss) 1914 rcvmem += 128; 1915 1916 do_div(rcvwin, advmss); 1917 rcvbuf = min_t(u64, rcvwin * rcvmem, 1918 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 1919 1920 if (rcvbuf > sk->sk_rcvbuf) { 1921 u32 window_clamp; 1922 1923 window_clamp = tcp_win_from_space(sk, rcvbuf); 1924 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 1925 1926 /* Make subflows follow along. If we do not do this, we 1927 * get drops at subflow level if skbs can't be moved to 1928 * the mptcp rx queue fast enough (announced rcv_win can 1929 * exceed ssk->sk_rcvbuf). 1930 */ 1931 mptcp_for_each_subflow(msk, subflow) { 1932 struct sock *ssk; 1933 bool slow; 1934 1935 ssk = mptcp_subflow_tcp_sock(subflow); 1936 slow = lock_sock_fast(ssk); 1937 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 1938 tcp_sk(ssk)->window_clamp = window_clamp; 1939 tcp_cleanup_rbuf(ssk, 1); 1940 unlock_sock_fast(ssk, slow); 1941 } 1942 } 1943 } 1944 1945 msk->rcvq_space.space = msk->rcvq_space.copied; 1946 new_measure: 1947 msk->rcvq_space.copied = 0; 1948 msk->rcvq_space.time = mstamp; 1949 } 1950 1951 static void __mptcp_update_rmem(struct sock *sk) 1952 { 1953 struct mptcp_sock *msk = mptcp_sk(sk); 1954 1955 if (!msk->rmem_released) 1956 return; 1957 1958 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc); 1959 mptcp_rmem_uncharge(sk, msk->rmem_released); 1960 WRITE_ONCE(msk->rmem_released, 0); 1961 } 1962 1963 static void __mptcp_splice_receive_queue(struct sock *sk) 1964 { 1965 struct mptcp_sock *msk = mptcp_sk(sk); 1966 1967 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue); 1968 } 1969 1970 static bool __mptcp_move_skbs(struct mptcp_sock *msk) 1971 { 1972 struct sock *sk = (struct sock *)msk; 1973 unsigned int moved = 0; 1974 bool ret, done; 1975 1976 do { 1977 struct sock *ssk = mptcp_subflow_recv_lookup(msk); 1978 bool slowpath; 1979 1980 /* we can have data pending in the subflows only if the msk 1981 * receive buffer was full at subflow_data_ready() time, 1982 * that is an unlikely slow path. 1983 */ 1984 if (likely(!ssk)) 1985 break; 1986 1987 slowpath = lock_sock_fast(ssk); 1988 mptcp_data_lock(sk); 1989 __mptcp_update_rmem(sk); 1990 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 1991 mptcp_data_unlock(sk); 1992 1993 if (unlikely(ssk->sk_err)) 1994 __mptcp_error_report(sk); 1995 unlock_sock_fast(ssk, slowpath); 1996 } while (!done); 1997 1998 /* acquire the data lock only if some input data is pending */ 1999 ret = moved > 0; 2000 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) || 2001 !skb_queue_empty_lockless(&sk->sk_receive_queue)) { 2002 mptcp_data_lock(sk); 2003 __mptcp_update_rmem(sk); 2004 ret |= __mptcp_ofo_queue(msk); 2005 __mptcp_splice_receive_queue(sk); 2006 mptcp_data_unlock(sk); 2007 } 2008 if (ret) 2009 mptcp_check_data_fin((struct sock *)msk); 2010 return !skb_queue_empty(&msk->receive_queue); 2011 } 2012 2013 static unsigned int mptcp_inq_hint(const struct sock *sk) 2014 { 2015 const struct mptcp_sock *msk = mptcp_sk(sk); 2016 const struct sk_buff *skb; 2017 2018 skb = skb_peek(&msk->receive_queue); 2019 if (skb) { 2020 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 2021 2022 if (hint_val >= INT_MAX) 2023 return INT_MAX; 2024 2025 return (unsigned int)hint_val; 2026 } 2027 2028 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2029 return 1; 2030 2031 return 0; 2032 } 2033 2034 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2035 int flags, int *addr_len) 2036 { 2037 struct mptcp_sock *msk = mptcp_sk(sk); 2038 struct scm_timestamping_internal tss; 2039 int copied = 0, cmsg_flags = 0; 2040 int target; 2041 long timeo; 2042 2043 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2044 if (unlikely(flags & MSG_ERRQUEUE)) 2045 return inet_recv_error(sk, msg, len, addr_len); 2046 2047 lock_sock(sk); 2048 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2049 copied = -ENOTCONN; 2050 goto out_err; 2051 } 2052 2053 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2054 2055 len = min_t(size_t, len, INT_MAX); 2056 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2057 2058 if (unlikely(msk->recvmsg_inq)) 2059 cmsg_flags = MPTCP_CMSG_INQ; 2060 2061 while (copied < len) { 2062 int bytes_read; 2063 2064 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags); 2065 if (unlikely(bytes_read < 0)) { 2066 if (!copied) 2067 copied = bytes_read; 2068 goto out_err; 2069 } 2070 2071 copied += bytes_read; 2072 2073 /* be sure to advertise window change */ 2074 mptcp_cleanup_rbuf(msk); 2075 2076 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk)) 2077 continue; 2078 2079 /* only the master socket status is relevant here. The exit 2080 * conditions mirror closely tcp_recvmsg() 2081 */ 2082 if (copied >= target) 2083 break; 2084 2085 if (copied) { 2086 if (sk->sk_err || 2087 sk->sk_state == TCP_CLOSE || 2088 (sk->sk_shutdown & RCV_SHUTDOWN) || 2089 !timeo || 2090 signal_pending(current)) 2091 break; 2092 } else { 2093 if (sk->sk_err) { 2094 copied = sock_error(sk); 2095 break; 2096 } 2097 2098 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags)) 2099 mptcp_check_for_eof(msk); 2100 2101 if (sk->sk_shutdown & RCV_SHUTDOWN) { 2102 /* race breaker: the shutdown could be after the 2103 * previous receive queue check 2104 */ 2105 if (__mptcp_move_skbs(msk)) 2106 continue; 2107 break; 2108 } 2109 2110 if (sk->sk_state == TCP_CLOSE) { 2111 copied = -ENOTCONN; 2112 break; 2113 } 2114 2115 if (!timeo) { 2116 copied = -EAGAIN; 2117 break; 2118 } 2119 2120 if (signal_pending(current)) { 2121 copied = sock_intr_errno(timeo); 2122 break; 2123 } 2124 } 2125 2126 pr_debug("block timeout %ld", timeo); 2127 sk_wait_data(sk, &timeo, NULL); 2128 } 2129 2130 out_err: 2131 if (cmsg_flags && copied >= 0) { 2132 if (cmsg_flags & MPTCP_CMSG_TS) 2133 tcp_recv_timestamp(msg, sk, &tss); 2134 2135 if (cmsg_flags & MPTCP_CMSG_INQ) { 2136 unsigned int inq = mptcp_inq_hint(sk); 2137 2138 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2139 } 2140 } 2141 2142 pr_debug("msk=%p rx queue empty=%d:%d copied=%d", 2143 msk, skb_queue_empty_lockless(&sk->sk_receive_queue), 2144 skb_queue_empty(&msk->receive_queue), copied); 2145 if (!(flags & MSG_PEEK)) 2146 mptcp_rcv_space_adjust(msk, copied); 2147 2148 release_sock(sk); 2149 return copied; 2150 } 2151 2152 static void mptcp_retransmit_timer(struct timer_list *t) 2153 { 2154 struct inet_connection_sock *icsk = from_timer(icsk, t, 2155 icsk_retransmit_timer); 2156 struct sock *sk = &icsk->icsk_inet.sk; 2157 struct mptcp_sock *msk = mptcp_sk(sk); 2158 2159 bh_lock_sock(sk); 2160 if (!sock_owned_by_user(sk)) { 2161 /* we need a process context to retransmit */ 2162 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2163 mptcp_schedule_work(sk); 2164 } else { 2165 /* delegate our work to tcp_release_cb() */ 2166 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2167 } 2168 bh_unlock_sock(sk); 2169 sock_put(sk); 2170 } 2171 2172 static struct mptcp_subflow_context * 2173 mp_fail_response_expect_subflow(struct mptcp_sock *msk) 2174 { 2175 struct mptcp_subflow_context *subflow, *ret = NULL; 2176 2177 mptcp_for_each_subflow(msk, subflow) { 2178 if (READ_ONCE(subflow->mp_fail_response_expect)) { 2179 ret = subflow; 2180 break; 2181 } 2182 } 2183 2184 return ret; 2185 } 2186 2187 static void mptcp_check_mp_fail_response(struct mptcp_sock *msk) 2188 { 2189 struct mptcp_subflow_context *subflow; 2190 struct sock *sk = (struct sock *)msk; 2191 2192 bh_lock_sock(sk); 2193 subflow = mp_fail_response_expect_subflow(msk); 2194 if (subflow) 2195 __set_bit(MPTCP_FAIL_NO_RESPONSE, &msk->flags); 2196 bh_unlock_sock(sk); 2197 } 2198 2199 static void mptcp_timeout_timer(struct timer_list *t) 2200 { 2201 struct sock *sk = from_timer(sk, t, sk_timer); 2202 2203 mptcp_check_mp_fail_response(mptcp_sk(sk)); 2204 mptcp_schedule_work(sk); 2205 sock_put(sk); 2206 } 2207 2208 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2209 * level. 2210 * 2211 * A backup subflow is returned only if that is the only kind available. 2212 */ 2213 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2214 { 2215 struct sock *backup = NULL, *pick = NULL; 2216 struct mptcp_subflow_context *subflow; 2217 int min_stale_count = INT_MAX; 2218 2219 sock_owned_by_me((const struct sock *)msk); 2220 2221 if (__mptcp_check_fallback(msk)) 2222 return NULL; 2223 2224 mptcp_for_each_subflow(msk, subflow) { 2225 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2226 2227 if (!__mptcp_subflow_active(subflow)) 2228 continue; 2229 2230 /* still data outstanding at TCP level? skip this */ 2231 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2232 mptcp_pm_subflow_chk_stale(msk, ssk); 2233 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2234 continue; 2235 } 2236 2237 if (subflow->backup) { 2238 if (!backup) 2239 backup = ssk; 2240 continue; 2241 } 2242 2243 if (!pick) 2244 pick = ssk; 2245 } 2246 2247 if (pick) 2248 return pick; 2249 2250 /* use backup only if there are no progresses anywhere */ 2251 return min_stale_count > 1 ? backup : NULL; 2252 } 2253 2254 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk) 2255 { 2256 if (msk->subflow) { 2257 iput(SOCK_INODE(msk->subflow)); 2258 msk->subflow = NULL; 2259 } 2260 } 2261 2262 bool __mptcp_retransmit_pending_data(struct sock *sk) 2263 { 2264 struct mptcp_data_frag *cur, *rtx_head; 2265 struct mptcp_sock *msk = mptcp_sk(sk); 2266 2267 if (__mptcp_check_fallback(mptcp_sk(sk))) 2268 return false; 2269 2270 if (tcp_rtx_and_write_queues_empty(sk)) 2271 return false; 2272 2273 /* the closing socket has some data untransmitted and/or unacked: 2274 * some data in the mptcp rtx queue has not really xmitted yet. 2275 * keep it simple and re-inject the whole mptcp level rtx queue 2276 */ 2277 mptcp_data_lock(sk); 2278 __mptcp_clean_una_wakeup(sk); 2279 rtx_head = mptcp_rtx_head(sk); 2280 if (!rtx_head) { 2281 mptcp_data_unlock(sk); 2282 return false; 2283 } 2284 2285 msk->recovery_snd_nxt = msk->snd_nxt; 2286 msk->recovery = true; 2287 mptcp_data_unlock(sk); 2288 2289 msk->first_pending = rtx_head; 2290 msk->snd_burst = 0; 2291 2292 /* be sure to clear the "sent status" on all re-injected fragments */ 2293 list_for_each_entry(cur, &msk->rtx_queue, list) { 2294 if (!cur->already_sent) 2295 break; 2296 cur->already_sent = 0; 2297 } 2298 2299 return true; 2300 } 2301 2302 /* flags for __mptcp_close_ssk() */ 2303 #define MPTCP_CF_PUSH BIT(1) 2304 #define MPTCP_CF_FASTCLOSE BIT(2) 2305 2306 /* subflow sockets can be either outgoing (connect) or incoming 2307 * (accept). 2308 * 2309 * Outgoing subflows use in-kernel sockets. 2310 * Incoming subflows do not have their own 'struct socket' allocated, 2311 * so we need to use tcp_close() after detaching them from the mptcp 2312 * parent socket. 2313 */ 2314 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2315 struct mptcp_subflow_context *subflow, 2316 unsigned int flags) 2317 { 2318 struct mptcp_sock *msk = mptcp_sk(sk); 2319 bool need_push, dispose_it; 2320 2321 dispose_it = !msk->subflow || ssk != msk->subflow->sk; 2322 if (dispose_it) 2323 list_del(&subflow->node); 2324 2325 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2326 2327 if (flags & MPTCP_CF_FASTCLOSE) 2328 subflow->send_fastclose = 1; 2329 2330 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2331 if (!dispose_it) { 2332 tcp_disconnect(ssk, 0); 2333 msk->subflow->state = SS_UNCONNECTED; 2334 mptcp_subflow_ctx_reset(subflow); 2335 release_sock(ssk); 2336 2337 goto out; 2338 } 2339 2340 /* if we are invoked by the msk cleanup code, the subflow is 2341 * already orphaned 2342 */ 2343 if (ssk->sk_socket) 2344 sock_orphan(ssk); 2345 2346 subflow->disposable = 1; 2347 2348 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2349 * the ssk has been already destroyed, we just need to release the 2350 * reference owned by msk; 2351 */ 2352 if (!inet_csk(ssk)->icsk_ulp_ops) { 2353 kfree_rcu(subflow, rcu); 2354 } else { 2355 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2356 __tcp_close(ssk, 0); 2357 2358 /* close acquired an extra ref */ 2359 __sock_put(ssk); 2360 } 2361 release_sock(ssk); 2362 2363 sock_put(ssk); 2364 2365 if (ssk == msk->first) 2366 msk->first = NULL; 2367 2368 out: 2369 if (ssk == msk->last_snd) 2370 msk->last_snd = NULL; 2371 2372 if (need_push) 2373 __mptcp_push_pending(sk, 0); 2374 } 2375 2376 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2377 struct mptcp_subflow_context *subflow) 2378 { 2379 if (sk->sk_state == TCP_ESTABLISHED) 2380 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2381 2382 /* subflow aborted before reaching the fully_established status 2383 * attempt the creation of the next subflow 2384 */ 2385 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow); 2386 2387 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2388 } 2389 2390 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2391 { 2392 return 0; 2393 } 2394 2395 static void __mptcp_close_subflow(struct mptcp_sock *msk) 2396 { 2397 struct mptcp_subflow_context *subflow, *tmp; 2398 2399 might_sleep(); 2400 2401 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) { 2402 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2403 2404 if (inet_sk_state_load(ssk) != TCP_CLOSE) 2405 continue; 2406 2407 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2408 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2409 continue; 2410 2411 mptcp_close_ssk((struct sock *)msk, ssk, subflow); 2412 } 2413 } 2414 2415 static bool mptcp_check_close_timeout(const struct sock *sk) 2416 { 2417 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp; 2418 struct mptcp_subflow_context *subflow; 2419 2420 if (delta >= TCP_TIMEWAIT_LEN) 2421 return true; 2422 2423 /* if all subflows are in closed status don't bother with additional 2424 * timeout 2425 */ 2426 mptcp_for_each_subflow(mptcp_sk(sk), subflow) { 2427 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) != 2428 TCP_CLOSE) 2429 return false; 2430 } 2431 return true; 2432 } 2433 2434 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2435 { 2436 struct mptcp_subflow_context *subflow, *tmp; 2437 struct sock *sk = &msk->sk.icsk_inet.sk; 2438 2439 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2440 return; 2441 2442 mptcp_token_destroy(msk); 2443 2444 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) { 2445 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2446 bool slow; 2447 2448 slow = lock_sock_fast(tcp_sk); 2449 if (tcp_sk->sk_state != TCP_CLOSE) { 2450 tcp_send_active_reset(tcp_sk, GFP_ATOMIC); 2451 tcp_set_state(tcp_sk, TCP_CLOSE); 2452 } 2453 unlock_sock_fast(tcp_sk, slow); 2454 } 2455 2456 inet_sk_state_store(sk, TCP_CLOSE); 2457 sk->sk_shutdown = SHUTDOWN_MASK; 2458 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2459 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2460 2461 mptcp_close_wake_up(sk); 2462 } 2463 2464 static void __mptcp_retrans(struct sock *sk) 2465 { 2466 struct mptcp_sock *msk = mptcp_sk(sk); 2467 struct mptcp_sendmsg_info info = {}; 2468 struct mptcp_data_frag *dfrag; 2469 size_t copied = 0; 2470 struct sock *ssk; 2471 int ret; 2472 2473 mptcp_clean_una_wakeup(sk); 2474 2475 /* first check ssk: need to kick "stale" logic */ 2476 ssk = mptcp_subflow_get_retrans(msk); 2477 dfrag = mptcp_rtx_head(sk); 2478 if (!dfrag) { 2479 if (mptcp_data_fin_enabled(msk)) { 2480 struct inet_connection_sock *icsk = inet_csk(sk); 2481 2482 icsk->icsk_retransmits++; 2483 mptcp_set_datafin_timeout(sk); 2484 mptcp_send_ack(msk); 2485 2486 goto reset_timer; 2487 } 2488 2489 if (!mptcp_send_head(sk)) 2490 return; 2491 2492 goto reset_timer; 2493 } 2494 2495 if (!ssk) 2496 goto reset_timer; 2497 2498 lock_sock(ssk); 2499 2500 /* limit retransmission to the bytes already sent on some subflows */ 2501 info.sent = 0; 2502 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent; 2503 while (info.sent < info.limit) { 2504 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2505 if (ret <= 0) 2506 break; 2507 2508 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2509 copied += ret; 2510 info.sent += ret; 2511 } 2512 if (copied) { 2513 dfrag->already_sent = max(dfrag->already_sent, info.sent); 2514 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2515 info.size_goal); 2516 WRITE_ONCE(msk->allow_infinite_fallback, false); 2517 } 2518 2519 release_sock(ssk); 2520 2521 reset_timer: 2522 mptcp_check_and_set_pending(sk); 2523 2524 mptcp_data_lock(sk); 2525 if (!mptcp_timer_pending(sk)) 2526 mptcp_reset_timer(sk); 2527 mptcp_data_unlock(sk); 2528 } 2529 2530 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2531 { 2532 struct mptcp_subflow_context *subflow; 2533 struct sock *ssk; 2534 bool slow; 2535 2536 subflow = mp_fail_response_expect_subflow(msk); 2537 if (subflow) { 2538 pr_debug("MP_FAIL doesn't respond, reset the subflow"); 2539 2540 ssk = mptcp_subflow_tcp_sock(subflow); 2541 slow = lock_sock_fast(ssk); 2542 mptcp_subflow_reset(ssk); 2543 unlock_sock_fast(ssk, slow); 2544 } 2545 } 2546 2547 static void mptcp_worker(struct work_struct *work) 2548 { 2549 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2550 struct sock *sk = &msk->sk.icsk_inet.sk; 2551 int state; 2552 2553 lock_sock(sk); 2554 state = sk->sk_state; 2555 if (unlikely(state == TCP_CLOSE)) 2556 goto unlock; 2557 2558 mptcp_check_data_fin_ack(sk); 2559 2560 mptcp_check_fastclose(msk); 2561 2562 mptcp_pm_nl_work(msk); 2563 2564 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags)) 2565 mptcp_check_for_eof(msk); 2566 2567 __mptcp_check_send_data_fin(sk); 2568 mptcp_check_data_fin(sk); 2569 2570 /* There is no point in keeping around an orphaned sk timedout or 2571 * closed, but we need the msk around to reply to incoming DATA_FIN, 2572 * even if it is orphaned and in FIN_WAIT2 state 2573 */ 2574 if (sock_flag(sk, SOCK_DEAD) && 2575 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) { 2576 inet_sk_state_store(sk, TCP_CLOSE); 2577 __mptcp_destroy_sock(sk); 2578 goto unlock; 2579 } 2580 2581 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2582 __mptcp_close_subflow(msk); 2583 2584 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2585 __mptcp_retrans(sk); 2586 2587 if (test_and_clear_bit(MPTCP_FAIL_NO_RESPONSE, &msk->flags)) 2588 mptcp_mp_fail_no_response(msk); 2589 2590 unlock: 2591 release_sock(sk); 2592 sock_put(sk); 2593 } 2594 2595 static int __mptcp_init_sock(struct sock *sk) 2596 { 2597 struct mptcp_sock *msk = mptcp_sk(sk); 2598 2599 INIT_LIST_HEAD(&msk->conn_list); 2600 INIT_LIST_HEAD(&msk->join_list); 2601 INIT_LIST_HEAD(&msk->rtx_queue); 2602 INIT_WORK(&msk->work, mptcp_worker); 2603 __skb_queue_head_init(&msk->receive_queue); 2604 msk->out_of_order_queue = RB_ROOT; 2605 msk->first_pending = NULL; 2606 msk->rmem_fwd_alloc = 0; 2607 WRITE_ONCE(msk->rmem_released, 0); 2608 msk->timer_ival = TCP_RTO_MIN; 2609 2610 msk->first = NULL; 2611 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2612 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2613 WRITE_ONCE(msk->allow_infinite_fallback, true); 2614 msk->recovery = false; 2615 2616 mptcp_pm_data_init(msk); 2617 2618 /* re-use the csk retrans timer for MPTCP-level retrans */ 2619 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2620 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0); 2621 2622 return 0; 2623 } 2624 2625 static void mptcp_ca_reset(struct sock *sk) 2626 { 2627 struct inet_connection_sock *icsk = inet_csk(sk); 2628 2629 tcp_assign_congestion_control(sk); 2630 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name); 2631 2632 /* no need to keep a reference to the ops, the name will suffice */ 2633 tcp_cleanup_congestion_control(sk); 2634 icsk->icsk_ca_ops = NULL; 2635 } 2636 2637 static int mptcp_init_sock(struct sock *sk) 2638 { 2639 struct net *net = sock_net(sk); 2640 int ret; 2641 2642 ret = __mptcp_init_sock(sk); 2643 if (ret) 2644 return ret; 2645 2646 if (!mptcp_is_enabled(net)) 2647 return -ENOPROTOOPT; 2648 2649 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2650 return -ENOMEM; 2651 2652 ret = __mptcp_socket_create(mptcp_sk(sk)); 2653 if (ret) 2654 return ret; 2655 2656 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2657 * propagate the correct value 2658 */ 2659 mptcp_ca_reset(sk); 2660 2661 sk_sockets_allocated_inc(sk); 2662 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1]; 2663 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1]; 2664 2665 return 0; 2666 } 2667 2668 static void __mptcp_clear_xmit(struct sock *sk) 2669 { 2670 struct mptcp_sock *msk = mptcp_sk(sk); 2671 struct mptcp_data_frag *dtmp, *dfrag; 2672 2673 WRITE_ONCE(msk->first_pending, NULL); 2674 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2675 dfrag_clear(sk, dfrag); 2676 } 2677 2678 static void mptcp_cancel_work(struct sock *sk) 2679 { 2680 struct mptcp_sock *msk = mptcp_sk(sk); 2681 2682 if (cancel_work_sync(&msk->work)) 2683 __sock_put(sk); 2684 } 2685 2686 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2687 { 2688 lock_sock(ssk); 2689 2690 switch (ssk->sk_state) { 2691 case TCP_LISTEN: 2692 if (!(how & RCV_SHUTDOWN)) 2693 break; 2694 fallthrough; 2695 case TCP_SYN_SENT: 2696 tcp_disconnect(ssk, O_NONBLOCK); 2697 break; 2698 default: 2699 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2700 pr_debug("Fallback"); 2701 ssk->sk_shutdown |= how; 2702 tcp_shutdown(ssk, how); 2703 } else { 2704 pr_debug("Sending DATA_FIN on subflow %p", ssk); 2705 tcp_send_ack(ssk); 2706 mptcp_data_lock(sk); 2707 if (!mptcp_timer_pending(sk)) 2708 mptcp_reset_timer(sk); 2709 mptcp_data_unlock(sk); 2710 } 2711 break; 2712 } 2713 2714 release_sock(ssk); 2715 } 2716 2717 static const unsigned char new_state[16] = { 2718 /* current state: new state: action: */ 2719 [0 /* (Invalid) */] = TCP_CLOSE, 2720 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2721 [TCP_SYN_SENT] = TCP_CLOSE, 2722 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2723 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2724 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2725 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2726 [TCP_CLOSE] = TCP_CLOSE, 2727 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2728 [TCP_LAST_ACK] = TCP_LAST_ACK, 2729 [TCP_LISTEN] = TCP_CLOSE, 2730 [TCP_CLOSING] = TCP_CLOSING, 2731 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2732 }; 2733 2734 static int mptcp_close_state(struct sock *sk) 2735 { 2736 int next = (int)new_state[sk->sk_state]; 2737 int ns = next & TCP_STATE_MASK; 2738 2739 inet_sk_state_store(sk, ns); 2740 2741 return next & TCP_ACTION_FIN; 2742 } 2743 2744 static void __mptcp_check_send_data_fin(struct sock *sk) 2745 { 2746 struct mptcp_subflow_context *subflow; 2747 struct mptcp_sock *msk = mptcp_sk(sk); 2748 2749 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu", 2750 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2751 msk->snd_nxt, msk->write_seq); 2752 2753 /* we still need to enqueue subflows or not really shutting down, 2754 * skip this 2755 */ 2756 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2757 mptcp_send_head(sk)) 2758 return; 2759 2760 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2761 2762 /* fallback socket will not get data_fin/ack, can move to the next 2763 * state now 2764 */ 2765 if (__mptcp_check_fallback(msk)) { 2766 WRITE_ONCE(msk->snd_una, msk->write_seq); 2767 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) { 2768 inet_sk_state_store(sk, TCP_CLOSE); 2769 mptcp_close_wake_up(sk); 2770 } else if (sk->sk_state == TCP_FIN_WAIT1) { 2771 inet_sk_state_store(sk, TCP_FIN_WAIT2); 2772 } 2773 } 2774 2775 mptcp_for_each_subflow(msk, subflow) { 2776 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2777 2778 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2779 } 2780 } 2781 2782 static void __mptcp_wr_shutdown(struct sock *sk) 2783 { 2784 struct mptcp_sock *msk = mptcp_sk(sk); 2785 2786 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d", 2787 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2788 !!mptcp_send_head(sk)); 2789 2790 /* will be ignored by fallback sockets */ 2791 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2792 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2793 2794 __mptcp_check_send_data_fin(sk); 2795 } 2796 2797 static void __mptcp_destroy_sock(struct sock *sk) 2798 { 2799 struct mptcp_subflow_context *subflow, *tmp; 2800 struct mptcp_sock *msk = mptcp_sk(sk); 2801 LIST_HEAD(conn_list); 2802 2803 pr_debug("msk=%p", msk); 2804 2805 might_sleep(); 2806 2807 /* join list will be eventually flushed (with rst) at sock lock release time*/ 2808 list_splice_init(&msk->conn_list, &conn_list); 2809 2810 mptcp_data_lock(sk); 2811 mptcp_stop_timer(sk); 2812 sk_stop_timer(sk, &sk->sk_timer); 2813 mptcp_data_unlock(sk); 2814 msk->pm.status = 0; 2815 2816 /* clears msk->subflow, allowing the following loop to close 2817 * even the initial subflow 2818 */ 2819 mptcp_dispose_initial_subflow(msk); 2820 list_for_each_entry_safe(subflow, tmp, &conn_list, node) { 2821 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2822 __mptcp_close_ssk(sk, ssk, subflow, 0); 2823 } 2824 2825 sk->sk_prot->destroy(sk); 2826 2827 WARN_ON_ONCE(msk->rmem_fwd_alloc); 2828 WARN_ON_ONCE(msk->rmem_released); 2829 sk_stream_kill_queues(sk); 2830 xfrm_sk_free_policy(sk); 2831 2832 sk_refcnt_debug_release(sk); 2833 sock_put(sk); 2834 } 2835 2836 static void mptcp_close(struct sock *sk, long timeout) 2837 { 2838 struct mptcp_subflow_context *subflow; 2839 bool do_cancel_work = false; 2840 2841 lock_sock(sk); 2842 sk->sk_shutdown = SHUTDOWN_MASK; 2843 2844 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 2845 inet_sk_state_store(sk, TCP_CLOSE); 2846 goto cleanup; 2847 } 2848 2849 if (mptcp_close_state(sk)) 2850 __mptcp_wr_shutdown(sk); 2851 2852 sk_stream_wait_close(sk, timeout); 2853 2854 cleanup: 2855 /* orphan all the subflows */ 2856 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32; 2857 mptcp_for_each_subflow(mptcp_sk(sk), subflow) { 2858 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2859 bool slow = lock_sock_fast_nested(ssk); 2860 2861 sock_orphan(ssk); 2862 unlock_sock_fast(ssk, slow); 2863 } 2864 sock_orphan(sk); 2865 2866 sock_hold(sk); 2867 pr_debug("msk=%p state=%d", sk, sk->sk_state); 2868 if (mptcp_sk(sk)->token) 2869 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL); 2870 2871 if (sk->sk_state == TCP_CLOSE) { 2872 __mptcp_destroy_sock(sk); 2873 do_cancel_work = true; 2874 } else { 2875 mptcp_data_lock(sk); 2876 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN); 2877 mptcp_data_unlock(sk); 2878 } 2879 release_sock(sk); 2880 if (do_cancel_work) 2881 mptcp_cancel_work(sk); 2882 2883 sock_put(sk); 2884 } 2885 2886 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 2887 { 2888 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2889 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 2890 struct ipv6_pinfo *msk6 = inet6_sk(msk); 2891 2892 msk->sk_v6_daddr = ssk->sk_v6_daddr; 2893 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 2894 2895 if (msk6 && ssk6) { 2896 msk6->saddr = ssk6->saddr; 2897 msk6->flow_label = ssk6->flow_label; 2898 } 2899 #endif 2900 2901 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 2902 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 2903 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 2904 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 2905 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 2906 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 2907 } 2908 2909 static int mptcp_disconnect(struct sock *sk, int flags) 2910 { 2911 struct mptcp_subflow_context *subflow; 2912 struct mptcp_sock *msk = mptcp_sk(sk); 2913 2914 inet_sk_state_store(sk, TCP_CLOSE); 2915 2916 mptcp_for_each_subflow(msk, subflow) { 2917 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2918 2919 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_FASTCLOSE); 2920 } 2921 2922 mptcp_data_lock(sk); 2923 mptcp_stop_timer(sk); 2924 sk_stop_timer(sk, &sk->sk_timer); 2925 mptcp_data_unlock(sk); 2926 2927 if (mptcp_sk(sk)->token) 2928 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL); 2929 2930 mptcp_destroy_common(msk); 2931 msk->last_snd = NULL; 2932 WRITE_ONCE(msk->flags, 0); 2933 msk->cb_flags = 0; 2934 msk->push_pending = 0; 2935 msk->recovery = false; 2936 msk->can_ack = false; 2937 msk->fully_established = false; 2938 msk->rcv_data_fin = false; 2939 msk->snd_data_fin_enable = false; 2940 msk->rcv_fastclose = false; 2941 msk->use_64bit_ack = false; 2942 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2943 mptcp_pm_data_reset(msk); 2944 mptcp_ca_reset(sk); 2945 2946 sk->sk_shutdown = 0; 2947 sk_error_report(sk); 2948 return 0; 2949 } 2950 2951 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2952 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 2953 { 2954 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo); 2955 2956 return (struct ipv6_pinfo *)(((u8 *)sk) + offset); 2957 } 2958 #endif 2959 2960 struct sock *mptcp_sk_clone(const struct sock *sk, 2961 const struct mptcp_options_received *mp_opt, 2962 struct request_sock *req) 2963 { 2964 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 2965 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 2966 struct mptcp_sock *msk; 2967 u64 ack_seq; 2968 2969 if (!nsk) 2970 return NULL; 2971 2972 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 2973 if (nsk->sk_family == AF_INET6) 2974 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 2975 #endif 2976 2977 __mptcp_init_sock(nsk); 2978 2979 msk = mptcp_sk(nsk); 2980 msk->local_key = subflow_req->local_key; 2981 msk->token = subflow_req->token; 2982 msk->subflow = NULL; 2983 WRITE_ONCE(msk->fully_established, false); 2984 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 2985 WRITE_ONCE(msk->csum_enabled, true); 2986 2987 msk->write_seq = subflow_req->idsn + 1; 2988 msk->snd_nxt = msk->write_seq; 2989 msk->snd_una = msk->write_seq; 2990 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd; 2991 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 2992 2993 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) { 2994 msk->can_ack = true; 2995 msk->remote_key = mp_opt->sndr_key; 2996 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq); 2997 ack_seq++; 2998 WRITE_ONCE(msk->ack_seq, ack_seq); 2999 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq); 3000 } 3001 3002 sock_reset_flag(nsk, SOCK_RCU_FREE); 3003 /* will be fully established after successful MPC subflow creation */ 3004 inet_sk_state_store(nsk, TCP_SYN_RECV); 3005 3006 security_inet_csk_clone(nsk, req); 3007 bh_unlock_sock(nsk); 3008 3009 /* keep a single reference */ 3010 __sock_put(nsk); 3011 return nsk; 3012 } 3013 3014 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3015 { 3016 const struct tcp_sock *tp = tcp_sk(ssk); 3017 3018 msk->rcvq_space.copied = 0; 3019 msk->rcvq_space.rtt_us = 0; 3020 3021 msk->rcvq_space.time = tp->tcp_mstamp; 3022 3023 /* initial rcv_space offering made to peer */ 3024 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3025 TCP_INIT_CWND * tp->advmss); 3026 if (msk->rcvq_space.space == 0) 3027 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3028 3029 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 3030 } 3031 3032 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err, 3033 bool kern) 3034 { 3035 struct mptcp_sock *msk = mptcp_sk(sk); 3036 struct socket *listener; 3037 struct sock *newsk; 3038 3039 listener = __mptcp_nmpc_socket(msk); 3040 if (WARN_ON_ONCE(!listener)) { 3041 *err = -EINVAL; 3042 return NULL; 3043 } 3044 3045 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk)); 3046 newsk = inet_csk_accept(listener->sk, flags, err, kern); 3047 if (!newsk) 3048 return NULL; 3049 3050 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk)); 3051 if (sk_is_mptcp(newsk)) { 3052 struct mptcp_subflow_context *subflow; 3053 struct sock *new_mptcp_sock; 3054 3055 subflow = mptcp_subflow_ctx(newsk); 3056 new_mptcp_sock = subflow->conn; 3057 3058 /* is_mptcp should be false if subflow->conn is missing, see 3059 * subflow_syn_recv_sock() 3060 */ 3061 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3062 tcp_sk(newsk)->is_mptcp = 0; 3063 goto out; 3064 } 3065 3066 /* acquire the 2nd reference for the owning socket */ 3067 sock_hold(new_mptcp_sock); 3068 newsk = new_mptcp_sock; 3069 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3070 } else { 3071 MPTCP_INC_STATS(sock_net(sk), 3072 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK); 3073 } 3074 3075 out: 3076 newsk->sk_kern_sock = kern; 3077 return newsk; 3078 } 3079 3080 void mptcp_destroy_common(struct mptcp_sock *msk) 3081 { 3082 struct sock *sk = (struct sock *)msk; 3083 3084 __mptcp_clear_xmit(sk); 3085 3086 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */ 3087 mptcp_data_lock(sk); 3088 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue); 3089 __skb_queue_purge(&sk->sk_receive_queue); 3090 skb_rbtree_purge(&msk->out_of_order_queue); 3091 mptcp_data_unlock(sk); 3092 3093 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3094 * inet_sock_destruct() will dispose it 3095 */ 3096 sk->sk_forward_alloc += msk->rmem_fwd_alloc; 3097 msk->rmem_fwd_alloc = 0; 3098 mptcp_token_destroy(msk); 3099 mptcp_pm_free_anno_list(msk); 3100 } 3101 3102 static void mptcp_destroy(struct sock *sk) 3103 { 3104 struct mptcp_sock *msk = mptcp_sk(sk); 3105 3106 mptcp_destroy_common(msk); 3107 sk_sockets_allocated_dec(sk); 3108 } 3109 3110 void __mptcp_data_acked(struct sock *sk) 3111 { 3112 if (!sock_owned_by_user(sk)) 3113 __mptcp_clean_una(sk); 3114 else 3115 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3116 3117 if (mptcp_pending_data_fin_ack(sk)) 3118 mptcp_schedule_work(sk); 3119 } 3120 3121 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3122 { 3123 if (!mptcp_send_head(sk)) 3124 return; 3125 3126 if (!sock_owned_by_user(sk)) { 3127 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk)); 3128 3129 if (xmit_ssk == ssk) 3130 __mptcp_subflow_push_pending(sk, ssk); 3131 else if (xmit_ssk) 3132 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND); 3133 } else { 3134 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3135 } 3136 } 3137 3138 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3139 BIT(MPTCP_RETRANSMIT) | \ 3140 BIT(MPTCP_FLUSH_JOIN_LIST)) 3141 3142 /* processes deferred events and flush wmem */ 3143 static void mptcp_release_cb(struct sock *sk) 3144 __must_hold(&sk->sk_lock.slock) 3145 { 3146 struct mptcp_sock *msk = mptcp_sk(sk); 3147 3148 for (;;) { 3149 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) | 3150 msk->push_pending; 3151 if (!flags) 3152 break; 3153 3154 /* the following actions acquire the subflow socket lock 3155 * 3156 * 1) can't be invoked in atomic scope 3157 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3158 * datapath acquires the msk socket spinlock while helding 3159 * the subflow socket lock 3160 */ 3161 msk->push_pending = 0; 3162 msk->cb_flags &= ~flags; 3163 spin_unlock_bh(&sk->sk_lock.slock); 3164 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3165 __mptcp_flush_join_list(sk); 3166 if (flags & BIT(MPTCP_PUSH_PENDING)) 3167 __mptcp_push_pending(sk, 0); 3168 if (flags & BIT(MPTCP_RETRANSMIT)) 3169 __mptcp_retrans(sk); 3170 3171 cond_resched(); 3172 spin_lock_bh(&sk->sk_lock.slock); 3173 } 3174 3175 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3176 __mptcp_clean_una_wakeup(sk); 3177 if (unlikely(&msk->cb_flags)) { 3178 /* be sure to set the current sk state before tacking actions 3179 * depending on sk_state, that is processing MPTCP_ERROR_REPORT 3180 */ 3181 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags)) 3182 __mptcp_set_connected(sk); 3183 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3184 __mptcp_error_report(sk); 3185 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags)) 3186 msk->last_snd = NULL; 3187 } 3188 3189 __mptcp_update_rmem(sk); 3190 } 3191 3192 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3193 * TCP can't schedule delack timer before the subflow is fully established. 3194 * MPTCP uses the delack timer to do 3rd ack retransmissions 3195 */ 3196 static void schedule_3rdack_retransmission(struct sock *ssk) 3197 { 3198 struct inet_connection_sock *icsk = inet_csk(ssk); 3199 struct tcp_sock *tp = tcp_sk(ssk); 3200 unsigned long timeout; 3201 3202 if (mptcp_subflow_ctx(ssk)->fully_established) 3203 return; 3204 3205 /* reschedule with a timeout above RTT, as we must look only for drop */ 3206 if (tp->srtt_us) 3207 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3208 else 3209 timeout = TCP_TIMEOUT_INIT; 3210 timeout += jiffies; 3211 3212 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3213 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3214 icsk->icsk_ack.timeout = timeout; 3215 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3216 } 3217 3218 void mptcp_subflow_process_delegated(struct sock *ssk) 3219 { 3220 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3221 struct sock *sk = subflow->conn; 3222 3223 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) { 3224 mptcp_data_lock(sk); 3225 if (!sock_owned_by_user(sk)) 3226 __mptcp_subflow_push_pending(sk, ssk); 3227 else 3228 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3229 mptcp_data_unlock(sk); 3230 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND); 3231 } 3232 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) { 3233 schedule_3rdack_retransmission(ssk); 3234 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK); 3235 } 3236 } 3237 3238 static int mptcp_hash(struct sock *sk) 3239 { 3240 /* should never be called, 3241 * we hash the TCP subflows not the master socket 3242 */ 3243 WARN_ON_ONCE(1); 3244 return 0; 3245 } 3246 3247 static void mptcp_unhash(struct sock *sk) 3248 { 3249 /* called from sk_common_release(), but nothing to do here */ 3250 } 3251 3252 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3253 { 3254 struct mptcp_sock *msk = mptcp_sk(sk); 3255 struct socket *ssock; 3256 3257 ssock = __mptcp_nmpc_socket(msk); 3258 pr_debug("msk=%p, subflow=%p", msk, ssock); 3259 if (WARN_ON_ONCE(!ssock)) 3260 return -EINVAL; 3261 3262 return inet_csk_get_port(ssock->sk, snum); 3263 } 3264 3265 void mptcp_finish_connect(struct sock *ssk) 3266 { 3267 struct mptcp_subflow_context *subflow; 3268 struct mptcp_sock *msk; 3269 struct sock *sk; 3270 u64 ack_seq; 3271 3272 subflow = mptcp_subflow_ctx(ssk); 3273 sk = subflow->conn; 3274 msk = mptcp_sk(sk); 3275 3276 pr_debug("msk=%p, token=%u", sk, subflow->token); 3277 3278 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq); 3279 ack_seq++; 3280 subflow->map_seq = ack_seq; 3281 subflow->map_subflow_seq = 1; 3282 3283 /* the socket is not connected yet, no msk/subflow ops can access/race 3284 * accessing the field below 3285 */ 3286 WRITE_ONCE(msk->remote_key, subflow->remote_key); 3287 WRITE_ONCE(msk->local_key, subflow->local_key); 3288 WRITE_ONCE(msk->write_seq, subflow->idsn + 1); 3289 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3290 WRITE_ONCE(msk->ack_seq, ack_seq); 3291 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq); 3292 WRITE_ONCE(msk->can_ack, 1); 3293 WRITE_ONCE(msk->snd_una, msk->write_seq); 3294 3295 mptcp_pm_new_connection(msk, ssk, 0); 3296 3297 mptcp_rcv_space_init(msk, ssk); 3298 } 3299 3300 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3301 { 3302 write_lock_bh(&sk->sk_callback_lock); 3303 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3304 sk_set_socket(sk, parent); 3305 sk->sk_uid = SOCK_INODE(parent)->i_uid; 3306 write_unlock_bh(&sk->sk_callback_lock); 3307 } 3308 3309 bool mptcp_finish_join(struct sock *ssk) 3310 { 3311 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3312 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3313 struct sock *parent = (void *)msk; 3314 bool ret = true; 3315 3316 pr_debug("msk=%p, subflow=%p", msk, subflow); 3317 3318 /* mptcp socket already closing? */ 3319 if (!mptcp_is_fully_established(parent)) { 3320 subflow->reset_reason = MPTCP_RST_EMPTCP; 3321 return false; 3322 } 3323 3324 if (!msk->pm.server_side) 3325 goto out; 3326 3327 if (!mptcp_pm_allow_new_subflow(msk)) 3328 goto err_prohibited; 3329 3330 if (WARN_ON_ONCE(!list_empty(&subflow->node))) 3331 goto err_prohibited; 3332 3333 /* active connections are already on conn_list. 3334 * If we can't acquire msk socket lock here, let the release callback 3335 * handle it 3336 */ 3337 mptcp_data_lock(parent); 3338 if (!sock_owned_by_user(parent)) { 3339 ret = __mptcp_finish_join(msk, ssk); 3340 if (ret) { 3341 sock_hold(ssk); 3342 list_add_tail(&subflow->node, &msk->conn_list); 3343 } 3344 } else { 3345 sock_hold(ssk); 3346 list_add_tail(&subflow->node, &msk->join_list); 3347 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3348 } 3349 mptcp_data_unlock(parent); 3350 3351 if (!ret) { 3352 err_prohibited: 3353 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3354 return false; 3355 } 3356 3357 subflow->map_seq = READ_ONCE(msk->ack_seq); 3358 WRITE_ONCE(msk->allow_infinite_fallback, false); 3359 3360 out: 3361 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 3362 return true; 3363 } 3364 3365 static void mptcp_shutdown(struct sock *sk, int how) 3366 { 3367 pr_debug("sk=%p, how=%d", sk, how); 3368 3369 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3370 __mptcp_wr_shutdown(sk); 3371 } 3372 3373 static int mptcp_forward_alloc_get(const struct sock *sk) 3374 { 3375 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc; 3376 } 3377 3378 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3379 { 3380 const struct sock *sk = (void *)msk; 3381 u64 delta; 3382 3383 if (sk->sk_state == TCP_LISTEN) 3384 return -EINVAL; 3385 3386 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3387 return 0; 3388 3389 delta = msk->write_seq - v; 3390 if (__mptcp_check_fallback(msk) && msk->first) { 3391 struct tcp_sock *tp = tcp_sk(msk->first); 3392 3393 /* the first subflow is disconnected after close - see 3394 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3395 * so ignore that status, too. 3396 */ 3397 if (!((1 << msk->first->sk_state) & 3398 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3399 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3400 } 3401 if (delta > INT_MAX) 3402 delta = INT_MAX; 3403 3404 return (int)delta; 3405 } 3406 3407 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg) 3408 { 3409 struct mptcp_sock *msk = mptcp_sk(sk); 3410 bool slow; 3411 int answ; 3412 3413 switch (cmd) { 3414 case SIOCINQ: 3415 if (sk->sk_state == TCP_LISTEN) 3416 return -EINVAL; 3417 3418 lock_sock(sk); 3419 __mptcp_move_skbs(msk); 3420 answ = mptcp_inq_hint(sk); 3421 release_sock(sk); 3422 break; 3423 case SIOCOUTQ: 3424 slow = lock_sock_fast(sk); 3425 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3426 unlock_sock_fast(sk, slow); 3427 break; 3428 case SIOCOUTQNSD: 3429 slow = lock_sock_fast(sk); 3430 answ = mptcp_ioctl_outq(msk, msk->snd_nxt); 3431 unlock_sock_fast(sk, slow); 3432 break; 3433 default: 3434 return -ENOIOCTLCMD; 3435 } 3436 3437 return put_user(answ, (int __user *)arg); 3438 } 3439 3440 static struct proto mptcp_prot = { 3441 .name = "MPTCP", 3442 .owner = THIS_MODULE, 3443 .init = mptcp_init_sock, 3444 .disconnect = mptcp_disconnect, 3445 .close = mptcp_close, 3446 .accept = mptcp_accept, 3447 .setsockopt = mptcp_setsockopt, 3448 .getsockopt = mptcp_getsockopt, 3449 .shutdown = mptcp_shutdown, 3450 .destroy = mptcp_destroy, 3451 .sendmsg = mptcp_sendmsg, 3452 .ioctl = mptcp_ioctl, 3453 .recvmsg = mptcp_recvmsg, 3454 .release_cb = mptcp_release_cb, 3455 .hash = mptcp_hash, 3456 .unhash = mptcp_unhash, 3457 .get_port = mptcp_get_port, 3458 .forward_alloc_get = mptcp_forward_alloc_get, 3459 .sockets_allocated = &mptcp_sockets_allocated, 3460 .memory_allocated = &tcp_memory_allocated, 3461 .memory_pressure = &tcp_memory_pressure, 3462 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3463 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3464 .sysctl_mem = sysctl_tcp_mem, 3465 .obj_size = sizeof(struct mptcp_sock), 3466 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3467 .no_autobind = true, 3468 }; 3469 3470 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3471 { 3472 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3473 struct socket *ssock; 3474 int err; 3475 3476 lock_sock(sock->sk); 3477 ssock = __mptcp_nmpc_socket(msk); 3478 if (!ssock) { 3479 err = -EINVAL; 3480 goto unlock; 3481 } 3482 3483 err = ssock->ops->bind(ssock, uaddr, addr_len); 3484 if (!err) 3485 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3486 3487 unlock: 3488 release_sock(sock->sk); 3489 return err; 3490 } 3491 3492 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk, 3493 struct mptcp_subflow_context *subflow) 3494 { 3495 subflow->request_mptcp = 0; 3496 __mptcp_do_fallback(msk); 3497 } 3498 3499 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr, 3500 int addr_len, int flags) 3501 { 3502 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3503 struct mptcp_subflow_context *subflow; 3504 struct socket *ssock; 3505 int err = -EINVAL; 3506 3507 lock_sock(sock->sk); 3508 if (uaddr) { 3509 if (addr_len < sizeof(uaddr->sa_family)) 3510 goto unlock; 3511 3512 if (uaddr->sa_family == AF_UNSPEC) { 3513 err = mptcp_disconnect(sock->sk, flags); 3514 sock->state = err ? SS_DISCONNECTING : SS_UNCONNECTED; 3515 goto unlock; 3516 } 3517 } 3518 3519 if (sock->state != SS_UNCONNECTED && msk->subflow) { 3520 /* pending connection or invalid state, let existing subflow 3521 * cope with that 3522 */ 3523 ssock = msk->subflow; 3524 goto do_connect; 3525 } 3526 3527 ssock = __mptcp_nmpc_socket(msk); 3528 if (!ssock) 3529 goto unlock; 3530 3531 mptcp_token_destroy(msk); 3532 inet_sk_state_store(sock->sk, TCP_SYN_SENT); 3533 subflow = mptcp_subflow_ctx(ssock->sk); 3534 #ifdef CONFIG_TCP_MD5SIG 3535 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3536 * TCP option space. 3537 */ 3538 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info)) 3539 mptcp_subflow_early_fallback(msk, subflow); 3540 #endif 3541 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) { 3542 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT); 3543 mptcp_subflow_early_fallback(msk, subflow); 3544 } 3545 if (likely(!__mptcp_check_fallback(msk))) 3546 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE); 3547 3548 do_connect: 3549 err = ssock->ops->connect(ssock, uaddr, addr_len, flags); 3550 sock->state = ssock->state; 3551 3552 /* on successful connect, the msk state will be moved to established by 3553 * subflow_finish_connect() 3554 */ 3555 if (!err || err == -EINPROGRESS) 3556 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3557 else 3558 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk)); 3559 3560 unlock: 3561 release_sock(sock->sk); 3562 return err; 3563 } 3564 3565 static int mptcp_listen(struct socket *sock, int backlog) 3566 { 3567 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3568 struct socket *ssock; 3569 int err; 3570 3571 pr_debug("msk=%p", msk); 3572 3573 lock_sock(sock->sk); 3574 ssock = __mptcp_nmpc_socket(msk); 3575 if (!ssock) { 3576 err = -EINVAL; 3577 goto unlock; 3578 } 3579 3580 mptcp_token_destroy(msk); 3581 inet_sk_state_store(sock->sk, TCP_LISTEN); 3582 sock_set_flag(sock->sk, SOCK_RCU_FREE); 3583 3584 err = ssock->ops->listen(ssock, backlog); 3585 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk)); 3586 if (!err) 3587 mptcp_copy_inaddrs(sock->sk, ssock->sk); 3588 3589 unlock: 3590 release_sock(sock->sk); 3591 return err; 3592 } 3593 3594 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3595 int flags, bool kern) 3596 { 3597 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3598 struct socket *ssock; 3599 int err; 3600 3601 pr_debug("msk=%p", msk); 3602 3603 ssock = __mptcp_nmpc_socket(msk); 3604 if (!ssock) 3605 return -EINVAL; 3606 3607 err = ssock->ops->accept(sock, newsock, flags, kern); 3608 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) { 3609 struct mptcp_sock *msk = mptcp_sk(newsock->sk); 3610 struct mptcp_subflow_context *subflow; 3611 struct sock *newsk = newsock->sk; 3612 3613 lock_sock(newsk); 3614 3615 /* PM/worker can now acquire the first subflow socket 3616 * lock without racing with listener queue cleanup, 3617 * we can notify it, if needed. 3618 * 3619 * Even if remote has reset the initial subflow by now 3620 * the refcnt is still at least one. 3621 */ 3622 subflow = mptcp_subflow_ctx(msk->first); 3623 list_add(&subflow->node, &msk->conn_list); 3624 sock_hold(msk->first); 3625 if (mptcp_is_fully_established(newsk)) 3626 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL); 3627 3628 mptcp_copy_inaddrs(newsk, msk->first); 3629 mptcp_rcv_space_init(msk, msk->first); 3630 mptcp_propagate_sndbuf(newsk, msk->first); 3631 3632 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3633 * This is needed so NOSPACE flag can be set from tcp stack. 3634 */ 3635 mptcp_for_each_subflow(msk, subflow) { 3636 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3637 3638 if (!ssk->sk_socket) 3639 mptcp_sock_graft(ssk, newsock); 3640 } 3641 release_sock(newsk); 3642 } 3643 3644 return err; 3645 } 3646 3647 static __poll_t mptcp_check_readable(struct mptcp_sock *msk) 3648 { 3649 /* Concurrent splices from sk_receive_queue into receive_queue will 3650 * always show at least one non-empty queue when checked in this order. 3651 */ 3652 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) && 3653 skb_queue_empty_lockless(&msk->receive_queue)) 3654 return 0; 3655 3656 return EPOLLIN | EPOLLRDNORM; 3657 } 3658 3659 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3660 { 3661 struct sock *sk = (struct sock *)msk; 3662 3663 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN)) 3664 return EPOLLOUT | EPOLLWRNORM; 3665 3666 if (sk_stream_is_writeable(sk)) 3667 return EPOLLOUT | EPOLLWRNORM; 3668 3669 mptcp_set_nospace(sk); 3670 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */ 3671 if (sk_stream_is_writeable(sk)) 3672 return EPOLLOUT | EPOLLWRNORM; 3673 3674 return 0; 3675 } 3676 3677 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3678 struct poll_table_struct *wait) 3679 { 3680 struct sock *sk = sock->sk; 3681 struct mptcp_sock *msk; 3682 __poll_t mask = 0; 3683 int state; 3684 3685 msk = mptcp_sk(sk); 3686 sock_poll_wait(file, sock, wait); 3687 3688 state = inet_sk_state_load(sk); 3689 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags); 3690 if (state == TCP_LISTEN) { 3691 if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk)) 3692 return 0; 3693 3694 return inet_csk_listen_poll(msk->subflow->sk); 3695 } 3696 3697 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 3698 mask |= mptcp_check_readable(msk); 3699 mask |= mptcp_check_writeable(msk); 3700 } 3701 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 3702 mask |= EPOLLHUP; 3703 if (sk->sk_shutdown & RCV_SHUTDOWN) 3704 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 3705 3706 /* This barrier is coupled with smp_wmb() in tcp_reset() */ 3707 smp_rmb(); 3708 if (sk->sk_err) 3709 mask |= EPOLLERR; 3710 3711 return mask; 3712 } 3713 3714 static const struct proto_ops mptcp_stream_ops = { 3715 .family = PF_INET, 3716 .owner = THIS_MODULE, 3717 .release = inet_release, 3718 .bind = mptcp_bind, 3719 .connect = mptcp_stream_connect, 3720 .socketpair = sock_no_socketpair, 3721 .accept = mptcp_stream_accept, 3722 .getname = inet_getname, 3723 .poll = mptcp_poll, 3724 .ioctl = inet_ioctl, 3725 .gettstamp = sock_gettstamp, 3726 .listen = mptcp_listen, 3727 .shutdown = inet_shutdown, 3728 .setsockopt = sock_common_setsockopt, 3729 .getsockopt = sock_common_getsockopt, 3730 .sendmsg = inet_sendmsg, 3731 .recvmsg = inet_recvmsg, 3732 .mmap = sock_no_mmap, 3733 .sendpage = inet_sendpage, 3734 }; 3735 3736 static struct inet_protosw mptcp_protosw = { 3737 .type = SOCK_STREAM, 3738 .protocol = IPPROTO_MPTCP, 3739 .prot = &mptcp_prot, 3740 .ops = &mptcp_stream_ops, 3741 .flags = INET_PROTOSW_ICSK, 3742 }; 3743 3744 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 3745 { 3746 struct mptcp_delegated_action *delegated; 3747 struct mptcp_subflow_context *subflow; 3748 int work_done = 0; 3749 3750 delegated = container_of(napi, struct mptcp_delegated_action, napi); 3751 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 3752 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3753 3754 bh_lock_sock_nested(ssk); 3755 if (!sock_owned_by_user(ssk) && 3756 mptcp_subflow_has_delegated_action(subflow)) 3757 mptcp_subflow_process_delegated(ssk); 3758 /* ... elsewhere tcp_release_cb_override already processed 3759 * the action or will do at next release_sock(). 3760 * In both case must dequeue the subflow here - on the same 3761 * CPU that scheduled it. 3762 */ 3763 bh_unlock_sock(ssk); 3764 sock_put(ssk); 3765 3766 if (++work_done == budget) 3767 return budget; 3768 } 3769 3770 /* always provide a 0 'work_done' argument, so that napi_complete_done 3771 * will not try accessing the NULL napi->dev ptr 3772 */ 3773 napi_complete_done(napi, 0); 3774 return work_done; 3775 } 3776 3777 void __init mptcp_proto_init(void) 3778 { 3779 struct mptcp_delegated_action *delegated; 3780 int cpu; 3781 3782 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 3783 3784 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 3785 panic("Failed to allocate MPTCP pcpu counter\n"); 3786 3787 init_dummy_netdev(&mptcp_napi_dev); 3788 for_each_possible_cpu(cpu) { 3789 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 3790 INIT_LIST_HEAD(&delegated->head); 3791 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll, 3792 NAPI_POLL_WEIGHT); 3793 napi_enable(&delegated->napi); 3794 } 3795 3796 mptcp_subflow_init(); 3797 mptcp_pm_init(); 3798 mptcp_token_init(); 3799 3800 if (proto_register(&mptcp_prot, 1) != 0) 3801 panic("Failed to register MPTCP proto.\n"); 3802 3803 inet_register_protosw(&mptcp_protosw); 3804 3805 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 3806 } 3807 3808 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3809 static const struct proto_ops mptcp_v6_stream_ops = { 3810 .family = PF_INET6, 3811 .owner = THIS_MODULE, 3812 .release = inet6_release, 3813 .bind = mptcp_bind, 3814 .connect = mptcp_stream_connect, 3815 .socketpair = sock_no_socketpair, 3816 .accept = mptcp_stream_accept, 3817 .getname = inet6_getname, 3818 .poll = mptcp_poll, 3819 .ioctl = inet6_ioctl, 3820 .gettstamp = sock_gettstamp, 3821 .listen = mptcp_listen, 3822 .shutdown = inet_shutdown, 3823 .setsockopt = sock_common_setsockopt, 3824 .getsockopt = sock_common_getsockopt, 3825 .sendmsg = inet6_sendmsg, 3826 .recvmsg = inet6_recvmsg, 3827 .mmap = sock_no_mmap, 3828 .sendpage = inet_sendpage, 3829 #ifdef CONFIG_COMPAT 3830 .compat_ioctl = inet6_compat_ioctl, 3831 #endif 3832 }; 3833 3834 static struct proto mptcp_v6_prot; 3835 3836 static void mptcp_v6_destroy(struct sock *sk) 3837 { 3838 mptcp_destroy(sk); 3839 inet6_destroy_sock(sk); 3840 } 3841 3842 static struct inet_protosw mptcp_v6_protosw = { 3843 .type = SOCK_STREAM, 3844 .protocol = IPPROTO_MPTCP, 3845 .prot = &mptcp_v6_prot, 3846 .ops = &mptcp_v6_stream_ops, 3847 .flags = INET_PROTOSW_ICSK, 3848 }; 3849 3850 int __init mptcp_proto_v6_init(void) 3851 { 3852 int err; 3853 3854 mptcp_v6_prot = mptcp_prot; 3855 strcpy(mptcp_v6_prot.name, "MPTCPv6"); 3856 mptcp_v6_prot.slab = NULL; 3857 mptcp_v6_prot.destroy = mptcp_v6_destroy; 3858 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 3859 3860 err = proto_register(&mptcp_v6_prot, 1); 3861 if (err) 3862 return err; 3863 3864 err = inet6_register_protosw(&mptcp_v6_protosw); 3865 if (err) 3866 proto_unregister(&mptcp_v6_prot); 3867 3868 return err; 3869 } 3870 #endif 3871