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_states.h> 19 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 20 #include <net/transp_v6.h> 21 #endif 22 #include <net/mptcp.h> 23 #include <net/hotdata.h> 24 #include <net/xfrm.h> 25 #include <asm/ioctls.h> 26 #include "protocol.h" 27 #include "mib.h" 28 29 #define CREATE_TRACE_POINTS 30 #include <trace/events/mptcp.h> 31 32 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 33 struct mptcp6_sock { 34 struct mptcp_sock msk; 35 struct ipv6_pinfo np; 36 }; 37 #endif 38 39 enum { 40 MPTCP_CMSG_TS = BIT(0), 41 MPTCP_CMSG_INQ = BIT(1), 42 }; 43 44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp; 45 46 static void __mptcp_destroy_sock(struct sock *sk); 47 static void mptcp_check_send_data_fin(struct sock *sk); 48 49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions) = { 50 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 51 }; 52 static struct net_device *mptcp_napi_dev; 53 54 /* Returns end sequence number of the receiver's advertised window */ 55 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 56 { 57 return READ_ONCE(msk->wnd_end); 58 } 59 60 static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk) 61 { 62 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 63 if (sk->sk_prot == &tcpv6_prot) 64 return &inet6_stream_ops; 65 #endif 66 WARN_ON_ONCE(sk->sk_prot != &tcp_prot); 67 return &inet_stream_ops; 68 } 69 70 static int __mptcp_socket_create(struct mptcp_sock *msk) 71 { 72 struct mptcp_subflow_context *subflow; 73 struct sock *sk = (struct sock *)msk; 74 struct socket *ssock; 75 int err; 76 77 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock); 78 if (err) 79 return err; 80 81 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio; 82 WRITE_ONCE(msk->first, ssock->sk); 83 subflow = mptcp_subflow_ctx(ssock->sk); 84 list_add(&subflow->node, &msk->conn_list); 85 sock_hold(ssock->sk); 86 subflow->request_mptcp = 1; 87 subflow->subflow_id = msk->subflow_id++; 88 89 /* This is the first subflow, always with id 0 */ 90 WRITE_ONCE(subflow->local_id, 0); 91 mptcp_sock_graft(msk->first, sk->sk_socket); 92 iput(SOCK_INODE(ssock)); 93 94 return 0; 95 } 96 97 /* If the MPC handshake is not started, returns the first subflow, 98 * eventually allocating it. 99 */ 100 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk) 101 { 102 struct sock *sk = (struct sock *)msk; 103 int ret; 104 105 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 106 return ERR_PTR(-EINVAL); 107 108 if (!msk->first) { 109 ret = __mptcp_socket_create(msk); 110 if (ret) 111 return ERR_PTR(ret); 112 } 113 114 return msk->first; 115 } 116 117 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 118 { 119 sk_drops_add(sk, skb); 120 __kfree_skb(skb); 121 } 122 123 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 124 struct sk_buff *from) 125 { 126 bool fragstolen; 127 int delta; 128 129 if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) || 130 MPTCP_SKB_CB(from)->offset || 131 ((to->len + from->len) > (sk->sk_rcvbuf >> 3)) || 132 !skb_try_coalesce(to, from, &fragstolen, &delta)) 133 return false; 134 135 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n", 136 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 137 to->len, MPTCP_SKB_CB(from)->end_seq); 138 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 139 140 /* note the fwd memory can reach a negative value after accounting 141 * for the delta, but the later skb free will restore a non 142 * negative one 143 */ 144 atomic_add(delta, &sk->sk_rmem_alloc); 145 sk_mem_charge(sk, delta); 146 kfree_skb_partial(from, fragstolen); 147 148 return true; 149 } 150 151 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 152 struct sk_buff *from) 153 { 154 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 155 return false; 156 157 return mptcp_try_coalesce((struct sock *)msk, to, from); 158 } 159 160 /* "inspired" by tcp_data_queue_ofo(), main differences: 161 * - use mptcp seqs 162 * - don't cope with sacks 163 */ 164 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 165 { 166 struct sock *sk = (struct sock *)msk; 167 struct rb_node **p, *parent; 168 u64 seq, end_seq, max_seq; 169 struct sk_buff *skb1; 170 171 seq = MPTCP_SKB_CB(skb)->map_seq; 172 end_seq = MPTCP_SKB_CB(skb)->end_seq; 173 max_seq = atomic64_read(&msk->rcv_wnd_sent); 174 175 pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq, 176 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 177 if (after64(end_seq, max_seq)) { 178 /* out of window */ 179 mptcp_drop(sk, skb); 180 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 181 (unsigned long long)end_seq - (unsigned long)max_seq, 182 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent)); 183 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 184 return; 185 } 186 187 p = &msk->out_of_order_queue.rb_node; 188 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 189 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 190 rb_link_node(&skb->rbnode, NULL, p); 191 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 192 msk->ooo_last_skb = skb; 193 goto end; 194 } 195 196 /* with 2 subflows, adding at end of ooo queue is quite likely 197 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 198 */ 199 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 202 return; 203 } 204 205 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 206 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 207 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 208 parent = &msk->ooo_last_skb->rbnode; 209 p = &parent->rb_right; 210 goto insert; 211 } 212 213 /* Find place to insert this segment. Handle overlaps on the way. */ 214 parent = NULL; 215 while (*p) { 216 parent = *p; 217 skb1 = rb_to_skb(parent); 218 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 219 p = &parent->rb_left; 220 continue; 221 } 222 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 223 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 224 /* All the bits are present. Drop. */ 225 mptcp_drop(sk, skb); 226 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 227 return; 228 } 229 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 230 /* partial overlap: 231 * | skb | 232 * | skb1 | 233 * continue traversing 234 */ 235 } else { 236 /* skb's seq == skb1's seq and skb covers skb1. 237 * Replace skb1 with skb. 238 */ 239 rb_replace_node(&skb1->rbnode, &skb->rbnode, 240 &msk->out_of_order_queue); 241 mptcp_drop(sk, skb1); 242 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 243 goto merge_right; 244 } 245 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 247 return; 248 } 249 p = &parent->rb_right; 250 } 251 252 insert: 253 /* Insert segment into RB tree. */ 254 rb_link_node(&skb->rbnode, parent, p); 255 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 256 257 merge_right: 258 /* Remove other segments covered by skb. */ 259 while ((skb1 = skb_rb_next(skb)) != NULL) { 260 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 261 break; 262 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 263 mptcp_drop(sk, skb1); 264 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 265 } 266 /* If there is no skb after us, we are the last_skb ! */ 267 if (!skb1) 268 msk->ooo_last_skb = skb; 269 270 end: 271 skb_condense(skb); 272 skb_set_owner_r(skb, sk); 273 } 274 275 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 276 struct sk_buff *skb, unsigned int offset, 277 size_t copy_len) 278 { 279 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 280 struct sock *sk = (struct sock *)msk; 281 struct sk_buff *tail; 282 bool has_rxtstamp; 283 284 __skb_unlink(skb, &ssk->sk_receive_queue); 285 286 skb_ext_reset(skb); 287 skb_orphan(skb); 288 289 /* try to fetch required memory from subflow */ 290 if (!sk_rmem_schedule(sk, skb, skb->truesize)) { 291 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 292 goto drop; 293 } 294 295 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 296 297 /* the skb map_seq accounts for the skb offset: 298 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 299 * value 300 */ 301 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 302 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 303 MPTCP_SKB_CB(skb)->offset = offset; 304 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 305 MPTCP_SKB_CB(skb)->cant_coalesce = 0; 306 307 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 308 /* in sequence */ 309 msk->bytes_received += copy_len; 310 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 311 tail = skb_peek_tail(&sk->sk_receive_queue); 312 if (tail && mptcp_try_coalesce(sk, tail, skb)) 313 return true; 314 315 skb_set_owner_r(skb, sk); 316 __skb_queue_tail(&sk->sk_receive_queue, skb); 317 return true; 318 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 319 mptcp_data_queue_ofo(msk, skb); 320 return false; 321 } 322 323 /* old data, keep it simple and drop the whole pkt, sender 324 * will retransmit as needed, if needed. 325 */ 326 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 327 drop: 328 mptcp_drop(sk, skb); 329 return false; 330 } 331 332 static void mptcp_stop_rtx_timer(struct sock *sk) 333 { 334 struct inet_connection_sock *icsk = inet_csk(sk); 335 336 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 337 mptcp_sk(sk)->timer_ival = 0; 338 } 339 340 static void mptcp_close_wake_up(struct sock *sk) 341 { 342 if (sock_flag(sk, SOCK_DEAD)) 343 return; 344 345 sk->sk_state_change(sk); 346 if (sk->sk_shutdown == SHUTDOWN_MASK || 347 sk->sk_state == TCP_CLOSE) 348 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 349 else 350 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 351 } 352 353 /* called under the msk socket lock */ 354 static bool mptcp_pending_data_fin_ack(struct sock *sk) 355 { 356 struct mptcp_sock *msk = mptcp_sk(sk); 357 358 return ((1 << sk->sk_state) & 359 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 360 msk->write_seq == READ_ONCE(msk->snd_una); 361 } 362 363 static void mptcp_check_data_fin_ack(struct sock *sk) 364 { 365 struct mptcp_sock *msk = mptcp_sk(sk); 366 367 /* Look for an acknowledged DATA_FIN */ 368 if (mptcp_pending_data_fin_ack(sk)) { 369 WRITE_ONCE(msk->snd_data_fin_enable, 0); 370 371 switch (sk->sk_state) { 372 case TCP_FIN_WAIT1: 373 mptcp_set_state(sk, TCP_FIN_WAIT2); 374 break; 375 case TCP_CLOSING: 376 case TCP_LAST_ACK: 377 mptcp_set_state(sk, TCP_CLOSE); 378 break; 379 } 380 381 mptcp_close_wake_up(sk); 382 } 383 } 384 385 /* can be called with no lock acquired */ 386 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 387 { 388 struct mptcp_sock *msk = mptcp_sk(sk); 389 390 if (READ_ONCE(msk->rcv_data_fin) && 391 ((1 << inet_sk_state_load(sk)) & 392 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 393 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 394 395 if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) { 396 if (seq) 397 *seq = rcv_data_fin_seq; 398 399 return true; 400 } 401 } 402 403 return false; 404 } 405 406 static void mptcp_set_datafin_timeout(struct sock *sk) 407 { 408 struct inet_connection_sock *icsk = inet_csk(sk); 409 u32 retransmits; 410 411 retransmits = min_t(u32, icsk->icsk_retransmits, 412 ilog2(TCP_RTO_MAX / TCP_RTO_MIN)); 413 414 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits; 415 } 416 417 static void __mptcp_set_timeout(struct sock *sk, long tout) 418 { 419 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 420 } 421 422 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 423 { 424 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 425 426 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 427 icsk_timeout(inet_csk(ssk)) - jiffies : 0; 428 } 429 430 static void mptcp_set_timeout(struct sock *sk) 431 { 432 struct mptcp_subflow_context *subflow; 433 long tout = 0; 434 435 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 436 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 437 __mptcp_set_timeout(sk, tout); 438 } 439 440 static inline bool tcp_can_send_ack(const struct sock *ssk) 441 { 442 return !((1 << inet_sk_state_load(ssk)) & 443 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 444 } 445 446 void __mptcp_subflow_send_ack(struct sock *ssk) 447 { 448 if (tcp_can_send_ack(ssk)) 449 tcp_send_ack(ssk); 450 } 451 452 static void mptcp_subflow_send_ack(struct sock *ssk) 453 { 454 bool slow; 455 456 slow = lock_sock_fast(ssk); 457 __mptcp_subflow_send_ack(ssk); 458 unlock_sock_fast(ssk, slow); 459 } 460 461 static void mptcp_send_ack(struct mptcp_sock *msk) 462 { 463 struct mptcp_subflow_context *subflow; 464 465 mptcp_for_each_subflow(msk, subflow) 466 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow)); 467 } 468 469 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied) 470 { 471 bool slow; 472 473 slow = lock_sock_fast(ssk); 474 if (tcp_can_send_ack(ssk)) 475 tcp_cleanup_rbuf(ssk, copied); 476 unlock_sock_fast(ssk, slow); 477 } 478 479 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 480 { 481 const struct inet_connection_sock *icsk = inet_csk(ssk); 482 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 483 const struct tcp_sock *tp = tcp_sk(ssk); 484 485 return (ack_pending & ICSK_ACK_SCHED) && 486 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 487 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 488 (rx_empty && ack_pending & 489 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 490 } 491 492 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied) 493 { 494 int old_space = READ_ONCE(msk->old_wspace); 495 struct mptcp_subflow_context *subflow; 496 struct sock *sk = (struct sock *)msk; 497 int space = __mptcp_space(sk); 498 bool cleanup, rx_empty; 499 500 cleanup = (space > 0) && (space >= (old_space << 1)) && copied; 501 rx_empty = !sk_rmem_alloc_get(sk) && copied; 502 503 mptcp_for_each_subflow(msk, subflow) { 504 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 505 506 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 507 mptcp_subflow_cleanup_rbuf(ssk, copied); 508 } 509 } 510 511 static bool mptcp_check_data_fin(struct sock *sk) 512 { 513 struct mptcp_sock *msk = mptcp_sk(sk); 514 u64 rcv_data_fin_seq; 515 bool ret = false; 516 517 /* Need to ack a DATA_FIN received from a peer while this side 518 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 519 * msk->rcv_data_fin was set when parsing the incoming options 520 * at the subflow level and the msk lock was not held, so this 521 * is the first opportunity to act on the DATA_FIN and change 522 * the msk state. 523 * 524 * If we are caught up to the sequence number of the incoming 525 * DATA_FIN, send the DATA_ACK now and do state transition. If 526 * not caught up, do nothing and let the recv code send DATA_ACK 527 * when catching up. 528 */ 529 530 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 531 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 532 WRITE_ONCE(msk->rcv_data_fin, 0); 533 534 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 535 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 536 537 switch (sk->sk_state) { 538 case TCP_ESTABLISHED: 539 mptcp_set_state(sk, TCP_CLOSE_WAIT); 540 break; 541 case TCP_FIN_WAIT1: 542 mptcp_set_state(sk, TCP_CLOSING); 543 break; 544 case TCP_FIN_WAIT2: 545 mptcp_set_state(sk, TCP_CLOSE); 546 break; 547 default: 548 /* Other states not expected */ 549 WARN_ON_ONCE(1); 550 break; 551 } 552 553 ret = true; 554 if (!__mptcp_check_fallback(msk)) 555 mptcp_send_ack(msk); 556 mptcp_close_wake_up(sk); 557 } 558 return ret; 559 } 560 561 static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk) 562 { 563 if (READ_ONCE(msk->allow_infinite_fallback)) { 564 MPTCP_INC_STATS(sock_net(ssk), 565 MPTCP_MIB_DSSCORRUPTIONFALLBACK); 566 mptcp_do_fallback(ssk); 567 } else { 568 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET); 569 mptcp_subflow_reset(ssk); 570 } 571 } 572 573 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 574 struct sock *ssk) 575 { 576 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 577 struct sock *sk = (struct sock *)msk; 578 bool more_data_avail; 579 struct tcp_sock *tp; 580 bool ret = false; 581 582 pr_debug("msk=%p ssk=%p\n", msk, ssk); 583 tp = tcp_sk(ssk); 584 do { 585 u32 map_remaining, offset; 586 u32 seq = tp->copied_seq; 587 struct sk_buff *skb; 588 bool fin; 589 590 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 591 break; 592 593 /* try to move as much data as available */ 594 map_remaining = subflow->map_data_len - 595 mptcp_subflow_get_map_offset(subflow); 596 597 skb = skb_peek(&ssk->sk_receive_queue); 598 if (unlikely(!skb)) 599 break; 600 601 if (__mptcp_check_fallback(msk)) { 602 /* Under fallback skbs have no MPTCP extension and TCP could 603 * collapse them between the dummy map creation and the 604 * current dequeue. Be sure to adjust the map size. 605 */ 606 map_remaining = skb->len; 607 subflow->map_data_len = skb->len; 608 } 609 610 offset = seq - TCP_SKB_CB(skb)->seq; 611 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 612 if (fin) 613 seq++; 614 615 if (offset < skb->len) { 616 size_t len = skb->len - offset; 617 618 ret = __mptcp_move_skb(msk, ssk, skb, offset, len) || ret; 619 seq += len; 620 621 if (unlikely(map_remaining < len)) { 622 DEBUG_NET_WARN_ON_ONCE(1); 623 mptcp_dss_corruption(msk, ssk); 624 } 625 } else { 626 if (unlikely(!fin)) { 627 DEBUG_NET_WARN_ON_ONCE(1); 628 mptcp_dss_corruption(msk, ssk); 629 } 630 631 sk_eat_skb(ssk, skb); 632 } 633 634 WRITE_ONCE(tp->copied_seq, seq); 635 more_data_avail = mptcp_subflow_data_available(ssk); 636 637 } while (more_data_avail); 638 639 if (ret) 640 msk->last_data_recv = tcp_jiffies32; 641 return ret; 642 } 643 644 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 645 { 646 struct sock *sk = (struct sock *)msk; 647 struct sk_buff *skb, *tail; 648 bool moved = false; 649 struct rb_node *p; 650 u64 end_seq; 651 652 p = rb_first(&msk->out_of_order_queue); 653 pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 654 while (p) { 655 skb = rb_to_skb(p); 656 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 657 break; 658 659 p = rb_next(p); 660 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 661 662 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 663 msk->ack_seq))) { 664 mptcp_drop(sk, skb); 665 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 666 continue; 667 } 668 669 end_seq = MPTCP_SKB_CB(skb)->end_seq; 670 tail = skb_peek_tail(&sk->sk_receive_queue); 671 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 672 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 673 674 /* skip overlapping data, if any */ 675 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n", 676 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 677 delta); 678 MPTCP_SKB_CB(skb)->offset += delta; 679 MPTCP_SKB_CB(skb)->map_seq += delta; 680 __skb_queue_tail(&sk->sk_receive_queue, skb); 681 } 682 msk->bytes_received += end_seq - msk->ack_seq; 683 WRITE_ONCE(msk->ack_seq, end_seq); 684 moved = true; 685 } 686 return moved; 687 } 688 689 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk) 690 { 691 int err = sock_error(ssk); 692 int ssk_state; 693 694 if (!err) 695 return false; 696 697 /* only propagate errors on fallen-back sockets or 698 * on MPC connect 699 */ 700 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk))) 701 return false; 702 703 /* We need to propagate only transition to CLOSE state. 704 * Orphaned socket will see such state change via 705 * subflow_sched_work_if_closed() and that path will properly 706 * destroy the msk as needed. 707 */ 708 ssk_state = inet_sk_state_load(ssk); 709 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD)) 710 mptcp_set_state(sk, ssk_state); 711 WRITE_ONCE(sk->sk_err, -err); 712 713 /* This barrier is coupled with smp_rmb() in mptcp_poll() */ 714 smp_wmb(); 715 sk_error_report(sk); 716 return true; 717 } 718 719 void __mptcp_error_report(struct sock *sk) 720 { 721 struct mptcp_subflow_context *subflow; 722 struct mptcp_sock *msk = mptcp_sk(sk); 723 724 mptcp_for_each_subflow(msk, subflow) 725 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow))) 726 break; 727 } 728 729 /* In most cases we will be able to lock the mptcp socket. If its already 730 * owned, we need to defer to the work queue to avoid ABBA deadlock. 731 */ 732 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 733 { 734 struct sock *sk = (struct sock *)msk; 735 bool moved; 736 737 moved = __mptcp_move_skbs_from_subflow(msk, ssk); 738 __mptcp_ofo_queue(msk); 739 if (unlikely(ssk->sk_err)) { 740 if (!sock_owned_by_user(sk)) 741 __mptcp_error_report(sk); 742 else 743 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags); 744 } 745 746 /* If the moves have caught up with the DATA_FIN sequence number 747 * it's time to ack the DATA_FIN and change socket state, but 748 * this is not a good place to change state. Let the workqueue 749 * do it. 750 */ 751 if (mptcp_pending_data_fin(sk, NULL)) 752 mptcp_schedule_work(sk); 753 return moved; 754 } 755 756 static void __mptcp_rcvbuf_update(struct sock *sk, struct sock *ssk) 757 { 758 if (unlikely(ssk->sk_rcvbuf > sk->sk_rcvbuf)) 759 WRITE_ONCE(sk->sk_rcvbuf, ssk->sk_rcvbuf); 760 } 761 762 static void __mptcp_data_ready(struct sock *sk, struct sock *ssk) 763 { 764 struct mptcp_sock *msk = mptcp_sk(sk); 765 766 __mptcp_rcvbuf_update(sk, ssk); 767 768 /* Wake-up the reader only for in-sequence data */ 769 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk)) 770 sk->sk_data_ready(sk); 771 } 772 773 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 774 { 775 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 776 777 /* The peer can send data while we are shutting down this 778 * subflow at msk destruction time, but we must avoid enqueuing 779 * more data to the msk receive queue 780 */ 781 if (unlikely(subflow->disposable)) 782 return; 783 784 mptcp_data_lock(sk); 785 if (!sock_owned_by_user(sk)) 786 __mptcp_data_ready(sk, ssk); 787 else 788 __set_bit(MPTCP_DEQUEUE, &mptcp_sk(sk)->cb_flags); 789 mptcp_data_unlock(sk); 790 } 791 792 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk) 793 { 794 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq); 795 WRITE_ONCE(msk->allow_infinite_fallback, false); 796 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 797 } 798 799 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk) 800 { 801 struct sock *sk = (struct sock *)msk; 802 803 if (sk->sk_state != TCP_ESTABLISHED) 804 return false; 805 806 /* attach to msk socket only after we are sure we will deal with it 807 * at close time 808 */ 809 if (sk->sk_socket && !ssk->sk_socket) 810 mptcp_sock_graft(ssk, sk->sk_socket); 811 812 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++; 813 mptcp_sockopt_sync_locked(msk, ssk); 814 mptcp_subflow_joined(msk, ssk); 815 mptcp_stop_tout_timer(sk); 816 __mptcp_propagate_sndbuf(sk, ssk); 817 return true; 818 } 819 820 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list) 821 { 822 struct mptcp_subflow_context *tmp, *subflow; 823 struct mptcp_sock *msk = mptcp_sk(sk); 824 825 list_for_each_entry_safe(subflow, tmp, join_list, node) { 826 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 827 bool slow = lock_sock_fast(ssk); 828 829 list_move_tail(&subflow->node, &msk->conn_list); 830 if (!__mptcp_finish_join(msk, ssk)) 831 mptcp_subflow_reset(ssk); 832 unlock_sock_fast(ssk, slow); 833 } 834 } 835 836 static bool mptcp_rtx_timer_pending(struct sock *sk) 837 { 838 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 839 } 840 841 static void mptcp_reset_rtx_timer(struct sock *sk) 842 { 843 struct inet_connection_sock *icsk = inet_csk(sk); 844 unsigned long tout; 845 846 /* prevent rescheduling on close */ 847 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 848 return; 849 850 tout = mptcp_sk(sk)->timer_ival; 851 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 852 } 853 854 bool mptcp_schedule_work(struct sock *sk) 855 { 856 if (inet_sk_state_load(sk) != TCP_CLOSE && 857 schedule_work(&mptcp_sk(sk)->work)) { 858 /* each subflow already holds a reference to the sk, and the 859 * workqueue is invoked by a subflow, so sk can't go away here. 860 */ 861 sock_hold(sk); 862 return true; 863 } 864 return false; 865 } 866 867 static bool mptcp_skb_can_collapse_to(u64 write_seq, 868 const struct sk_buff *skb, 869 const struct mptcp_ext *mpext) 870 { 871 if (!tcp_skb_can_collapse_to(skb)) 872 return false; 873 874 /* can collapse only if MPTCP level sequence is in order and this 875 * mapping has not been xmitted yet 876 */ 877 return mpext && mpext->data_seq + mpext->data_len == write_seq && 878 !mpext->frozen; 879 } 880 881 /* we can append data to the given data frag if: 882 * - there is space available in the backing page_frag 883 * - the data frag tail matches the current page_frag free offset 884 * - the data frag end sequence number matches the current write seq 885 */ 886 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 887 const struct page_frag *pfrag, 888 const struct mptcp_data_frag *df) 889 { 890 return df && pfrag->page == df->page && 891 pfrag->size - pfrag->offset > 0 && 892 pfrag->offset == (df->offset + df->data_len) && 893 df->data_seq + df->data_len == msk->write_seq; 894 } 895 896 static void dfrag_uncharge(struct sock *sk, int len) 897 { 898 sk_mem_uncharge(sk, len); 899 sk_wmem_queued_add(sk, -len); 900 } 901 902 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 903 { 904 int len = dfrag->data_len + dfrag->overhead; 905 906 list_del(&dfrag->list); 907 dfrag_uncharge(sk, len); 908 put_page(dfrag->page); 909 } 910 911 /* called under both the msk socket lock and the data lock */ 912 static void __mptcp_clean_una(struct sock *sk) 913 { 914 struct mptcp_sock *msk = mptcp_sk(sk); 915 struct mptcp_data_frag *dtmp, *dfrag; 916 u64 snd_una; 917 918 snd_una = msk->snd_una; 919 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 920 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 921 break; 922 923 if (unlikely(dfrag == msk->first_pending)) { 924 /* in recovery mode can see ack after the current snd head */ 925 if (WARN_ON_ONCE(!msk->recovery)) 926 break; 927 928 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 929 } 930 931 dfrag_clear(sk, dfrag); 932 } 933 934 dfrag = mptcp_rtx_head(sk); 935 if (dfrag && after64(snd_una, dfrag->data_seq)) { 936 u64 delta = snd_una - dfrag->data_seq; 937 938 /* prevent wrap around in recovery mode */ 939 if (unlikely(delta > dfrag->already_sent)) { 940 if (WARN_ON_ONCE(!msk->recovery)) 941 goto out; 942 if (WARN_ON_ONCE(delta > dfrag->data_len)) 943 goto out; 944 dfrag->already_sent += delta - dfrag->already_sent; 945 } 946 947 dfrag->data_seq += delta; 948 dfrag->offset += delta; 949 dfrag->data_len -= delta; 950 dfrag->already_sent -= delta; 951 952 dfrag_uncharge(sk, delta); 953 } 954 955 /* all retransmitted data acked, recovery completed */ 956 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 957 msk->recovery = false; 958 959 out: 960 if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) { 961 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 962 mptcp_stop_rtx_timer(sk); 963 } else { 964 mptcp_reset_rtx_timer(sk); 965 } 966 967 if (mptcp_pending_data_fin_ack(sk)) 968 mptcp_schedule_work(sk); 969 } 970 971 static void __mptcp_clean_una_wakeup(struct sock *sk) 972 { 973 lockdep_assert_held_once(&sk->sk_lock.slock); 974 975 __mptcp_clean_una(sk); 976 mptcp_write_space(sk); 977 } 978 979 static void mptcp_clean_una_wakeup(struct sock *sk) 980 { 981 mptcp_data_lock(sk); 982 __mptcp_clean_una_wakeup(sk); 983 mptcp_data_unlock(sk); 984 } 985 986 static void mptcp_enter_memory_pressure(struct sock *sk) 987 { 988 struct mptcp_subflow_context *subflow; 989 struct mptcp_sock *msk = mptcp_sk(sk); 990 bool first = true; 991 992 mptcp_for_each_subflow(msk, subflow) { 993 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 994 995 if (first) 996 tcp_enter_memory_pressure(ssk); 997 sk_stream_moderate_sndbuf(ssk); 998 999 first = false; 1000 } 1001 __mptcp_sync_sndbuf(sk); 1002 } 1003 1004 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1005 * data 1006 */ 1007 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1008 { 1009 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1010 pfrag, sk->sk_allocation))) 1011 return true; 1012 1013 mptcp_enter_memory_pressure(sk); 1014 return false; 1015 } 1016 1017 static struct mptcp_data_frag * 1018 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1019 int orig_offset) 1020 { 1021 int offset = ALIGN(orig_offset, sizeof(long)); 1022 struct mptcp_data_frag *dfrag; 1023 1024 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1025 dfrag->data_len = 0; 1026 dfrag->data_seq = msk->write_seq; 1027 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1028 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1029 dfrag->already_sent = 0; 1030 dfrag->page = pfrag->page; 1031 1032 return dfrag; 1033 } 1034 1035 struct mptcp_sendmsg_info { 1036 int mss_now; 1037 int size_goal; 1038 u16 limit; 1039 u16 sent; 1040 unsigned int flags; 1041 bool data_lock_held; 1042 }; 1043 1044 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk, 1045 u64 data_seq, int avail_size) 1046 { 1047 u64 window_end = mptcp_wnd_end(msk); 1048 u64 mptcp_snd_wnd; 1049 1050 if (__mptcp_check_fallback(msk)) 1051 return avail_size; 1052 1053 mptcp_snd_wnd = window_end - data_seq; 1054 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size); 1055 1056 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) { 1057 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd); 1058 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED); 1059 } 1060 1061 return avail_size; 1062 } 1063 1064 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1065 { 1066 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1067 1068 if (!mpext) 1069 return false; 1070 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1071 return true; 1072 } 1073 1074 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1075 { 1076 struct sk_buff *skb; 1077 1078 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1079 if (likely(skb)) { 1080 if (likely(__mptcp_add_ext(skb, gfp))) { 1081 skb_reserve(skb, MAX_TCP_HEADER); 1082 skb->ip_summed = CHECKSUM_PARTIAL; 1083 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1084 return skb; 1085 } 1086 __kfree_skb(skb); 1087 } else { 1088 mptcp_enter_memory_pressure(sk); 1089 } 1090 return NULL; 1091 } 1092 1093 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1094 { 1095 struct sk_buff *skb; 1096 1097 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1098 if (!skb) 1099 return NULL; 1100 1101 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1102 tcp_skb_entail(ssk, skb); 1103 return skb; 1104 } 1105 tcp_skb_tsorted_anchor_cleanup(skb); 1106 kfree_skb(skb); 1107 return NULL; 1108 } 1109 1110 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1111 { 1112 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1113 1114 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1115 } 1116 1117 /* note: this always recompute the csum on the whole skb, even 1118 * if we just appended a single frag. More status info needed 1119 */ 1120 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1121 { 1122 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1123 __wsum csum = ~csum_unfold(mpext->csum); 1124 int offset = skb->len - added; 1125 1126 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1127 } 1128 1129 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1130 struct sock *ssk, 1131 struct mptcp_ext *mpext) 1132 { 1133 if (!mpext) 1134 return; 1135 1136 mpext->infinite_map = 1; 1137 mpext->data_len = 0; 1138 1139 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX); 1140 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1141 pr_fallback(msk); 1142 mptcp_do_fallback(ssk); 1143 } 1144 1145 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1)) 1146 1147 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1148 struct mptcp_data_frag *dfrag, 1149 struct mptcp_sendmsg_info *info) 1150 { 1151 u64 data_seq = dfrag->data_seq + info->sent; 1152 int offset = dfrag->offset + info->sent; 1153 struct mptcp_sock *msk = mptcp_sk(sk); 1154 bool zero_window_probe = false; 1155 struct mptcp_ext *mpext = NULL; 1156 bool can_coalesce = false; 1157 bool reuse_skb = true; 1158 struct sk_buff *skb; 1159 size_t copy; 1160 int i; 1161 1162 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n", 1163 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1164 1165 if (WARN_ON_ONCE(info->sent > info->limit || 1166 info->limit > dfrag->data_len)) 1167 return 0; 1168 1169 if (unlikely(!__tcp_can_send(ssk))) 1170 return -EAGAIN; 1171 1172 /* compute send limit */ 1173 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE)) 1174 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE; 1175 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1176 copy = info->size_goal; 1177 1178 skb = tcp_write_queue_tail(ssk); 1179 if (skb && copy > skb->len) { 1180 /* Limit the write to the size available in the 1181 * current skb, if any, so that we create at most a new skb. 1182 * Explicitly tells TCP internals to avoid collapsing on later 1183 * queue management operation, to avoid breaking the ext <-> 1184 * SSN association set here 1185 */ 1186 mpext = mptcp_get_ext(skb); 1187 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1188 TCP_SKB_CB(skb)->eor = 1; 1189 tcp_mark_push(tcp_sk(ssk), skb); 1190 goto alloc_skb; 1191 } 1192 1193 i = skb_shinfo(skb)->nr_frags; 1194 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1195 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) { 1196 tcp_mark_push(tcp_sk(ssk), skb); 1197 goto alloc_skb; 1198 } 1199 1200 copy -= skb->len; 1201 } else { 1202 alloc_skb: 1203 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1204 if (!skb) 1205 return -ENOMEM; 1206 1207 i = skb_shinfo(skb)->nr_frags; 1208 reuse_skb = false; 1209 mpext = mptcp_get_ext(skb); 1210 } 1211 1212 /* Zero window and all data acked? Probe. */ 1213 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy); 1214 if (copy == 0) { 1215 u64 snd_una = READ_ONCE(msk->snd_una); 1216 1217 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) { 1218 tcp_remove_empty_skb(ssk); 1219 return 0; 1220 } 1221 1222 zero_window_probe = true; 1223 data_seq = snd_una - 1; 1224 copy = 1; 1225 } 1226 1227 copy = min_t(size_t, copy, info->limit - info->sent); 1228 if (!sk_wmem_schedule(ssk, copy)) { 1229 tcp_remove_empty_skb(ssk); 1230 return -ENOMEM; 1231 } 1232 1233 if (can_coalesce) { 1234 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1235 } else { 1236 get_page(dfrag->page); 1237 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1238 } 1239 1240 skb->len += copy; 1241 skb->data_len += copy; 1242 skb->truesize += copy; 1243 sk_wmem_queued_add(ssk, copy); 1244 sk_mem_charge(ssk, copy); 1245 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1246 TCP_SKB_CB(skb)->end_seq += copy; 1247 tcp_skb_pcount_set(skb, 0); 1248 1249 /* on skb reuse we just need to update the DSS len */ 1250 if (reuse_skb) { 1251 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1252 mpext->data_len += copy; 1253 goto out; 1254 } 1255 1256 memset(mpext, 0, sizeof(*mpext)); 1257 mpext->data_seq = data_seq; 1258 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1259 mpext->data_len = copy; 1260 mpext->use_map = 1; 1261 mpext->dsn64 = 1; 1262 1263 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n", 1264 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1265 mpext->dsn64); 1266 1267 if (zero_window_probe) { 1268 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1269 mpext->frozen = 1; 1270 if (READ_ONCE(msk->csum_enabled)) 1271 mptcp_update_data_checksum(skb, copy); 1272 tcp_push_pending_frames(ssk); 1273 return 0; 1274 } 1275 out: 1276 if (READ_ONCE(msk->csum_enabled)) 1277 mptcp_update_data_checksum(skb, copy); 1278 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1279 mptcp_update_infinite_map(msk, ssk, mpext); 1280 trace_mptcp_sendmsg_frag(mpext); 1281 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1282 return copy; 1283 } 1284 1285 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1286 sizeof(struct tcphdr) - \ 1287 MAX_TCP_OPTION_SPACE - \ 1288 sizeof(struct ipv6hdr) - \ 1289 sizeof(struct frag_hdr)) 1290 1291 struct subflow_send_info { 1292 struct sock *ssk; 1293 u64 linger_time; 1294 }; 1295 1296 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1297 { 1298 if (!subflow->stale) 1299 return; 1300 1301 subflow->stale = 0; 1302 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1303 } 1304 1305 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1306 { 1307 if (unlikely(subflow->stale)) { 1308 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1309 1310 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1311 return false; 1312 1313 mptcp_subflow_set_active(subflow); 1314 } 1315 return __mptcp_subflow_active(subflow); 1316 } 1317 1318 #define SSK_MODE_ACTIVE 0 1319 #define SSK_MODE_BACKUP 1 1320 #define SSK_MODE_MAX 2 1321 1322 /* implement the mptcp packet scheduler; 1323 * returns the subflow that will transmit the next DSS 1324 * additionally updates the rtx timeout 1325 */ 1326 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1327 { 1328 struct subflow_send_info send_info[SSK_MODE_MAX]; 1329 struct mptcp_subflow_context *subflow; 1330 struct sock *sk = (struct sock *)msk; 1331 u32 pace, burst, wmem; 1332 int i, nr_active = 0; 1333 struct sock *ssk; 1334 u64 linger_time; 1335 long tout = 0; 1336 1337 /* pick the subflow with the lower wmem/wspace ratio */ 1338 for (i = 0; i < SSK_MODE_MAX; ++i) { 1339 send_info[i].ssk = NULL; 1340 send_info[i].linger_time = -1; 1341 } 1342 1343 mptcp_for_each_subflow(msk, subflow) { 1344 bool backup = subflow->backup || subflow->request_bkup; 1345 1346 trace_mptcp_subflow_get_send(subflow); 1347 ssk = mptcp_subflow_tcp_sock(subflow); 1348 if (!mptcp_subflow_active(subflow)) 1349 continue; 1350 1351 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1352 nr_active += !backup; 1353 pace = subflow->avg_pacing_rate; 1354 if (unlikely(!pace)) { 1355 /* init pacing rate from socket */ 1356 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1357 pace = subflow->avg_pacing_rate; 1358 if (!pace) 1359 continue; 1360 } 1361 1362 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1363 if (linger_time < send_info[backup].linger_time) { 1364 send_info[backup].ssk = ssk; 1365 send_info[backup].linger_time = linger_time; 1366 } 1367 } 1368 __mptcp_set_timeout(sk, tout); 1369 1370 /* pick the best backup if no other subflow is active */ 1371 if (!nr_active) 1372 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1373 1374 /* According to the blest algorithm, to avoid HoL blocking for the 1375 * faster flow, we need to: 1376 * - estimate the faster flow linger time 1377 * - use the above to estimate the amount of byte transferred 1378 * by the faster flow 1379 * - check that the amount of queued data is greter than the above, 1380 * otherwise do not use the picked, slower, subflow 1381 * We select the subflow with the shorter estimated time to flush 1382 * the queued mem, which basically ensure the above. We just need 1383 * to check that subflow has a non empty cwin. 1384 */ 1385 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1386 if (!ssk || !sk_stream_memory_free(ssk)) 1387 return NULL; 1388 1389 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt); 1390 wmem = READ_ONCE(ssk->sk_wmem_queued); 1391 if (!burst) 1392 return ssk; 1393 1394 subflow = mptcp_subflow_ctx(ssk); 1395 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1396 READ_ONCE(ssk->sk_pacing_rate) * burst, 1397 burst + wmem); 1398 msk->snd_burst = burst; 1399 return ssk; 1400 } 1401 1402 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1403 { 1404 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1405 release_sock(ssk); 1406 } 1407 1408 static void mptcp_update_post_push(struct mptcp_sock *msk, 1409 struct mptcp_data_frag *dfrag, 1410 u32 sent) 1411 { 1412 u64 snd_nxt_new = dfrag->data_seq; 1413 1414 dfrag->already_sent += sent; 1415 1416 msk->snd_burst -= sent; 1417 1418 snd_nxt_new += dfrag->already_sent; 1419 1420 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1421 * is recovering after a failover. In that event, this re-sends 1422 * old segments. 1423 * 1424 * Thus compute snd_nxt_new candidate based on 1425 * the dfrag->data_seq that was sent and the data 1426 * that has been handed to the subflow for transmission 1427 * and skip update in case it was old dfrag. 1428 */ 1429 if (likely(after64(snd_nxt_new, msk->snd_nxt))) { 1430 msk->bytes_sent += snd_nxt_new - msk->snd_nxt; 1431 WRITE_ONCE(msk->snd_nxt, snd_nxt_new); 1432 } 1433 } 1434 1435 void mptcp_check_and_set_pending(struct sock *sk) 1436 { 1437 if (mptcp_send_head(sk)) { 1438 mptcp_data_lock(sk); 1439 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING); 1440 mptcp_data_unlock(sk); 1441 } 1442 } 1443 1444 static int __subflow_push_pending(struct sock *sk, struct sock *ssk, 1445 struct mptcp_sendmsg_info *info) 1446 { 1447 struct mptcp_sock *msk = mptcp_sk(sk); 1448 struct mptcp_data_frag *dfrag; 1449 int len, copied = 0, err = 0; 1450 1451 while ((dfrag = mptcp_send_head(sk))) { 1452 info->sent = dfrag->already_sent; 1453 info->limit = dfrag->data_len; 1454 len = dfrag->data_len - dfrag->already_sent; 1455 while (len > 0) { 1456 int ret = 0; 1457 1458 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info); 1459 if (ret <= 0) { 1460 err = copied ? : ret; 1461 goto out; 1462 } 1463 1464 info->sent += ret; 1465 copied += ret; 1466 len -= ret; 1467 1468 mptcp_update_post_push(msk, dfrag, ret); 1469 } 1470 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1471 1472 if (msk->snd_burst <= 0 || 1473 !sk_stream_memory_free(ssk) || 1474 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) { 1475 err = copied; 1476 goto out; 1477 } 1478 mptcp_set_timeout(sk); 1479 } 1480 err = copied; 1481 1482 out: 1483 if (err > 0) 1484 msk->last_data_sent = tcp_jiffies32; 1485 return err; 1486 } 1487 1488 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1489 { 1490 struct sock *prev_ssk = NULL, *ssk = NULL; 1491 struct mptcp_sock *msk = mptcp_sk(sk); 1492 struct mptcp_sendmsg_info info = { 1493 .flags = flags, 1494 }; 1495 bool do_check_data_fin = false; 1496 int push_count = 1; 1497 1498 while (mptcp_send_head(sk) && (push_count > 0)) { 1499 struct mptcp_subflow_context *subflow; 1500 int ret = 0; 1501 1502 if (mptcp_sched_get_send(msk)) 1503 break; 1504 1505 push_count = 0; 1506 1507 mptcp_for_each_subflow(msk, subflow) { 1508 if (READ_ONCE(subflow->scheduled)) { 1509 mptcp_subflow_set_scheduled(subflow, false); 1510 1511 prev_ssk = ssk; 1512 ssk = mptcp_subflow_tcp_sock(subflow); 1513 if (ssk != prev_ssk) { 1514 /* First check. If the ssk has changed since 1515 * the last round, release prev_ssk 1516 */ 1517 if (prev_ssk) 1518 mptcp_push_release(prev_ssk, &info); 1519 1520 /* Need to lock the new subflow only if different 1521 * from the previous one, otherwise we are still 1522 * helding the relevant lock 1523 */ 1524 lock_sock(ssk); 1525 } 1526 1527 push_count++; 1528 1529 ret = __subflow_push_pending(sk, ssk, &info); 1530 if (ret <= 0) { 1531 if (ret != -EAGAIN || 1532 (1 << ssk->sk_state) & 1533 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE)) 1534 push_count--; 1535 continue; 1536 } 1537 do_check_data_fin = true; 1538 } 1539 } 1540 } 1541 1542 /* at this point we held the socket lock for the last subflow we used */ 1543 if (ssk) 1544 mptcp_push_release(ssk, &info); 1545 1546 /* ensure the rtx timer is running */ 1547 if (!mptcp_rtx_timer_pending(sk)) 1548 mptcp_reset_rtx_timer(sk); 1549 if (do_check_data_fin) 1550 mptcp_check_send_data_fin(sk); 1551 } 1552 1553 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first) 1554 { 1555 struct mptcp_sock *msk = mptcp_sk(sk); 1556 struct mptcp_sendmsg_info info = { 1557 .data_lock_held = true, 1558 }; 1559 bool keep_pushing = true; 1560 struct sock *xmit_ssk; 1561 int copied = 0; 1562 1563 info.flags = 0; 1564 while (mptcp_send_head(sk) && keep_pushing) { 1565 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1566 int ret = 0; 1567 1568 /* check for a different subflow usage only after 1569 * spooling the first chunk of data 1570 */ 1571 if (first) { 1572 mptcp_subflow_set_scheduled(subflow, false); 1573 ret = __subflow_push_pending(sk, ssk, &info); 1574 first = false; 1575 if (ret <= 0) 1576 break; 1577 copied += ret; 1578 continue; 1579 } 1580 1581 if (mptcp_sched_get_send(msk)) 1582 goto out; 1583 1584 if (READ_ONCE(subflow->scheduled)) { 1585 mptcp_subflow_set_scheduled(subflow, false); 1586 ret = __subflow_push_pending(sk, ssk, &info); 1587 if (ret <= 0) 1588 keep_pushing = false; 1589 copied += ret; 1590 } 1591 1592 mptcp_for_each_subflow(msk, subflow) { 1593 if (READ_ONCE(subflow->scheduled)) { 1594 xmit_ssk = mptcp_subflow_tcp_sock(subflow); 1595 if (xmit_ssk != ssk) { 1596 mptcp_subflow_delegate(subflow, 1597 MPTCP_DELEGATE_SEND); 1598 keep_pushing = false; 1599 } 1600 } 1601 } 1602 } 1603 1604 out: 1605 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1606 * not going to flush it via release_sock() 1607 */ 1608 if (copied) { 1609 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1610 info.size_goal); 1611 if (!mptcp_rtx_timer_pending(sk)) 1612 mptcp_reset_rtx_timer(sk); 1613 1614 if (msk->snd_data_fin_enable && 1615 msk->snd_nxt + 1 == msk->write_seq) 1616 mptcp_schedule_work(sk); 1617 } 1618 } 1619 1620 static int mptcp_disconnect(struct sock *sk, int flags); 1621 1622 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1623 size_t len, int *copied_syn) 1624 { 1625 unsigned int saved_flags = msg->msg_flags; 1626 struct mptcp_sock *msk = mptcp_sk(sk); 1627 struct sock *ssk; 1628 int ret; 1629 1630 /* on flags based fastopen the mptcp is supposed to create the 1631 * first subflow right now. Otherwise we are in the defer_connect 1632 * path, and the first subflow must be already present. 1633 * Since the defer_connect flag is cleared after the first succsful 1634 * fastopen attempt, no need to check for additional subflow status. 1635 */ 1636 if (msg->msg_flags & MSG_FASTOPEN) { 1637 ssk = __mptcp_nmpc_sk(msk); 1638 if (IS_ERR(ssk)) 1639 return PTR_ERR(ssk); 1640 } 1641 if (!msk->first) 1642 return -EINVAL; 1643 1644 ssk = msk->first; 1645 1646 lock_sock(ssk); 1647 msg->msg_flags |= MSG_DONTWAIT; 1648 msk->fastopening = 1; 1649 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL); 1650 msk->fastopening = 0; 1651 msg->msg_flags = saved_flags; 1652 release_sock(ssk); 1653 1654 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */ 1655 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) { 1656 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1657 msg->msg_namelen, msg->msg_flags, 1); 1658 1659 /* Keep the same behaviour of plain TCP: zero the copied bytes in 1660 * case of any error, except timeout or signal 1661 */ 1662 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR) 1663 *copied_syn = 0; 1664 } else if (ret && ret != -EINPROGRESS) { 1665 /* The disconnect() op called by tcp_sendmsg_fastopen()/ 1666 * __inet_stream_connect() can fail, due to looking check, 1667 * see mptcp_disconnect(). 1668 * Attempt it again outside the problematic scope. 1669 */ 1670 if (!mptcp_disconnect(sk, 0)) { 1671 sk->sk_disconnects++; 1672 sk->sk_socket->state = SS_UNCONNECTED; 1673 } 1674 } 1675 inet_clear_bit(DEFER_CONNECT, sk); 1676 1677 return ret; 1678 } 1679 1680 static int do_copy_data_nocache(struct sock *sk, int copy, 1681 struct iov_iter *from, char *to) 1682 { 1683 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) { 1684 if (!copy_from_iter_full_nocache(to, copy, from)) 1685 return -EFAULT; 1686 } else if (!copy_from_iter_full(to, copy, from)) { 1687 return -EFAULT; 1688 } 1689 return 0; 1690 } 1691 1692 /* open-code sk_stream_memory_free() plus sent limit computation to 1693 * avoid indirect calls in fast-path. 1694 * Called under the msk socket lock, so we can avoid a bunch of ONCE 1695 * annotations. 1696 */ 1697 static u32 mptcp_send_limit(const struct sock *sk) 1698 { 1699 const struct mptcp_sock *msk = mptcp_sk(sk); 1700 u32 limit, not_sent; 1701 1702 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf)) 1703 return 0; 1704 1705 limit = mptcp_notsent_lowat(sk); 1706 if (limit == UINT_MAX) 1707 return UINT_MAX; 1708 1709 not_sent = msk->write_seq - msk->snd_nxt; 1710 if (not_sent >= limit) 1711 return 0; 1712 1713 return limit - not_sent; 1714 } 1715 1716 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1717 { 1718 struct mptcp_sock *msk = mptcp_sk(sk); 1719 struct page_frag *pfrag; 1720 size_t copied = 0; 1721 int ret = 0; 1722 long timeo; 1723 1724 /* silently ignore everything else */ 1725 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN; 1726 1727 lock_sock(sk); 1728 1729 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) || 1730 msg->msg_flags & MSG_FASTOPEN)) { 1731 int copied_syn = 0; 1732 1733 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn); 1734 copied += copied_syn; 1735 if (ret == -EINPROGRESS && copied_syn > 0) 1736 goto out; 1737 else if (ret) 1738 goto do_error; 1739 } 1740 1741 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1742 1743 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1744 ret = sk_stream_wait_connect(sk, &timeo); 1745 if (ret) 1746 goto do_error; 1747 } 1748 1749 ret = -EPIPE; 1750 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))) 1751 goto do_error; 1752 1753 pfrag = sk_page_frag(sk); 1754 1755 while (msg_data_left(msg)) { 1756 int total_ts, frag_truesize = 0; 1757 struct mptcp_data_frag *dfrag; 1758 bool dfrag_collapsed; 1759 size_t psize, offset; 1760 u32 copy_limit; 1761 1762 /* ensure fitting the notsent_lowat() constraint */ 1763 copy_limit = mptcp_send_limit(sk); 1764 if (!copy_limit) 1765 goto wait_for_memory; 1766 1767 /* reuse tail pfrag, if possible, or carve a new one from the 1768 * page allocator 1769 */ 1770 dfrag = mptcp_pending_tail(sk); 1771 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1772 if (!dfrag_collapsed) { 1773 if (!mptcp_page_frag_refill(sk, pfrag)) 1774 goto wait_for_memory; 1775 1776 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1777 frag_truesize = dfrag->overhead; 1778 } 1779 1780 /* we do not bound vs wspace, to allow a single packet. 1781 * memory accounting will prevent execessive memory usage 1782 * anyway 1783 */ 1784 offset = dfrag->offset + dfrag->data_len; 1785 psize = pfrag->size - offset; 1786 psize = min_t(size_t, psize, msg_data_left(msg)); 1787 psize = min_t(size_t, psize, copy_limit); 1788 total_ts = psize + frag_truesize; 1789 1790 if (!sk_wmem_schedule(sk, total_ts)) 1791 goto wait_for_memory; 1792 1793 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter, 1794 page_address(dfrag->page) + offset); 1795 if (ret) 1796 goto do_error; 1797 1798 /* data successfully copied into the write queue */ 1799 sk_forward_alloc_add(sk, -total_ts); 1800 copied += psize; 1801 dfrag->data_len += psize; 1802 frag_truesize += psize; 1803 pfrag->offset += frag_truesize; 1804 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1805 1806 /* charge data on mptcp pending queue to the msk socket 1807 * Note: we charge such data both to sk and ssk 1808 */ 1809 sk_wmem_queued_add(sk, frag_truesize); 1810 if (!dfrag_collapsed) { 1811 get_page(dfrag->page); 1812 list_add_tail(&dfrag->list, &msk->rtx_queue); 1813 if (!msk->first_pending) 1814 WRITE_ONCE(msk->first_pending, dfrag); 1815 } 1816 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk, 1817 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1818 !dfrag_collapsed); 1819 1820 continue; 1821 1822 wait_for_memory: 1823 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1824 __mptcp_push_pending(sk, msg->msg_flags); 1825 ret = sk_stream_wait_memory(sk, &timeo); 1826 if (ret) 1827 goto do_error; 1828 } 1829 1830 if (copied) 1831 __mptcp_push_pending(sk, msg->msg_flags); 1832 1833 out: 1834 release_sock(sk); 1835 return copied; 1836 1837 do_error: 1838 if (copied) 1839 goto out; 1840 1841 copied = sk_stream_error(sk, msg->msg_flags, ret); 1842 goto out; 1843 } 1844 1845 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied); 1846 1847 static int __mptcp_recvmsg_mskq(struct sock *sk, 1848 struct msghdr *msg, 1849 size_t len, int flags, 1850 struct scm_timestamping_internal *tss, 1851 int *cmsg_flags) 1852 { 1853 struct mptcp_sock *msk = mptcp_sk(sk); 1854 struct sk_buff *skb, *tmp; 1855 int copied = 0; 1856 1857 skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) { 1858 u32 offset = MPTCP_SKB_CB(skb)->offset; 1859 u32 data_len = skb->len - offset; 1860 u32 count = min_t(size_t, len - copied, data_len); 1861 int err; 1862 1863 if (!(flags & MSG_TRUNC)) { 1864 err = skb_copy_datagram_msg(skb, offset, msg, count); 1865 if (unlikely(err < 0)) { 1866 if (!copied) 1867 return err; 1868 break; 1869 } 1870 } 1871 1872 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 1873 tcp_update_recv_tstamps(skb, tss); 1874 *cmsg_flags |= MPTCP_CMSG_TS; 1875 } 1876 1877 copied += count; 1878 1879 if (count < data_len) { 1880 if (!(flags & MSG_PEEK)) { 1881 MPTCP_SKB_CB(skb)->offset += count; 1882 MPTCP_SKB_CB(skb)->map_seq += count; 1883 msk->bytes_consumed += count; 1884 } 1885 break; 1886 } 1887 1888 if (!(flags & MSG_PEEK)) { 1889 /* avoid the indirect call, we know the destructor is sock_wfree */ 1890 skb->destructor = NULL; 1891 atomic_sub(skb->truesize, &sk->sk_rmem_alloc); 1892 sk_mem_uncharge(sk, skb->truesize); 1893 __skb_unlink(skb, &sk->sk_receive_queue); 1894 __kfree_skb(skb); 1895 msk->bytes_consumed += count; 1896 } 1897 1898 if (copied >= len) 1899 break; 1900 } 1901 1902 mptcp_rcv_space_adjust(msk, copied); 1903 return copied; 1904 } 1905 1906 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1907 * 1908 * Only difference: Use highest rtt estimate of the subflows in use. 1909 */ 1910 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1911 { 1912 struct mptcp_subflow_context *subflow; 1913 struct sock *sk = (struct sock *)msk; 1914 u8 scaling_ratio = U8_MAX; 1915 u32 time, advmss = 1; 1916 u64 rtt_us, mstamp; 1917 1918 msk_owned_by_me(msk); 1919 1920 if (copied <= 0) 1921 return; 1922 1923 if (!msk->rcvspace_init) 1924 mptcp_rcv_space_init(msk, msk->first); 1925 1926 msk->rcvq_space.copied += copied; 1927 1928 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1929 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1930 1931 rtt_us = msk->rcvq_space.rtt_us; 1932 if (rtt_us && time < (rtt_us >> 3)) 1933 return; 1934 1935 rtt_us = 0; 1936 mptcp_for_each_subflow(msk, subflow) { 1937 const struct tcp_sock *tp; 1938 u64 sf_rtt_us; 1939 u32 sf_advmss; 1940 1941 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1942 1943 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1944 sf_advmss = READ_ONCE(tp->advmss); 1945 1946 rtt_us = max(sf_rtt_us, rtt_us); 1947 advmss = max(sf_advmss, advmss); 1948 scaling_ratio = min(tp->scaling_ratio, scaling_ratio); 1949 } 1950 1951 msk->rcvq_space.rtt_us = rtt_us; 1952 msk->scaling_ratio = scaling_ratio; 1953 if (time < (rtt_us >> 3) || rtt_us == 0) 1954 return; 1955 1956 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 1957 goto new_measure; 1958 1959 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 1960 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 1961 u64 rcvwin, grow; 1962 int rcvbuf; 1963 1964 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 1965 1966 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 1967 1968 do_div(grow, msk->rcvq_space.space); 1969 rcvwin += (grow << 1); 1970 1971 rcvbuf = min_t(u64, mptcp_space_from_win(sk, rcvwin), 1972 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 1973 1974 if (rcvbuf > sk->sk_rcvbuf) { 1975 u32 window_clamp; 1976 1977 window_clamp = mptcp_win_from_space(sk, rcvbuf); 1978 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 1979 1980 /* Make subflows follow along. If we do not do this, we 1981 * get drops at subflow level if skbs can't be moved to 1982 * the mptcp rx queue fast enough (announced rcv_win can 1983 * exceed ssk->sk_rcvbuf). 1984 */ 1985 mptcp_for_each_subflow(msk, subflow) { 1986 struct sock *ssk; 1987 bool slow; 1988 1989 ssk = mptcp_subflow_tcp_sock(subflow); 1990 slow = lock_sock_fast(ssk); 1991 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 1992 WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp); 1993 if (tcp_can_send_ack(ssk)) 1994 tcp_cleanup_rbuf(ssk, 1); 1995 unlock_sock_fast(ssk, slow); 1996 } 1997 } 1998 } 1999 2000 msk->rcvq_space.space = msk->rcvq_space.copied; 2001 new_measure: 2002 msk->rcvq_space.copied = 0; 2003 msk->rcvq_space.time = mstamp; 2004 } 2005 2006 static struct mptcp_subflow_context * 2007 __mptcp_first_ready_from(struct mptcp_sock *msk, 2008 struct mptcp_subflow_context *subflow) 2009 { 2010 struct mptcp_subflow_context *start_subflow = subflow; 2011 2012 while (!READ_ONCE(subflow->data_avail)) { 2013 subflow = mptcp_next_subflow(msk, subflow); 2014 if (subflow == start_subflow) 2015 return NULL; 2016 } 2017 return subflow; 2018 } 2019 2020 static bool __mptcp_move_skbs(struct sock *sk) 2021 { 2022 struct mptcp_subflow_context *subflow; 2023 struct mptcp_sock *msk = mptcp_sk(sk); 2024 bool ret = false; 2025 2026 if (list_empty(&msk->conn_list)) 2027 return false; 2028 2029 /* verify we can move any data from the subflow, eventually updating */ 2030 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 2031 mptcp_for_each_subflow(msk, subflow) 2032 __mptcp_rcvbuf_update(sk, subflow->tcp_sock); 2033 2034 subflow = list_first_entry(&msk->conn_list, 2035 struct mptcp_subflow_context, node); 2036 for (;;) { 2037 struct sock *ssk; 2038 bool slowpath; 2039 2040 /* 2041 * As an optimization avoid traversing the subflows list 2042 * and ev. acquiring the subflow socket lock before baling out 2043 */ 2044 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 2045 break; 2046 2047 subflow = __mptcp_first_ready_from(msk, subflow); 2048 if (!subflow) 2049 break; 2050 2051 ssk = mptcp_subflow_tcp_sock(subflow); 2052 slowpath = lock_sock_fast(ssk); 2053 ret = __mptcp_move_skbs_from_subflow(msk, ssk) || ret; 2054 if (unlikely(ssk->sk_err)) 2055 __mptcp_error_report(sk); 2056 unlock_sock_fast(ssk, slowpath); 2057 2058 subflow = mptcp_next_subflow(msk, subflow); 2059 } 2060 2061 __mptcp_ofo_queue(msk); 2062 if (ret) 2063 mptcp_check_data_fin((struct sock *)msk); 2064 return ret; 2065 } 2066 2067 static unsigned int mptcp_inq_hint(const struct sock *sk) 2068 { 2069 const struct mptcp_sock *msk = mptcp_sk(sk); 2070 const struct sk_buff *skb; 2071 2072 skb = skb_peek(&sk->sk_receive_queue); 2073 if (skb) { 2074 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq; 2075 2076 if (hint_val >= INT_MAX) 2077 return INT_MAX; 2078 2079 return (unsigned int)hint_val; 2080 } 2081 2082 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2083 return 1; 2084 2085 return 0; 2086 } 2087 2088 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2089 int flags, int *addr_len) 2090 { 2091 struct mptcp_sock *msk = mptcp_sk(sk); 2092 struct scm_timestamping_internal tss; 2093 int copied = 0, cmsg_flags = 0; 2094 int target; 2095 long timeo; 2096 2097 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2098 if (unlikely(flags & MSG_ERRQUEUE)) 2099 return inet_recv_error(sk, msg, len, addr_len); 2100 2101 lock_sock(sk); 2102 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2103 copied = -ENOTCONN; 2104 goto out_err; 2105 } 2106 2107 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2108 2109 len = min_t(size_t, len, INT_MAX); 2110 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2111 2112 if (unlikely(msk->recvmsg_inq)) 2113 cmsg_flags = MPTCP_CMSG_INQ; 2114 2115 while (copied < len) { 2116 int err, bytes_read; 2117 2118 bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags, &tss, &cmsg_flags); 2119 if (unlikely(bytes_read < 0)) { 2120 if (!copied) 2121 copied = bytes_read; 2122 goto out_err; 2123 } 2124 2125 copied += bytes_read; 2126 2127 if (skb_queue_empty(&sk->sk_receive_queue) && __mptcp_move_skbs(sk)) 2128 continue; 2129 2130 /* only the MPTCP socket status is relevant here. The exit 2131 * conditions mirror closely tcp_recvmsg() 2132 */ 2133 if (copied >= target) 2134 break; 2135 2136 if (copied) { 2137 if (sk->sk_err || 2138 sk->sk_state == TCP_CLOSE || 2139 (sk->sk_shutdown & RCV_SHUTDOWN) || 2140 !timeo || 2141 signal_pending(current)) 2142 break; 2143 } else { 2144 if (sk->sk_err) { 2145 copied = sock_error(sk); 2146 break; 2147 } 2148 2149 if (sk->sk_shutdown & RCV_SHUTDOWN) { 2150 /* race breaker: the shutdown could be after the 2151 * previous receive queue check 2152 */ 2153 if (__mptcp_move_skbs(sk)) 2154 continue; 2155 break; 2156 } 2157 2158 if (sk->sk_state == TCP_CLOSE) { 2159 copied = -ENOTCONN; 2160 break; 2161 } 2162 2163 if (!timeo) { 2164 copied = -EAGAIN; 2165 break; 2166 } 2167 2168 if (signal_pending(current)) { 2169 copied = sock_intr_errno(timeo); 2170 break; 2171 } 2172 } 2173 2174 pr_debug("block timeout %ld\n", timeo); 2175 mptcp_cleanup_rbuf(msk, copied); 2176 err = sk_wait_data(sk, &timeo, NULL); 2177 if (err < 0) { 2178 err = copied ? : err; 2179 goto out_err; 2180 } 2181 } 2182 2183 mptcp_cleanup_rbuf(msk, copied); 2184 2185 out_err: 2186 if (cmsg_flags && copied >= 0) { 2187 if (cmsg_flags & MPTCP_CMSG_TS) 2188 tcp_recv_timestamp(msg, sk, &tss); 2189 2190 if (cmsg_flags & MPTCP_CMSG_INQ) { 2191 unsigned int inq = mptcp_inq_hint(sk); 2192 2193 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2194 } 2195 } 2196 2197 pr_debug("msk=%p rx queue empty=%d copied=%d\n", 2198 msk, skb_queue_empty(&sk->sk_receive_queue), copied); 2199 2200 release_sock(sk); 2201 return copied; 2202 } 2203 2204 static void mptcp_retransmit_timer(struct timer_list *t) 2205 { 2206 struct inet_connection_sock *icsk = timer_container_of(icsk, t, 2207 icsk_retransmit_timer); 2208 struct sock *sk = &icsk->icsk_inet.sk; 2209 struct mptcp_sock *msk = mptcp_sk(sk); 2210 2211 bh_lock_sock(sk); 2212 if (!sock_owned_by_user(sk)) { 2213 /* we need a process context to retransmit */ 2214 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2215 mptcp_schedule_work(sk); 2216 } else { 2217 /* delegate our work to tcp_release_cb() */ 2218 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2219 } 2220 bh_unlock_sock(sk); 2221 sock_put(sk); 2222 } 2223 2224 static void mptcp_tout_timer(struct timer_list *t) 2225 { 2226 struct sock *sk = timer_container_of(sk, t, sk_timer); 2227 2228 mptcp_schedule_work(sk); 2229 sock_put(sk); 2230 } 2231 2232 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2233 * level. 2234 * 2235 * A backup subflow is returned only if that is the only kind available. 2236 */ 2237 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2238 { 2239 struct sock *backup = NULL, *pick = NULL; 2240 struct mptcp_subflow_context *subflow; 2241 int min_stale_count = INT_MAX; 2242 2243 mptcp_for_each_subflow(msk, subflow) { 2244 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2245 2246 if (!__mptcp_subflow_active(subflow)) 2247 continue; 2248 2249 /* still data outstanding at TCP level? skip this */ 2250 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2251 mptcp_pm_subflow_chk_stale(msk, ssk); 2252 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2253 continue; 2254 } 2255 2256 if (subflow->backup || subflow->request_bkup) { 2257 if (!backup) 2258 backup = ssk; 2259 continue; 2260 } 2261 2262 if (!pick) 2263 pick = ssk; 2264 } 2265 2266 if (pick) 2267 return pick; 2268 2269 /* use backup only if there are no progresses anywhere */ 2270 return min_stale_count > 1 ? backup : NULL; 2271 } 2272 2273 bool __mptcp_retransmit_pending_data(struct sock *sk) 2274 { 2275 struct mptcp_data_frag *cur, *rtx_head; 2276 struct mptcp_sock *msk = mptcp_sk(sk); 2277 2278 if (__mptcp_check_fallback(msk)) 2279 return false; 2280 2281 /* the closing socket has some data untransmitted and/or unacked: 2282 * some data in the mptcp rtx queue has not really xmitted yet. 2283 * keep it simple and re-inject the whole mptcp level rtx queue 2284 */ 2285 mptcp_data_lock(sk); 2286 __mptcp_clean_una_wakeup(sk); 2287 rtx_head = mptcp_rtx_head(sk); 2288 if (!rtx_head) { 2289 mptcp_data_unlock(sk); 2290 return false; 2291 } 2292 2293 msk->recovery_snd_nxt = msk->snd_nxt; 2294 msk->recovery = true; 2295 mptcp_data_unlock(sk); 2296 2297 msk->first_pending = rtx_head; 2298 msk->snd_burst = 0; 2299 2300 /* be sure to clear the "sent status" on all re-injected fragments */ 2301 list_for_each_entry(cur, &msk->rtx_queue, list) { 2302 if (!cur->already_sent) 2303 break; 2304 cur->already_sent = 0; 2305 } 2306 2307 return true; 2308 } 2309 2310 /* flags for __mptcp_close_ssk() */ 2311 #define MPTCP_CF_PUSH BIT(1) 2312 #define MPTCP_CF_FASTCLOSE BIT(2) 2313 2314 /* be sure to send a reset only if the caller asked for it, also 2315 * clean completely the subflow status when the subflow reaches 2316 * TCP_CLOSE state 2317 */ 2318 static void __mptcp_subflow_disconnect(struct sock *ssk, 2319 struct mptcp_subflow_context *subflow, 2320 unsigned int flags) 2321 { 2322 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 2323 (flags & MPTCP_CF_FASTCLOSE)) { 2324 /* The MPTCP code never wait on the subflow sockets, TCP-level 2325 * disconnect should never fail 2326 */ 2327 WARN_ON_ONCE(tcp_disconnect(ssk, 0)); 2328 mptcp_subflow_ctx_reset(subflow); 2329 } else { 2330 tcp_shutdown(ssk, SEND_SHUTDOWN); 2331 } 2332 } 2333 2334 /* subflow sockets can be either outgoing (connect) or incoming 2335 * (accept). 2336 * 2337 * Outgoing subflows use in-kernel sockets. 2338 * Incoming subflows do not have their own 'struct socket' allocated, 2339 * so we need to use tcp_close() after detaching them from the mptcp 2340 * parent socket. 2341 */ 2342 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2343 struct mptcp_subflow_context *subflow, 2344 unsigned int flags) 2345 { 2346 struct mptcp_sock *msk = mptcp_sk(sk); 2347 bool dispose_it, need_push = false; 2348 2349 /* If the first subflow moved to a close state before accept, e.g. due 2350 * to an incoming reset or listener shutdown, the subflow socket is 2351 * already deleted by inet_child_forget() and the mptcp socket can't 2352 * survive too. 2353 */ 2354 if (msk->in_accept_queue && msk->first == ssk && 2355 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) { 2356 /* ensure later check in mptcp_worker() will dispose the msk */ 2357 sock_set_flag(sk, SOCK_DEAD); 2358 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1)); 2359 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2360 mptcp_subflow_drop_ctx(ssk); 2361 goto out_release; 2362 } 2363 2364 dispose_it = msk->free_first || ssk != msk->first; 2365 if (dispose_it) 2366 list_del(&subflow->node); 2367 2368 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2369 2370 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) { 2371 /* be sure to force the tcp_close path 2372 * to generate the egress reset 2373 */ 2374 ssk->sk_lingertime = 0; 2375 sock_set_flag(ssk, SOCK_LINGER); 2376 subflow->send_fastclose = 1; 2377 } 2378 2379 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2380 if (!dispose_it) { 2381 __mptcp_subflow_disconnect(ssk, subflow, flags); 2382 release_sock(ssk); 2383 2384 goto out; 2385 } 2386 2387 subflow->disposable = 1; 2388 2389 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2390 * the ssk has been already destroyed, we just need to release the 2391 * reference owned by msk; 2392 */ 2393 if (!inet_csk(ssk)->icsk_ulp_ops) { 2394 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD)); 2395 kfree_rcu(subflow, rcu); 2396 } else { 2397 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2398 __tcp_close(ssk, 0); 2399 2400 /* close acquired an extra ref */ 2401 __sock_put(ssk); 2402 } 2403 2404 out_release: 2405 __mptcp_subflow_error_report(sk, ssk); 2406 release_sock(ssk); 2407 2408 sock_put(ssk); 2409 2410 if (ssk == msk->first) 2411 WRITE_ONCE(msk->first, NULL); 2412 2413 out: 2414 __mptcp_sync_sndbuf(sk); 2415 if (need_push) 2416 __mptcp_push_pending(sk, 0); 2417 2418 /* Catch every 'all subflows closed' scenario, including peers silently 2419 * closing them, e.g. due to timeout. 2420 * For established sockets, allow an additional timeout before closing, 2421 * as the protocol can still create more subflows. 2422 */ 2423 if (list_is_singular(&msk->conn_list) && msk->first && 2424 inet_sk_state_load(msk->first) == TCP_CLOSE) { 2425 if (sk->sk_state != TCP_ESTABLISHED || 2426 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) { 2427 mptcp_set_state(sk, TCP_CLOSE); 2428 mptcp_close_wake_up(sk); 2429 } else { 2430 mptcp_start_tout_timer(sk); 2431 } 2432 } 2433 } 2434 2435 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2436 struct mptcp_subflow_context *subflow) 2437 { 2438 /* The first subflow can already be closed and still in the list */ 2439 if (subflow->close_event_done) 2440 return; 2441 2442 subflow->close_event_done = true; 2443 2444 if (sk->sk_state == TCP_ESTABLISHED) 2445 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2446 2447 /* subflow aborted before reaching the fully_established status 2448 * attempt the creation of the next subflow 2449 */ 2450 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow); 2451 2452 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2453 } 2454 2455 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2456 { 2457 return 0; 2458 } 2459 2460 static void __mptcp_close_subflow(struct sock *sk) 2461 { 2462 struct mptcp_subflow_context *subflow, *tmp; 2463 struct mptcp_sock *msk = mptcp_sk(sk); 2464 2465 might_sleep(); 2466 2467 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2468 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2469 int ssk_state = inet_sk_state_load(ssk); 2470 2471 if (ssk_state != TCP_CLOSE && 2472 (ssk_state != TCP_CLOSE_WAIT || 2473 inet_sk_state_load(sk) != TCP_ESTABLISHED)) 2474 continue; 2475 2476 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2477 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2478 continue; 2479 2480 mptcp_close_ssk(sk, ssk, subflow); 2481 } 2482 2483 } 2484 2485 static bool mptcp_close_tout_expired(const struct sock *sk) 2486 { 2487 if (!inet_csk(sk)->icsk_mtup.probe_timestamp || 2488 sk->sk_state == TCP_CLOSE) 2489 return false; 2490 2491 return time_after32(tcp_jiffies32, 2492 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk)); 2493 } 2494 2495 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2496 { 2497 struct mptcp_subflow_context *subflow, *tmp; 2498 struct sock *sk = (struct sock *)msk; 2499 2500 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2501 return; 2502 2503 mptcp_token_destroy(msk); 2504 2505 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2506 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2507 bool slow; 2508 2509 slow = lock_sock_fast(tcp_sk); 2510 if (tcp_sk->sk_state != TCP_CLOSE) { 2511 mptcp_send_active_reset_reason(tcp_sk); 2512 tcp_set_state(tcp_sk, TCP_CLOSE); 2513 } 2514 unlock_sock_fast(tcp_sk, slow); 2515 } 2516 2517 /* Mirror the tcp_reset() error propagation */ 2518 switch (sk->sk_state) { 2519 case TCP_SYN_SENT: 2520 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 2521 break; 2522 case TCP_CLOSE_WAIT: 2523 WRITE_ONCE(sk->sk_err, EPIPE); 2524 break; 2525 case TCP_CLOSE: 2526 return; 2527 default: 2528 WRITE_ONCE(sk->sk_err, ECONNRESET); 2529 } 2530 2531 mptcp_set_state(sk, TCP_CLOSE); 2532 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2533 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2534 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2535 2536 /* the calling mptcp_worker will properly destroy the socket */ 2537 if (sock_flag(sk, SOCK_DEAD)) 2538 return; 2539 2540 sk->sk_state_change(sk); 2541 sk_error_report(sk); 2542 } 2543 2544 static void __mptcp_retrans(struct sock *sk) 2545 { 2546 struct mptcp_sock *msk = mptcp_sk(sk); 2547 struct mptcp_subflow_context *subflow; 2548 struct mptcp_sendmsg_info info = {}; 2549 struct mptcp_data_frag *dfrag; 2550 struct sock *ssk; 2551 int ret, err; 2552 u16 len = 0; 2553 2554 mptcp_clean_una_wakeup(sk); 2555 2556 /* first check ssk: need to kick "stale" logic */ 2557 err = mptcp_sched_get_retrans(msk); 2558 dfrag = mptcp_rtx_head(sk); 2559 if (!dfrag) { 2560 if (mptcp_data_fin_enabled(msk)) { 2561 struct inet_connection_sock *icsk = inet_csk(sk); 2562 2563 icsk->icsk_retransmits++; 2564 mptcp_set_datafin_timeout(sk); 2565 mptcp_send_ack(msk); 2566 2567 goto reset_timer; 2568 } 2569 2570 if (!mptcp_send_head(sk)) 2571 return; 2572 2573 goto reset_timer; 2574 } 2575 2576 if (err) 2577 goto reset_timer; 2578 2579 mptcp_for_each_subflow(msk, subflow) { 2580 if (READ_ONCE(subflow->scheduled)) { 2581 u16 copied = 0; 2582 2583 mptcp_subflow_set_scheduled(subflow, false); 2584 2585 ssk = mptcp_subflow_tcp_sock(subflow); 2586 2587 lock_sock(ssk); 2588 2589 /* limit retransmission to the bytes already sent on some subflows */ 2590 info.sent = 0; 2591 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : 2592 dfrag->already_sent; 2593 while (info.sent < info.limit) { 2594 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2595 if (ret <= 0) 2596 break; 2597 2598 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2599 copied += ret; 2600 info.sent += ret; 2601 } 2602 if (copied) { 2603 len = max(copied, len); 2604 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2605 info.size_goal); 2606 WRITE_ONCE(msk->allow_infinite_fallback, false); 2607 } 2608 2609 release_sock(ssk); 2610 } 2611 } 2612 2613 msk->bytes_retrans += len; 2614 dfrag->already_sent = max(dfrag->already_sent, len); 2615 2616 reset_timer: 2617 mptcp_check_and_set_pending(sk); 2618 2619 if (!mptcp_rtx_timer_pending(sk)) 2620 mptcp_reset_rtx_timer(sk); 2621 } 2622 2623 /* schedule the timeout timer for the relevant event: either close timeout 2624 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one 2625 */ 2626 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout) 2627 { 2628 struct sock *sk = (struct sock *)msk; 2629 unsigned long timeout, close_timeout; 2630 2631 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp) 2632 return; 2633 2634 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp - 2635 tcp_jiffies32 + jiffies + mptcp_close_timeout(sk); 2636 2637 /* the close timeout takes precedence on the fail one, and here at least one of 2638 * them is active 2639 */ 2640 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout; 2641 2642 sk_reset_timer(sk, &sk->sk_timer, timeout); 2643 } 2644 2645 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2646 { 2647 struct sock *ssk = msk->first; 2648 bool slow; 2649 2650 if (!ssk) 2651 return; 2652 2653 pr_debug("MP_FAIL doesn't respond, reset the subflow\n"); 2654 2655 slow = lock_sock_fast(ssk); 2656 mptcp_subflow_reset(ssk); 2657 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0); 2658 unlock_sock_fast(ssk, slow); 2659 } 2660 2661 static void mptcp_do_fastclose(struct sock *sk) 2662 { 2663 struct mptcp_subflow_context *subflow, *tmp; 2664 struct mptcp_sock *msk = mptcp_sk(sk); 2665 2666 mptcp_set_state(sk, TCP_CLOSE); 2667 mptcp_for_each_subflow_safe(msk, subflow, tmp) 2668 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), 2669 subflow, MPTCP_CF_FASTCLOSE); 2670 } 2671 2672 static void mptcp_worker(struct work_struct *work) 2673 { 2674 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2675 struct sock *sk = (struct sock *)msk; 2676 unsigned long fail_tout; 2677 int state; 2678 2679 lock_sock(sk); 2680 state = sk->sk_state; 2681 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN))) 2682 goto unlock; 2683 2684 mptcp_check_fastclose(msk); 2685 2686 mptcp_pm_worker(msk); 2687 2688 mptcp_check_send_data_fin(sk); 2689 mptcp_check_data_fin_ack(sk); 2690 mptcp_check_data_fin(sk); 2691 2692 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2693 __mptcp_close_subflow(sk); 2694 2695 if (mptcp_close_tout_expired(sk)) { 2696 mptcp_do_fastclose(sk); 2697 mptcp_close_wake_up(sk); 2698 } 2699 2700 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) { 2701 __mptcp_destroy_sock(sk); 2702 goto unlock; 2703 } 2704 2705 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2706 __mptcp_retrans(sk); 2707 2708 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0; 2709 if (fail_tout && time_after(jiffies, fail_tout)) 2710 mptcp_mp_fail_no_response(msk); 2711 2712 unlock: 2713 release_sock(sk); 2714 sock_put(sk); 2715 } 2716 2717 static void __mptcp_init_sock(struct sock *sk) 2718 { 2719 struct mptcp_sock *msk = mptcp_sk(sk); 2720 2721 INIT_LIST_HEAD(&msk->conn_list); 2722 INIT_LIST_HEAD(&msk->join_list); 2723 INIT_LIST_HEAD(&msk->rtx_queue); 2724 INIT_WORK(&msk->work, mptcp_worker); 2725 msk->out_of_order_queue = RB_ROOT; 2726 msk->first_pending = NULL; 2727 msk->timer_ival = TCP_RTO_MIN; 2728 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO; 2729 2730 WRITE_ONCE(msk->first, NULL); 2731 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2732 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2733 WRITE_ONCE(msk->allow_infinite_fallback, true); 2734 msk->recovery = false; 2735 msk->subflow_id = 1; 2736 msk->last_data_sent = tcp_jiffies32; 2737 msk->last_data_recv = tcp_jiffies32; 2738 msk->last_ack_recv = tcp_jiffies32; 2739 2740 mptcp_pm_data_init(msk); 2741 2742 /* re-use the csk retrans timer for MPTCP-level retrans */ 2743 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2744 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0); 2745 } 2746 2747 static void mptcp_ca_reset(struct sock *sk) 2748 { 2749 struct inet_connection_sock *icsk = inet_csk(sk); 2750 2751 tcp_assign_congestion_control(sk); 2752 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name, 2753 sizeof(mptcp_sk(sk)->ca_name)); 2754 2755 /* no need to keep a reference to the ops, the name will suffice */ 2756 tcp_cleanup_congestion_control(sk); 2757 icsk->icsk_ca_ops = NULL; 2758 } 2759 2760 static int mptcp_init_sock(struct sock *sk) 2761 { 2762 struct net *net = sock_net(sk); 2763 int ret; 2764 2765 __mptcp_init_sock(sk); 2766 2767 if (!mptcp_is_enabled(net)) 2768 return -ENOPROTOOPT; 2769 2770 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2771 return -ENOMEM; 2772 2773 rcu_read_lock(); 2774 ret = mptcp_init_sched(mptcp_sk(sk), 2775 mptcp_sched_find(mptcp_get_scheduler(net))); 2776 rcu_read_unlock(); 2777 if (ret) 2778 return ret; 2779 2780 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); 2781 2782 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2783 * propagate the correct value 2784 */ 2785 mptcp_ca_reset(sk); 2786 2787 sk_sockets_allocated_inc(sk); 2788 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]); 2789 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]); 2790 2791 return 0; 2792 } 2793 2794 static void __mptcp_clear_xmit(struct sock *sk) 2795 { 2796 struct mptcp_sock *msk = mptcp_sk(sk); 2797 struct mptcp_data_frag *dtmp, *dfrag; 2798 2799 WRITE_ONCE(msk->first_pending, NULL); 2800 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2801 dfrag_clear(sk, dfrag); 2802 } 2803 2804 void mptcp_cancel_work(struct sock *sk) 2805 { 2806 struct mptcp_sock *msk = mptcp_sk(sk); 2807 2808 if (cancel_work_sync(&msk->work)) 2809 __sock_put(sk); 2810 } 2811 2812 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2813 { 2814 lock_sock(ssk); 2815 2816 switch (ssk->sk_state) { 2817 case TCP_LISTEN: 2818 if (!(how & RCV_SHUTDOWN)) 2819 break; 2820 fallthrough; 2821 case TCP_SYN_SENT: 2822 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK)); 2823 break; 2824 default: 2825 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2826 pr_debug("Fallback\n"); 2827 ssk->sk_shutdown |= how; 2828 tcp_shutdown(ssk, how); 2829 2830 /* simulate the data_fin ack reception to let the state 2831 * machine move forward 2832 */ 2833 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt); 2834 mptcp_schedule_work(sk); 2835 } else { 2836 pr_debug("Sending DATA_FIN on subflow %p\n", ssk); 2837 tcp_send_ack(ssk); 2838 if (!mptcp_rtx_timer_pending(sk)) 2839 mptcp_reset_rtx_timer(sk); 2840 } 2841 break; 2842 } 2843 2844 release_sock(ssk); 2845 } 2846 2847 void mptcp_set_state(struct sock *sk, int state) 2848 { 2849 int oldstate = sk->sk_state; 2850 2851 switch (state) { 2852 case TCP_ESTABLISHED: 2853 if (oldstate != TCP_ESTABLISHED) 2854 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2855 break; 2856 case TCP_CLOSE_WAIT: 2857 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state: 2858 * MPTCP "accepted" sockets will be created later on. So no 2859 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT. 2860 */ 2861 break; 2862 default: 2863 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 2864 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2865 } 2866 2867 inet_sk_state_store(sk, state); 2868 } 2869 2870 static const unsigned char new_state[16] = { 2871 /* current state: new state: action: */ 2872 [0 /* (Invalid) */] = TCP_CLOSE, 2873 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2874 [TCP_SYN_SENT] = TCP_CLOSE, 2875 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2876 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2877 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2878 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2879 [TCP_CLOSE] = TCP_CLOSE, 2880 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2881 [TCP_LAST_ACK] = TCP_LAST_ACK, 2882 [TCP_LISTEN] = TCP_CLOSE, 2883 [TCP_CLOSING] = TCP_CLOSING, 2884 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2885 }; 2886 2887 static int mptcp_close_state(struct sock *sk) 2888 { 2889 int next = (int)new_state[sk->sk_state]; 2890 int ns = next & TCP_STATE_MASK; 2891 2892 mptcp_set_state(sk, ns); 2893 2894 return next & TCP_ACTION_FIN; 2895 } 2896 2897 static void mptcp_check_send_data_fin(struct sock *sk) 2898 { 2899 struct mptcp_subflow_context *subflow; 2900 struct mptcp_sock *msk = mptcp_sk(sk); 2901 2902 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n", 2903 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2904 msk->snd_nxt, msk->write_seq); 2905 2906 /* we still need to enqueue subflows or not really shutting down, 2907 * skip this 2908 */ 2909 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2910 mptcp_send_head(sk)) 2911 return; 2912 2913 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2914 2915 mptcp_for_each_subflow(msk, subflow) { 2916 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2917 2918 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2919 } 2920 } 2921 2922 static void __mptcp_wr_shutdown(struct sock *sk) 2923 { 2924 struct mptcp_sock *msk = mptcp_sk(sk); 2925 2926 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n", 2927 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2928 !!mptcp_send_head(sk)); 2929 2930 /* will be ignored by fallback sockets */ 2931 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2932 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2933 2934 mptcp_check_send_data_fin(sk); 2935 } 2936 2937 static void __mptcp_destroy_sock(struct sock *sk) 2938 { 2939 struct mptcp_sock *msk = mptcp_sk(sk); 2940 2941 pr_debug("msk=%p\n", msk); 2942 2943 might_sleep(); 2944 2945 mptcp_stop_rtx_timer(sk); 2946 sk_stop_timer(sk, &sk->sk_timer); 2947 msk->pm.status = 0; 2948 mptcp_release_sched(msk); 2949 2950 sk->sk_prot->destroy(sk); 2951 2952 sk_stream_kill_queues(sk); 2953 xfrm_sk_free_policy(sk); 2954 2955 sock_put(sk); 2956 } 2957 2958 void __mptcp_unaccepted_force_close(struct sock *sk) 2959 { 2960 sock_set_flag(sk, SOCK_DEAD); 2961 mptcp_do_fastclose(sk); 2962 __mptcp_destroy_sock(sk); 2963 } 2964 2965 static __poll_t mptcp_check_readable(struct sock *sk) 2966 { 2967 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0; 2968 } 2969 2970 static void mptcp_check_listen_stop(struct sock *sk) 2971 { 2972 struct sock *ssk; 2973 2974 if (inet_sk_state_load(sk) != TCP_LISTEN) 2975 return; 2976 2977 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 2978 ssk = mptcp_sk(sk)->first; 2979 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN)) 2980 return; 2981 2982 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2983 tcp_set_state(ssk, TCP_CLOSE); 2984 mptcp_subflow_queue_clean(sk, ssk); 2985 inet_csk_listen_stop(ssk); 2986 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED); 2987 release_sock(ssk); 2988 } 2989 2990 bool __mptcp_close(struct sock *sk, long timeout) 2991 { 2992 struct mptcp_subflow_context *subflow; 2993 struct mptcp_sock *msk = mptcp_sk(sk); 2994 bool do_cancel_work = false; 2995 int subflows_alive = 0; 2996 2997 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2998 2999 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 3000 mptcp_check_listen_stop(sk); 3001 mptcp_set_state(sk, TCP_CLOSE); 3002 goto cleanup; 3003 } 3004 3005 if (mptcp_data_avail(msk) || timeout < 0) { 3006 /* If the msk has read data, or the caller explicitly ask it, 3007 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose 3008 */ 3009 mptcp_do_fastclose(sk); 3010 timeout = 0; 3011 } else if (mptcp_close_state(sk)) { 3012 __mptcp_wr_shutdown(sk); 3013 } 3014 3015 sk_stream_wait_close(sk, timeout); 3016 3017 cleanup: 3018 /* orphan all the subflows */ 3019 mptcp_for_each_subflow(msk, subflow) { 3020 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3021 bool slow = lock_sock_fast_nested(ssk); 3022 3023 subflows_alive += ssk->sk_state != TCP_CLOSE; 3024 3025 /* since the close timeout takes precedence on the fail one, 3026 * cancel the latter 3027 */ 3028 if (ssk == msk->first) 3029 subflow->fail_tout = 0; 3030 3031 /* detach from the parent socket, but allow data_ready to 3032 * push incoming data into the mptcp stack, to properly ack it 3033 */ 3034 ssk->sk_socket = NULL; 3035 ssk->sk_wq = NULL; 3036 unlock_sock_fast(ssk, slow); 3037 } 3038 sock_orphan(sk); 3039 3040 /* all the subflows are closed, only timeout can change the msk 3041 * state, let's not keep resources busy for no reasons 3042 */ 3043 if (subflows_alive == 0) 3044 mptcp_set_state(sk, TCP_CLOSE); 3045 3046 sock_hold(sk); 3047 pr_debug("msk=%p state=%d\n", sk, sk->sk_state); 3048 mptcp_pm_connection_closed(msk); 3049 3050 if (sk->sk_state == TCP_CLOSE) { 3051 __mptcp_destroy_sock(sk); 3052 do_cancel_work = true; 3053 } else { 3054 mptcp_start_tout_timer(sk); 3055 } 3056 3057 return do_cancel_work; 3058 } 3059 3060 static void mptcp_close(struct sock *sk, long timeout) 3061 { 3062 bool do_cancel_work; 3063 3064 lock_sock(sk); 3065 3066 do_cancel_work = __mptcp_close(sk, timeout); 3067 release_sock(sk); 3068 if (do_cancel_work) 3069 mptcp_cancel_work(sk); 3070 3071 sock_put(sk); 3072 } 3073 3074 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 3075 { 3076 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3077 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 3078 struct ipv6_pinfo *msk6 = inet6_sk(msk); 3079 3080 msk->sk_v6_daddr = ssk->sk_v6_daddr; 3081 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 3082 3083 if (msk6 && ssk6) { 3084 msk6->saddr = ssk6->saddr; 3085 msk6->flow_label = ssk6->flow_label; 3086 } 3087 #endif 3088 3089 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 3090 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 3091 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 3092 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 3093 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 3094 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 3095 } 3096 3097 static int mptcp_disconnect(struct sock *sk, int flags) 3098 { 3099 struct mptcp_sock *msk = mptcp_sk(sk); 3100 3101 /* We are on the fastopen error path. We can't call straight into the 3102 * subflows cleanup code due to lock nesting (we are already under 3103 * msk->firstsocket lock). 3104 */ 3105 if (msk->fastopening) 3106 return -EBUSY; 3107 3108 mptcp_check_listen_stop(sk); 3109 mptcp_set_state(sk, TCP_CLOSE); 3110 3111 mptcp_stop_rtx_timer(sk); 3112 mptcp_stop_tout_timer(sk); 3113 3114 mptcp_pm_connection_closed(msk); 3115 3116 /* msk->subflow is still intact, the following will not free the first 3117 * subflow 3118 */ 3119 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE); 3120 WRITE_ONCE(msk->flags, 0); 3121 msk->cb_flags = 0; 3122 msk->recovery = false; 3123 WRITE_ONCE(msk->can_ack, false); 3124 WRITE_ONCE(msk->fully_established, false); 3125 WRITE_ONCE(msk->rcv_data_fin, false); 3126 WRITE_ONCE(msk->snd_data_fin_enable, false); 3127 WRITE_ONCE(msk->rcv_fastclose, false); 3128 WRITE_ONCE(msk->use_64bit_ack, false); 3129 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 3130 mptcp_pm_data_reset(msk); 3131 mptcp_ca_reset(sk); 3132 msk->bytes_consumed = 0; 3133 msk->bytes_acked = 0; 3134 msk->bytes_received = 0; 3135 msk->bytes_sent = 0; 3136 msk->bytes_retrans = 0; 3137 msk->rcvspace_init = 0; 3138 3139 WRITE_ONCE(sk->sk_shutdown, 0); 3140 sk_error_report(sk); 3141 return 0; 3142 } 3143 3144 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3145 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 3146 { 3147 struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk); 3148 3149 return &msk6->np; 3150 } 3151 3152 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk) 3153 { 3154 const struct ipv6_pinfo *np = inet6_sk(sk); 3155 struct ipv6_txoptions *opt; 3156 struct ipv6_pinfo *newnp; 3157 3158 newnp = inet6_sk(newsk); 3159 3160 rcu_read_lock(); 3161 opt = rcu_dereference(np->opt); 3162 if (opt) { 3163 opt = ipv6_dup_options(newsk, opt); 3164 if (!opt) 3165 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__); 3166 } 3167 RCU_INIT_POINTER(newnp->opt, opt); 3168 rcu_read_unlock(); 3169 } 3170 #endif 3171 3172 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk) 3173 { 3174 struct ip_options_rcu *inet_opt, *newopt = NULL; 3175 const struct inet_sock *inet = inet_sk(sk); 3176 struct inet_sock *newinet; 3177 3178 newinet = inet_sk(newsk); 3179 3180 rcu_read_lock(); 3181 inet_opt = rcu_dereference(inet->inet_opt); 3182 if (inet_opt) { 3183 newopt = sock_kmemdup(newsk, inet_opt, sizeof(*inet_opt) + 3184 inet_opt->opt.optlen, GFP_ATOMIC); 3185 if (!newopt) 3186 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__); 3187 } 3188 RCU_INIT_POINTER(newinet->inet_opt, newopt); 3189 rcu_read_unlock(); 3190 } 3191 3192 struct sock *mptcp_sk_clone_init(const struct sock *sk, 3193 const struct mptcp_options_received *mp_opt, 3194 struct sock *ssk, 3195 struct request_sock *req) 3196 { 3197 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 3198 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 3199 struct mptcp_subflow_context *subflow; 3200 struct mptcp_sock *msk; 3201 3202 if (!nsk) 3203 return NULL; 3204 3205 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3206 if (nsk->sk_family == AF_INET6) 3207 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 3208 #endif 3209 3210 __mptcp_init_sock(nsk); 3211 3212 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3213 if (nsk->sk_family == AF_INET6) 3214 mptcp_copy_ip6_options(nsk, sk); 3215 else 3216 #endif 3217 mptcp_copy_ip_options(nsk, sk); 3218 3219 msk = mptcp_sk(nsk); 3220 WRITE_ONCE(msk->local_key, subflow_req->local_key); 3221 WRITE_ONCE(msk->token, subflow_req->token); 3222 msk->in_accept_queue = 1; 3223 WRITE_ONCE(msk->fully_established, false); 3224 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 3225 WRITE_ONCE(msk->csum_enabled, true); 3226 3227 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1); 3228 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3229 WRITE_ONCE(msk->snd_una, msk->write_seq); 3230 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 3231 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 3232 mptcp_init_sched(msk, mptcp_sk(sk)->sched); 3233 3234 /* passive msk is created after the first/MPC subflow */ 3235 msk->subflow_id = 2; 3236 3237 sock_reset_flag(nsk, SOCK_RCU_FREE); 3238 security_inet_csk_clone(nsk, req); 3239 3240 /* this can't race with mptcp_close(), as the msk is 3241 * not yet exposted to user-space 3242 */ 3243 mptcp_set_state(nsk, TCP_ESTABLISHED); 3244 3245 /* The msk maintain a ref to each subflow in the connections list */ 3246 WRITE_ONCE(msk->first, ssk); 3247 subflow = mptcp_subflow_ctx(ssk); 3248 list_add(&subflow->node, &msk->conn_list); 3249 sock_hold(ssk); 3250 3251 /* new mpc subflow takes ownership of the newly 3252 * created mptcp socket 3253 */ 3254 mptcp_token_accept(subflow_req, msk); 3255 3256 /* set msk addresses early to ensure mptcp_pm_get_local_id() 3257 * uses the correct data 3258 */ 3259 mptcp_copy_inaddrs(nsk, ssk); 3260 __mptcp_propagate_sndbuf(nsk, ssk); 3261 3262 mptcp_rcv_space_init(msk, ssk); 3263 3264 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 3265 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 3266 bh_unlock_sock(nsk); 3267 3268 /* note: the newly allocated socket refcount is 2 now */ 3269 return nsk; 3270 } 3271 3272 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3273 { 3274 const struct tcp_sock *tp = tcp_sk(ssk); 3275 3276 msk->rcvspace_init = 1; 3277 msk->rcvq_space.copied = 0; 3278 msk->rcvq_space.rtt_us = 0; 3279 3280 msk->rcvq_space.time = tp->tcp_mstamp; 3281 3282 /* initial rcv_space offering made to peer */ 3283 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3284 TCP_INIT_CWND * tp->advmss); 3285 if (msk->rcvq_space.space == 0) 3286 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3287 } 3288 3289 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags) 3290 { 3291 struct mptcp_subflow_context *subflow, *tmp; 3292 struct sock *sk = (struct sock *)msk; 3293 3294 __mptcp_clear_xmit(sk); 3295 3296 /* join list will be eventually flushed (with rst) at sock lock release time */ 3297 mptcp_for_each_subflow_safe(msk, subflow, tmp) 3298 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags); 3299 3300 __skb_queue_purge(&sk->sk_receive_queue); 3301 skb_rbtree_purge(&msk->out_of_order_queue); 3302 3303 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3304 * inet_sock_destruct() will dispose it 3305 */ 3306 mptcp_token_destroy(msk); 3307 mptcp_pm_destroy(msk); 3308 } 3309 3310 static void mptcp_destroy(struct sock *sk) 3311 { 3312 struct mptcp_sock *msk = mptcp_sk(sk); 3313 3314 /* allow the following to close even the initial subflow */ 3315 msk->free_first = 1; 3316 mptcp_destroy_common(msk, 0); 3317 sk_sockets_allocated_dec(sk); 3318 } 3319 3320 void __mptcp_data_acked(struct sock *sk) 3321 { 3322 if (!sock_owned_by_user(sk)) 3323 __mptcp_clean_una(sk); 3324 else 3325 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3326 } 3327 3328 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3329 { 3330 if (!mptcp_send_head(sk)) 3331 return; 3332 3333 if (!sock_owned_by_user(sk)) 3334 __mptcp_subflow_push_pending(sk, ssk, false); 3335 else 3336 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3337 } 3338 3339 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3340 BIT(MPTCP_RETRANSMIT) | \ 3341 BIT(MPTCP_FLUSH_JOIN_LIST) | \ 3342 BIT(MPTCP_DEQUEUE)) 3343 3344 /* processes deferred events and flush wmem */ 3345 static void mptcp_release_cb(struct sock *sk) 3346 __must_hold(&sk->sk_lock.slock) 3347 { 3348 struct mptcp_sock *msk = mptcp_sk(sk); 3349 3350 for (;;) { 3351 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED); 3352 struct list_head join_list; 3353 3354 if (!flags) 3355 break; 3356 3357 INIT_LIST_HEAD(&join_list); 3358 list_splice_init(&msk->join_list, &join_list); 3359 3360 /* the following actions acquire the subflow socket lock 3361 * 3362 * 1) can't be invoked in atomic scope 3363 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3364 * datapath acquires the msk socket spinlock while helding 3365 * the subflow socket lock 3366 */ 3367 msk->cb_flags &= ~flags; 3368 spin_unlock_bh(&sk->sk_lock.slock); 3369 3370 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3371 __mptcp_flush_join_list(sk, &join_list); 3372 if (flags & BIT(MPTCP_PUSH_PENDING)) 3373 __mptcp_push_pending(sk, 0); 3374 if (flags & BIT(MPTCP_RETRANSMIT)) 3375 __mptcp_retrans(sk); 3376 if ((flags & BIT(MPTCP_DEQUEUE)) && __mptcp_move_skbs(sk)) { 3377 /* notify ack seq update */ 3378 mptcp_cleanup_rbuf(msk, 0); 3379 sk->sk_data_ready(sk); 3380 } 3381 3382 cond_resched(); 3383 spin_lock_bh(&sk->sk_lock.slock); 3384 } 3385 3386 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3387 __mptcp_clean_una_wakeup(sk); 3388 if (unlikely(msk->cb_flags)) { 3389 /* be sure to sync the msk state before taking actions 3390 * depending on sk_state (MPTCP_ERROR_REPORT) 3391 * On sk release avoid actions depending on the first subflow 3392 */ 3393 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first) 3394 __mptcp_sync_state(sk, msk->pending_state); 3395 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3396 __mptcp_error_report(sk); 3397 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags)) 3398 __mptcp_sync_sndbuf(sk); 3399 } 3400 } 3401 3402 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3403 * TCP can't schedule delack timer before the subflow is fully established. 3404 * MPTCP uses the delack timer to do 3rd ack retransmissions 3405 */ 3406 static void schedule_3rdack_retransmission(struct sock *ssk) 3407 { 3408 struct inet_connection_sock *icsk = inet_csk(ssk); 3409 struct tcp_sock *tp = tcp_sk(ssk); 3410 unsigned long timeout; 3411 3412 if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established)) 3413 return; 3414 3415 /* reschedule with a timeout above RTT, as we must look only for drop */ 3416 if (tp->srtt_us) 3417 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3418 else 3419 timeout = TCP_TIMEOUT_INIT; 3420 timeout += jiffies; 3421 3422 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3423 smp_store_release(&icsk->icsk_ack.pending, 3424 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER); 3425 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3426 } 3427 3428 void mptcp_subflow_process_delegated(struct sock *ssk, long status) 3429 { 3430 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3431 struct sock *sk = subflow->conn; 3432 3433 if (status & BIT(MPTCP_DELEGATE_SEND)) { 3434 mptcp_data_lock(sk); 3435 if (!sock_owned_by_user(sk)) 3436 __mptcp_subflow_push_pending(sk, ssk, true); 3437 else 3438 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3439 mptcp_data_unlock(sk); 3440 } 3441 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) { 3442 mptcp_data_lock(sk); 3443 if (!sock_owned_by_user(sk)) 3444 __mptcp_sync_sndbuf(sk); 3445 else 3446 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags); 3447 mptcp_data_unlock(sk); 3448 } 3449 if (status & BIT(MPTCP_DELEGATE_ACK)) 3450 schedule_3rdack_retransmission(ssk); 3451 } 3452 3453 static int mptcp_hash(struct sock *sk) 3454 { 3455 /* should never be called, 3456 * we hash the TCP subflows not the MPTCP socket 3457 */ 3458 WARN_ON_ONCE(1); 3459 return 0; 3460 } 3461 3462 static void mptcp_unhash(struct sock *sk) 3463 { 3464 /* called from sk_common_release(), but nothing to do here */ 3465 } 3466 3467 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3468 { 3469 struct mptcp_sock *msk = mptcp_sk(sk); 3470 3471 pr_debug("msk=%p, ssk=%p\n", msk, msk->first); 3472 if (WARN_ON_ONCE(!msk->first)) 3473 return -EINVAL; 3474 3475 return inet_csk_get_port(msk->first, snum); 3476 } 3477 3478 void mptcp_finish_connect(struct sock *ssk) 3479 { 3480 struct mptcp_subflow_context *subflow; 3481 struct mptcp_sock *msk; 3482 struct sock *sk; 3483 3484 subflow = mptcp_subflow_ctx(ssk); 3485 sk = subflow->conn; 3486 msk = mptcp_sk(sk); 3487 3488 pr_debug("msk=%p, token=%u\n", sk, subflow->token); 3489 3490 subflow->map_seq = subflow->iasn; 3491 subflow->map_subflow_seq = 1; 3492 3493 /* the socket is not connected yet, no msk/subflow ops can access/race 3494 * accessing the field below 3495 */ 3496 WRITE_ONCE(msk->local_key, subflow->local_key); 3497 3498 mptcp_pm_new_connection(msk, ssk, 0); 3499 } 3500 3501 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3502 { 3503 write_lock_bh(&sk->sk_callback_lock); 3504 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3505 sk_set_socket(sk, parent); 3506 sk->sk_uid = SOCK_INODE(parent)->i_uid; 3507 write_unlock_bh(&sk->sk_callback_lock); 3508 } 3509 3510 bool mptcp_finish_join(struct sock *ssk) 3511 { 3512 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3513 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3514 struct sock *parent = (void *)msk; 3515 bool ret = true; 3516 3517 pr_debug("msk=%p, subflow=%p\n", msk, subflow); 3518 3519 /* mptcp socket already closing? */ 3520 if (!mptcp_is_fully_established(parent)) { 3521 subflow->reset_reason = MPTCP_RST_EMPTCP; 3522 return false; 3523 } 3524 3525 /* active subflow, already present inside the conn_list */ 3526 if (!list_empty(&subflow->node)) { 3527 mptcp_subflow_joined(msk, ssk); 3528 mptcp_propagate_sndbuf(parent, ssk); 3529 return true; 3530 } 3531 3532 if (!mptcp_pm_allow_new_subflow(msk)) { 3533 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_JOINREJECTED); 3534 goto err_prohibited; 3535 } 3536 3537 /* If we can't acquire msk socket lock here, let the release callback 3538 * handle it 3539 */ 3540 mptcp_data_lock(parent); 3541 if (!sock_owned_by_user(parent)) { 3542 ret = __mptcp_finish_join(msk, ssk); 3543 if (ret) { 3544 sock_hold(ssk); 3545 list_add_tail(&subflow->node, &msk->conn_list); 3546 } 3547 } else { 3548 sock_hold(ssk); 3549 list_add_tail(&subflow->node, &msk->join_list); 3550 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3551 } 3552 mptcp_data_unlock(parent); 3553 3554 if (!ret) { 3555 err_prohibited: 3556 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3557 return false; 3558 } 3559 3560 return true; 3561 } 3562 3563 static void mptcp_shutdown(struct sock *sk, int how) 3564 { 3565 pr_debug("sk=%p, how=%d\n", sk, how); 3566 3567 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3568 __mptcp_wr_shutdown(sk); 3569 } 3570 3571 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3572 { 3573 const struct sock *sk = (void *)msk; 3574 u64 delta; 3575 3576 if (sk->sk_state == TCP_LISTEN) 3577 return -EINVAL; 3578 3579 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3580 return 0; 3581 3582 delta = msk->write_seq - v; 3583 if (__mptcp_check_fallback(msk) && msk->first) { 3584 struct tcp_sock *tp = tcp_sk(msk->first); 3585 3586 /* the first subflow is disconnected after close - see 3587 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3588 * so ignore that status, too. 3589 */ 3590 if (!((1 << msk->first->sk_state) & 3591 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3592 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3593 } 3594 if (delta > INT_MAX) 3595 delta = INT_MAX; 3596 3597 return (int)delta; 3598 } 3599 3600 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg) 3601 { 3602 struct mptcp_sock *msk = mptcp_sk(sk); 3603 bool slow; 3604 3605 switch (cmd) { 3606 case SIOCINQ: 3607 if (sk->sk_state == TCP_LISTEN) 3608 return -EINVAL; 3609 3610 lock_sock(sk); 3611 if (__mptcp_move_skbs(sk)) 3612 mptcp_cleanup_rbuf(msk, 0); 3613 *karg = mptcp_inq_hint(sk); 3614 release_sock(sk); 3615 break; 3616 case SIOCOUTQ: 3617 slow = lock_sock_fast(sk); 3618 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3619 unlock_sock_fast(sk, slow); 3620 break; 3621 case SIOCOUTQNSD: 3622 slow = lock_sock_fast(sk); 3623 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt); 3624 unlock_sock_fast(sk, slow); 3625 break; 3626 default: 3627 return -ENOIOCTLCMD; 3628 } 3629 3630 return 0; 3631 } 3632 3633 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) 3634 { 3635 struct mptcp_subflow_context *subflow; 3636 struct mptcp_sock *msk = mptcp_sk(sk); 3637 int err = -EINVAL; 3638 struct sock *ssk; 3639 3640 ssk = __mptcp_nmpc_sk(msk); 3641 if (IS_ERR(ssk)) 3642 return PTR_ERR(ssk); 3643 3644 mptcp_set_state(sk, TCP_SYN_SENT); 3645 subflow = mptcp_subflow_ctx(ssk); 3646 #ifdef CONFIG_TCP_MD5SIG 3647 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3648 * TCP option space. 3649 */ 3650 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info)) 3651 mptcp_subflow_early_fallback(msk, subflow); 3652 #endif 3653 if (subflow->request_mptcp) { 3654 if (mptcp_active_should_disable(sk)) { 3655 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEACTIVEDISABLED); 3656 mptcp_subflow_early_fallback(msk, subflow); 3657 } else if (mptcp_token_new_connect(ssk) < 0) { 3658 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT); 3659 mptcp_subflow_early_fallback(msk, subflow); 3660 } 3661 } 3662 3663 WRITE_ONCE(msk->write_seq, subflow->idsn); 3664 WRITE_ONCE(msk->snd_nxt, subflow->idsn); 3665 WRITE_ONCE(msk->snd_una, subflow->idsn); 3666 if (likely(!__mptcp_check_fallback(msk))) 3667 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE); 3668 3669 /* if reaching here via the fastopen/sendmsg path, the caller already 3670 * acquired the subflow socket lock, too. 3671 */ 3672 if (!msk->fastopening) 3673 lock_sock(ssk); 3674 3675 /* the following mirrors closely a very small chunk of code from 3676 * __inet_stream_connect() 3677 */ 3678 if (ssk->sk_state != TCP_CLOSE) 3679 goto out; 3680 3681 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) { 3682 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len); 3683 if (err) 3684 goto out; 3685 } 3686 3687 err = ssk->sk_prot->connect(ssk, uaddr, addr_len); 3688 if (err < 0) 3689 goto out; 3690 3691 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk)); 3692 3693 out: 3694 if (!msk->fastopening) 3695 release_sock(ssk); 3696 3697 /* on successful connect, the msk state will be moved to established by 3698 * subflow_finish_connect() 3699 */ 3700 if (unlikely(err)) { 3701 /* avoid leaving a dangling token in an unconnected socket */ 3702 mptcp_token_destroy(msk); 3703 mptcp_set_state(sk, TCP_CLOSE); 3704 return err; 3705 } 3706 3707 mptcp_copy_inaddrs(sk, ssk); 3708 return 0; 3709 } 3710 3711 static struct proto mptcp_prot = { 3712 .name = "MPTCP", 3713 .owner = THIS_MODULE, 3714 .init = mptcp_init_sock, 3715 .connect = mptcp_connect, 3716 .disconnect = mptcp_disconnect, 3717 .close = mptcp_close, 3718 .setsockopt = mptcp_setsockopt, 3719 .getsockopt = mptcp_getsockopt, 3720 .shutdown = mptcp_shutdown, 3721 .destroy = mptcp_destroy, 3722 .sendmsg = mptcp_sendmsg, 3723 .ioctl = mptcp_ioctl, 3724 .recvmsg = mptcp_recvmsg, 3725 .release_cb = mptcp_release_cb, 3726 .hash = mptcp_hash, 3727 .unhash = mptcp_unhash, 3728 .get_port = mptcp_get_port, 3729 .stream_memory_free = mptcp_stream_memory_free, 3730 .sockets_allocated = &mptcp_sockets_allocated, 3731 3732 .memory_allocated = &tcp_memory_allocated, 3733 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc, 3734 3735 .memory_pressure = &tcp_memory_pressure, 3736 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3737 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3738 .sysctl_mem = sysctl_tcp_mem, 3739 .obj_size = sizeof(struct mptcp_sock), 3740 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3741 .no_autobind = true, 3742 }; 3743 3744 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3745 { 3746 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3747 struct sock *ssk, *sk = sock->sk; 3748 int err = -EINVAL; 3749 3750 lock_sock(sk); 3751 ssk = __mptcp_nmpc_sk(msk); 3752 if (IS_ERR(ssk)) { 3753 err = PTR_ERR(ssk); 3754 goto unlock; 3755 } 3756 3757 if (sk->sk_family == AF_INET) 3758 err = inet_bind_sk(ssk, uaddr, addr_len); 3759 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3760 else if (sk->sk_family == AF_INET6) 3761 err = inet6_bind_sk(ssk, uaddr, addr_len); 3762 #endif 3763 if (!err) 3764 mptcp_copy_inaddrs(sk, ssk); 3765 3766 unlock: 3767 release_sock(sk); 3768 return err; 3769 } 3770 3771 static int mptcp_listen(struct socket *sock, int backlog) 3772 { 3773 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3774 struct sock *sk = sock->sk; 3775 struct sock *ssk; 3776 int err; 3777 3778 pr_debug("msk=%p\n", msk); 3779 3780 lock_sock(sk); 3781 3782 err = -EINVAL; 3783 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM) 3784 goto unlock; 3785 3786 ssk = __mptcp_nmpc_sk(msk); 3787 if (IS_ERR(ssk)) { 3788 err = PTR_ERR(ssk); 3789 goto unlock; 3790 } 3791 3792 mptcp_set_state(sk, TCP_LISTEN); 3793 sock_set_flag(sk, SOCK_RCU_FREE); 3794 3795 lock_sock(ssk); 3796 err = __inet_listen_sk(ssk, backlog); 3797 release_sock(ssk); 3798 mptcp_set_state(sk, inet_sk_state_load(ssk)); 3799 3800 if (!err) { 3801 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 3802 mptcp_copy_inaddrs(sk, ssk); 3803 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED); 3804 } 3805 3806 unlock: 3807 release_sock(sk); 3808 return err; 3809 } 3810 3811 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3812 struct proto_accept_arg *arg) 3813 { 3814 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3815 struct sock *ssk, *newsk; 3816 3817 pr_debug("msk=%p\n", msk); 3818 3819 /* Buggy applications can call accept on socket states other then LISTEN 3820 * but no need to allocate the first subflow just to error out. 3821 */ 3822 ssk = READ_ONCE(msk->first); 3823 if (!ssk) 3824 return -EINVAL; 3825 3826 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk)); 3827 newsk = inet_csk_accept(ssk, arg); 3828 if (!newsk) 3829 return arg->err; 3830 3831 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk)); 3832 if (sk_is_mptcp(newsk)) { 3833 struct mptcp_subflow_context *subflow; 3834 struct sock *new_mptcp_sock; 3835 3836 subflow = mptcp_subflow_ctx(newsk); 3837 new_mptcp_sock = subflow->conn; 3838 3839 /* is_mptcp should be false if subflow->conn is missing, see 3840 * subflow_syn_recv_sock() 3841 */ 3842 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3843 tcp_sk(newsk)->is_mptcp = 0; 3844 goto tcpfallback; 3845 } 3846 3847 newsk = new_mptcp_sock; 3848 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3849 3850 newsk->sk_kern_sock = arg->kern; 3851 lock_sock(newsk); 3852 __inet_accept(sock, newsock, newsk); 3853 3854 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags); 3855 msk = mptcp_sk(newsk); 3856 msk->in_accept_queue = 0; 3857 3858 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3859 * This is needed so NOSPACE flag can be set from tcp stack. 3860 */ 3861 mptcp_for_each_subflow(msk, subflow) { 3862 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3863 3864 if (!ssk->sk_socket) 3865 mptcp_sock_graft(ssk, newsock); 3866 } 3867 3868 /* Do late cleanup for the first subflow as necessary. Also 3869 * deal with bad peers not doing a complete shutdown. 3870 */ 3871 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) { 3872 __mptcp_close_ssk(newsk, msk->first, 3873 mptcp_subflow_ctx(msk->first), 0); 3874 if (unlikely(list_is_singular(&msk->conn_list))) 3875 mptcp_set_state(newsk, TCP_CLOSE); 3876 } 3877 } else { 3878 tcpfallback: 3879 newsk->sk_kern_sock = arg->kern; 3880 lock_sock(newsk); 3881 __inet_accept(sock, newsock, newsk); 3882 /* we are being invoked after accepting a non-mp-capable 3883 * flow: sk is a tcp_sk, not an mptcp one. 3884 * 3885 * Hand the socket over to tcp so all further socket ops 3886 * bypass mptcp. 3887 */ 3888 WRITE_ONCE(newsock->sk->sk_socket->ops, 3889 mptcp_fallback_tcp_ops(newsock->sk)); 3890 } 3891 release_sock(newsk); 3892 3893 return 0; 3894 } 3895 3896 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3897 { 3898 struct sock *sk = (struct sock *)msk; 3899 3900 if (__mptcp_stream_is_writeable(sk, 1)) 3901 return EPOLLOUT | EPOLLWRNORM; 3902 3903 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 3904 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */ 3905 if (__mptcp_stream_is_writeable(sk, 1)) 3906 return EPOLLOUT | EPOLLWRNORM; 3907 3908 return 0; 3909 } 3910 3911 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3912 struct poll_table_struct *wait) 3913 { 3914 struct sock *sk = sock->sk; 3915 struct mptcp_sock *msk; 3916 __poll_t mask = 0; 3917 u8 shutdown; 3918 int state; 3919 3920 msk = mptcp_sk(sk); 3921 sock_poll_wait(file, sock, wait); 3922 3923 state = inet_sk_state_load(sk); 3924 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags); 3925 if (state == TCP_LISTEN) { 3926 struct sock *ssk = READ_ONCE(msk->first); 3927 3928 if (WARN_ON_ONCE(!ssk)) 3929 return 0; 3930 3931 return inet_csk_listen_poll(ssk); 3932 } 3933 3934 shutdown = READ_ONCE(sk->sk_shutdown); 3935 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 3936 mask |= EPOLLHUP; 3937 if (shutdown & RCV_SHUTDOWN) 3938 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 3939 3940 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 3941 mask |= mptcp_check_readable(sk); 3942 if (shutdown & SEND_SHUTDOWN) 3943 mask |= EPOLLOUT | EPOLLWRNORM; 3944 else 3945 mask |= mptcp_check_writeable(msk); 3946 } else if (state == TCP_SYN_SENT && 3947 inet_test_bit(DEFER_CONNECT, sk)) { 3948 /* cf tcp_poll() note about TFO */ 3949 mask |= EPOLLOUT | EPOLLWRNORM; 3950 } 3951 3952 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */ 3953 smp_rmb(); 3954 if (READ_ONCE(sk->sk_err)) 3955 mask |= EPOLLERR; 3956 3957 return mask; 3958 } 3959 3960 static const struct proto_ops mptcp_stream_ops = { 3961 .family = PF_INET, 3962 .owner = THIS_MODULE, 3963 .release = inet_release, 3964 .bind = mptcp_bind, 3965 .connect = inet_stream_connect, 3966 .socketpair = sock_no_socketpair, 3967 .accept = mptcp_stream_accept, 3968 .getname = inet_getname, 3969 .poll = mptcp_poll, 3970 .ioctl = inet_ioctl, 3971 .gettstamp = sock_gettstamp, 3972 .listen = mptcp_listen, 3973 .shutdown = inet_shutdown, 3974 .setsockopt = sock_common_setsockopt, 3975 .getsockopt = sock_common_getsockopt, 3976 .sendmsg = inet_sendmsg, 3977 .recvmsg = inet_recvmsg, 3978 .mmap = sock_no_mmap, 3979 .set_rcvlowat = mptcp_set_rcvlowat, 3980 }; 3981 3982 static struct inet_protosw mptcp_protosw = { 3983 .type = SOCK_STREAM, 3984 .protocol = IPPROTO_MPTCP, 3985 .prot = &mptcp_prot, 3986 .ops = &mptcp_stream_ops, 3987 .flags = INET_PROTOSW_ICSK, 3988 }; 3989 3990 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 3991 { 3992 struct mptcp_delegated_action *delegated; 3993 struct mptcp_subflow_context *subflow; 3994 int work_done = 0; 3995 3996 delegated = container_of(napi, struct mptcp_delegated_action, napi); 3997 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 3998 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3999 4000 bh_lock_sock_nested(ssk); 4001 if (!sock_owned_by_user(ssk)) { 4002 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0)); 4003 } else { 4004 /* tcp_release_cb_override already processed 4005 * the action or will do at next release_sock(). 4006 * In both case must dequeue the subflow here - on the same 4007 * CPU that scheduled it. 4008 */ 4009 smp_wmb(); 4010 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status); 4011 } 4012 bh_unlock_sock(ssk); 4013 sock_put(ssk); 4014 4015 if (++work_done == budget) 4016 return budget; 4017 } 4018 4019 /* always provide a 0 'work_done' argument, so that napi_complete_done 4020 * will not try accessing the NULL napi->dev ptr 4021 */ 4022 napi_complete_done(napi, 0); 4023 return work_done; 4024 } 4025 4026 void __init mptcp_proto_init(void) 4027 { 4028 struct mptcp_delegated_action *delegated; 4029 int cpu; 4030 4031 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 4032 4033 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 4034 panic("Failed to allocate MPTCP pcpu counter\n"); 4035 4036 mptcp_napi_dev = alloc_netdev_dummy(0); 4037 if (!mptcp_napi_dev) 4038 panic("Failed to allocate MPTCP dummy netdev\n"); 4039 for_each_possible_cpu(cpu) { 4040 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 4041 INIT_LIST_HEAD(&delegated->head); 4042 netif_napi_add_tx(mptcp_napi_dev, &delegated->napi, 4043 mptcp_napi_poll); 4044 napi_enable(&delegated->napi); 4045 } 4046 4047 mptcp_subflow_init(); 4048 mptcp_pm_init(); 4049 mptcp_sched_init(); 4050 mptcp_token_init(); 4051 4052 if (proto_register(&mptcp_prot, 1) != 0) 4053 panic("Failed to register MPTCP proto.\n"); 4054 4055 inet_register_protosw(&mptcp_protosw); 4056 4057 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 4058 } 4059 4060 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 4061 static const struct proto_ops mptcp_v6_stream_ops = { 4062 .family = PF_INET6, 4063 .owner = THIS_MODULE, 4064 .release = inet6_release, 4065 .bind = mptcp_bind, 4066 .connect = inet_stream_connect, 4067 .socketpair = sock_no_socketpair, 4068 .accept = mptcp_stream_accept, 4069 .getname = inet6_getname, 4070 .poll = mptcp_poll, 4071 .ioctl = inet6_ioctl, 4072 .gettstamp = sock_gettstamp, 4073 .listen = mptcp_listen, 4074 .shutdown = inet_shutdown, 4075 .setsockopt = sock_common_setsockopt, 4076 .getsockopt = sock_common_getsockopt, 4077 .sendmsg = inet6_sendmsg, 4078 .recvmsg = inet6_recvmsg, 4079 .mmap = sock_no_mmap, 4080 #ifdef CONFIG_COMPAT 4081 .compat_ioctl = inet6_compat_ioctl, 4082 #endif 4083 .set_rcvlowat = mptcp_set_rcvlowat, 4084 }; 4085 4086 static struct proto mptcp_v6_prot; 4087 4088 static struct inet_protosw mptcp_v6_protosw = { 4089 .type = SOCK_STREAM, 4090 .protocol = IPPROTO_MPTCP, 4091 .prot = &mptcp_v6_prot, 4092 .ops = &mptcp_v6_stream_ops, 4093 .flags = INET_PROTOSW_ICSK, 4094 }; 4095 4096 int __init mptcp_proto_v6_init(void) 4097 { 4098 int err; 4099 4100 mptcp_v6_prot = mptcp_prot; 4101 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name)); 4102 mptcp_v6_prot.slab = NULL; 4103 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 4104 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np); 4105 4106 err = proto_register(&mptcp_v6_prot, 1); 4107 if (err) 4108 return err; 4109 4110 err = inet6_register_protosw(&mptcp_v6_protosw); 4111 if (err) 4112 proto_unregister(&mptcp_v6_prot); 4113 4114 return err; 4115 } 4116 #endif 4117