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/aligned_data.h> 15 #include <net/sock.h> 16 #include <net/inet_common.h> 17 #include <net/inet_hashtables.h> 18 #include <net/protocol.h> 19 #include <net/tcp_states.h> 20 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 21 #include <net/transp_v6.h> 22 #endif 23 #include <net/mptcp.h> 24 #include <net/hotdata.h> 25 #include <net/xfrm.h> 26 #include <asm/ioctls.h> 27 #include "protocol.h" 28 #include "mib.h" 29 30 #define CREATE_TRACE_POINTS 31 #include <trace/events/mptcp.h> 32 33 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 34 struct mptcp6_sock { 35 struct mptcp_sock msk; 36 struct ipv6_pinfo np; 37 }; 38 #endif 39 40 enum { 41 MPTCP_CMSG_TS = BIT(0), 42 MPTCP_CMSG_INQ = BIT(1), 43 }; 44 45 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp; 46 47 static void __mptcp_destroy_sock(struct sock *sk); 48 static void mptcp_check_send_data_fin(struct sock *sk); 49 50 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions) = { 51 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 52 }; 53 static struct net_device *mptcp_napi_dev; 54 55 /* Returns end sequence number of the receiver's advertised window */ 56 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 57 { 58 return READ_ONCE(msk->wnd_end); 59 } 60 61 static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk) 62 { 63 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 64 if (sk->sk_prot == &tcpv6_prot) 65 return &inet6_stream_ops; 66 #endif 67 WARN_ON_ONCE(sk->sk_prot != &tcp_prot); 68 return &inet_stream_ops; 69 } 70 71 bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib) 72 { 73 struct net *net = sock_net((struct sock *)msk); 74 75 if (__mptcp_check_fallback(msk)) 76 return true; 77 78 spin_lock_bh(&msk->fallback_lock); 79 if (!msk->allow_infinite_fallback) { 80 spin_unlock_bh(&msk->fallback_lock); 81 return false; 82 } 83 84 msk->allow_subflows = false; 85 set_bit(MPTCP_FALLBACK_DONE, &msk->flags); 86 __MPTCP_INC_STATS(net, fb_mib); 87 spin_unlock_bh(&msk->fallback_lock); 88 return true; 89 } 90 91 static int __mptcp_socket_create(struct mptcp_sock *msk) 92 { 93 struct mptcp_subflow_context *subflow; 94 struct sock *sk = (struct sock *)msk; 95 struct socket *ssock; 96 int err; 97 98 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock); 99 if (err) 100 return err; 101 102 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio; 103 WRITE_ONCE(msk->first, ssock->sk); 104 subflow = mptcp_subflow_ctx(ssock->sk); 105 list_add(&subflow->node, &msk->conn_list); 106 sock_hold(ssock->sk); 107 subflow->request_mptcp = 1; 108 subflow->subflow_id = msk->subflow_id++; 109 110 /* This is the first subflow, always with id 0 */ 111 WRITE_ONCE(subflow->local_id, 0); 112 mptcp_sock_graft(msk->first, sk->sk_socket); 113 iput(SOCK_INODE(ssock)); 114 115 return 0; 116 } 117 118 /* If the MPC handshake is not started, returns the first subflow, 119 * eventually allocating it. 120 */ 121 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk) 122 { 123 struct sock *sk = (struct sock *)msk; 124 int ret; 125 126 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 127 return ERR_PTR(-EINVAL); 128 129 if (!msk->first) { 130 ret = __mptcp_socket_create(msk); 131 if (ret) 132 return ERR_PTR(ret); 133 } 134 135 return msk->first; 136 } 137 138 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 139 { 140 sk_drops_add(sk, skb); 141 __kfree_skb(skb); 142 } 143 144 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 145 struct sk_buff *from) 146 { 147 bool fragstolen; 148 int delta; 149 150 if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) || 151 MPTCP_SKB_CB(from)->offset || 152 ((to->len + from->len) > (sk->sk_rcvbuf >> 3)) || 153 !skb_try_coalesce(to, from, &fragstolen, &delta)) 154 return false; 155 156 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n", 157 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 158 to->len, MPTCP_SKB_CB(from)->end_seq); 159 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 160 161 /* note the fwd memory can reach a negative value after accounting 162 * for the delta, but the later skb free will restore a non 163 * negative one 164 */ 165 atomic_add(delta, &sk->sk_rmem_alloc); 166 sk_mem_charge(sk, delta); 167 kfree_skb_partial(from, fragstolen); 168 169 return true; 170 } 171 172 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 173 struct sk_buff *from) 174 { 175 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 176 return false; 177 178 return mptcp_try_coalesce((struct sock *)msk, to, from); 179 } 180 181 /* "inspired" by tcp_data_queue_ofo(), main differences: 182 * - use mptcp seqs 183 * - don't cope with sacks 184 */ 185 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 186 { 187 struct sock *sk = (struct sock *)msk; 188 struct rb_node **p, *parent; 189 u64 seq, end_seq, max_seq; 190 struct sk_buff *skb1; 191 192 seq = MPTCP_SKB_CB(skb)->map_seq; 193 end_seq = MPTCP_SKB_CB(skb)->end_seq; 194 max_seq = atomic64_read(&msk->rcv_wnd_sent); 195 196 pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq, 197 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 198 if (after64(end_seq, max_seq)) { 199 /* out of window */ 200 mptcp_drop(sk, skb); 201 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 202 (unsigned long long)end_seq - (unsigned long)max_seq, 203 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent)); 204 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 205 return; 206 } 207 208 p = &msk->out_of_order_queue.rb_node; 209 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 210 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 211 rb_link_node(&skb->rbnode, NULL, p); 212 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 213 msk->ooo_last_skb = skb; 214 goto end; 215 } 216 217 /* with 2 subflows, adding at end of ooo queue is quite likely 218 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 219 */ 220 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 221 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 222 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 223 return; 224 } 225 226 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 227 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 228 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 229 parent = &msk->ooo_last_skb->rbnode; 230 p = &parent->rb_right; 231 goto insert; 232 } 233 234 /* Find place to insert this segment. Handle overlaps on the way. */ 235 parent = NULL; 236 while (*p) { 237 parent = *p; 238 skb1 = rb_to_skb(parent); 239 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 240 p = &parent->rb_left; 241 continue; 242 } 243 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 244 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 245 /* All the bits are present. Drop. */ 246 mptcp_drop(sk, skb); 247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 248 return; 249 } 250 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 251 /* partial overlap: 252 * | skb | 253 * | skb1 | 254 * continue traversing 255 */ 256 } else { 257 /* skb's seq == skb1's seq and skb covers skb1. 258 * Replace skb1 with skb. 259 */ 260 rb_replace_node(&skb1->rbnode, &skb->rbnode, 261 &msk->out_of_order_queue); 262 mptcp_drop(sk, skb1); 263 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 264 goto merge_right; 265 } 266 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 267 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 268 return; 269 } 270 p = &parent->rb_right; 271 } 272 273 insert: 274 /* Insert segment into RB tree. */ 275 rb_link_node(&skb->rbnode, parent, p); 276 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 277 278 merge_right: 279 /* Remove other segments covered by skb. */ 280 while ((skb1 = skb_rb_next(skb)) != NULL) { 281 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 282 break; 283 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 284 mptcp_drop(sk, skb1); 285 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 286 } 287 /* If there is no skb after us, we are the last_skb ! */ 288 if (!skb1) 289 msk->ooo_last_skb = skb; 290 291 end: 292 skb_condense(skb); 293 skb_set_owner_r(skb, sk); 294 } 295 296 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 297 struct sk_buff *skb, unsigned int offset, 298 size_t copy_len) 299 { 300 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 301 struct sock *sk = (struct sock *)msk; 302 struct sk_buff *tail; 303 bool has_rxtstamp; 304 305 __skb_unlink(skb, &ssk->sk_receive_queue); 306 307 skb_ext_reset(skb); 308 skb_orphan(skb); 309 310 /* try to fetch required memory from subflow */ 311 if (!sk_rmem_schedule(sk, skb, skb->truesize)) { 312 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 313 goto drop; 314 } 315 316 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 317 318 /* the skb map_seq accounts for the skb offset: 319 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 320 * value 321 */ 322 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 323 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 324 MPTCP_SKB_CB(skb)->offset = offset; 325 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 326 MPTCP_SKB_CB(skb)->cant_coalesce = 0; 327 328 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 329 /* in sequence */ 330 msk->bytes_received += copy_len; 331 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 332 tail = skb_peek_tail(&sk->sk_receive_queue); 333 if (tail && mptcp_try_coalesce(sk, tail, skb)) 334 return true; 335 336 skb_set_owner_r(skb, sk); 337 __skb_queue_tail(&sk->sk_receive_queue, skb); 338 return true; 339 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 340 mptcp_data_queue_ofo(msk, skb); 341 return false; 342 } 343 344 /* old data, keep it simple and drop the whole pkt, sender 345 * will retransmit as needed, if needed. 346 */ 347 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 348 drop: 349 mptcp_drop(sk, skb); 350 return false; 351 } 352 353 static void mptcp_stop_rtx_timer(struct sock *sk) 354 { 355 struct inet_connection_sock *icsk = inet_csk(sk); 356 357 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 358 mptcp_sk(sk)->timer_ival = 0; 359 } 360 361 static void mptcp_close_wake_up(struct sock *sk) 362 { 363 if (sock_flag(sk, SOCK_DEAD)) 364 return; 365 366 sk->sk_state_change(sk); 367 if (sk->sk_shutdown == SHUTDOWN_MASK || 368 sk->sk_state == TCP_CLOSE) 369 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 370 else 371 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 372 } 373 374 static void mptcp_shutdown_subflows(struct mptcp_sock *msk) 375 { 376 struct mptcp_subflow_context *subflow; 377 378 mptcp_for_each_subflow(msk, subflow) { 379 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 380 bool slow; 381 382 slow = lock_sock_fast(ssk); 383 tcp_shutdown(ssk, SEND_SHUTDOWN); 384 unlock_sock_fast(ssk, slow); 385 } 386 } 387 388 /* called under the msk socket lock */ 389 static bool mptcp_pending_data_fin_ack(struct sock *sk) 390 { 391 struct mptcp_sock *msk = mptcp_sk(sk); 392 393 return ((1 << sk->sk_state) & 394 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 395 msk->write_seq == READ_ONCE(msk->snd_una); 396 } 397 398 static void mptcp_check_data_fin_ack(struct sock *sk) 399 { 400 struct mptcp_sock *msk = mptcp_sk(sk); 401 402 /* Look for an acknowledged DATA_FIN */ 403 if (mptcp_pending_data_fin_ack(sk)) { 404 WRITE_ONCE(msk->snd_data_fin_enable, 0); 405 406 switch (sk->sk_state) { 407 case TCP_FIN_WAIT1: 408 mptcp_set_state(sk, TCP_FIN_WAIT2); 409 break; 410 case TCP_CLOSING: 411 case TCP_LAST_ACK: 412 mptcp_shutdown_subflows(msk); 413 mptcp_set_state(sk, TCP_CLOSE); 414 break; 415 } 416 417 mptcp_close_wake_up(sk); 418 } 419 } 420 421 /* can be called with no lock acquired */ 422 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 423 { 424 struct mptcp_sock *msk = mptcp_sk(sk); 425 426 if (READ_ONCE(msk->rcv_data_fin) && 427 ((1 << inet_sk_state_load(sk)) & 428 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 429 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 430 431 if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) { 432 if (seq) 433 *seq = rcv_data_fin_seq; 434 435 return true; 436 } 437 } 438 439 return false; 440 } 441 442 static void mptcp_set_datafin_timeout(struct sock *sk) 443 { 444 struct inet_connection_sock *icsk = inet_csk(sk); 445 u32 retransmits; 446 447 retransmits = min_t(u32, icsk->icsk_retransmits, 448 ilog2(TCP_RTO_MAX / TCP_RTO_MIN)); 449 450 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits; 451 } 452 453 static void __mptcp_set_timeout(struct sock *sk, long tout) 454 { 455 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 456 } 457 458 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 459 { 460 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 461 462 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 463 icsk_timeout(inet_csk(ssk)) - jiffies : 0; 464 } 465 466 static void mptcp_set_timeout(struct sock *sk) 467 { 468 struct mptcp_subflow_context *subflow; 469 long tout = 0; 470 471 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 472 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 473 __mptcp_set_timeout(sk, tout); 474 } 475 476 static inline bool tcp_can_send_ack(const struct sock *ssk) 477 { 478 return !((1 << inet_sk_state_load(ssk)) & 479 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 480 } 481 482 void __mptcp_subflow_send_ack(struct sock *ssk) 483 { 484 if (tcp_can_send_ack(ssk)) 485 tcp_send_ack(ssk); 486 } 487 488 static void mptcp_subflow_send_ack(struct sock *ssk) 489 { 490 bool slow; 491 492 slow = lock_sock_fast(ssk); 493 __mptcp_subflow_send_ack(ssk); 494 unlock_sock_fast(ssk, slow); 495 } 496 497 static void mptcp_send_ack(struct mptcp_sock *msk) 498 { 499 struct mptcp_subflow_context *subflow; 500 501 mptcp_for_each_subflow(msk, subflow) 502 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow)); 503 } 504 505 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied) 506 { 507 bool slow; 508 509 slow = lock_sock_fast(ssk); 510 if (tcp_can_send_ack(ssk)) 511 tcp_cleanup_rbuf(ssk, copied); 512 unlock_sock_fast(ssk, slow); 513 } 514 515 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 516 { 517 const struct inet_connection_sock *icsk = inet_csk(ssk); 518 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 519 const struct tcp_sock *tp = tcp_sk(ssk); 520 521 return (ack_pending & ICSK_ACK_SCHED) && 522 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 523 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 524 (rx_empty && ack_pending & 525 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 526 } 527 528 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied) 529 { 530 int old_space = READ_ONCE(msk->old_wspace); 531 struct mptcp_subflow_context *subflow; 532 struct sock *sk = (struct sock *)msk; 533 int space = __mptcp_space(sk); 534 bool cleanup, rx_empty; 535 536 cleanup = (space > 0) && (space >= (old_space << 1)) && copied; 537 rx_empty = !sk_rmem_alloc_get(sk) && copied; 538 539 mptcp_for_each_subflow(msk, subflow) { 540 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 541 542 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 543 mptcp_subflow_cleanup_rbuf(ssk, copied); 544 } 545 } 546 547 static bool mptcp_check_data_fin(struct sock *sk) 548 { 549 struct mptcp_sock *msk = mptcp_sk(sk); 550 u64 rcv_data_fin_seq; 551 bool ret = false; 552 553 /* Need to ack a DATA_FIN received from a peer while this side 554 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 555 * msk->rcv_data_fin was set when parsing the incoming options 556 * at the subflow level and the msk lock was not held, so this 557 * is the first opportunity to act on the DATA_FIN and change 558 * the msk state. 559 * 560 * If we are caught up to the sequence number of the incoming 561 * DATA_FIN, send the DATA_ACK now and do state transition. If 562 * not caught up, do nothing and let the recv code send DATA_ACK 563 * when catching up. 564 */ 565 566 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 567 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 568 WRITE_ONCE(msk->rcv_data_fin, 0); 569 570 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 571 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 572 573 switch (sk->sk_state) { 574 case TCP_ESTABLISHED: 575 mptcp_set_state(sk, TCP_CLOSE_WAIT); 576 break; 577 case TCP_FIN_WAIT1: 578 mptcp_set_state(sk, TCP_CLOSING); 579 break; 580 case TCP_FIN_WAIT2: 581 mptcp_shutdown_subflows(msk); 582 mptcp_set_state(sk, TCP_CLOSE); 583 break; 584 default: 585 /* Other states not expected */ 586 WARN_ON_ONCE(1); 587 break; 588 } 589 590 ret = true; 591 if (!__mptcp_check_fallback(msk)) 592 mptcp_send_ack(msk); 593 mptcp_close_wake_up(sk); 594 } 595 return ret; 596 } 597 598 static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk) 599 { 600 if (!mptcp_try_fallback(ssk, MPTCP_MIB_DSSCORRUPTIONFALLBACK)) { 601 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET); 602 mptcp_subflow_reset(ssk); 603 } 604 } 605 606 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 607 struct sock *ssk) 608 { 609 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 610 struct sock *sk = (struct sock *)msk; 611 bool more_data_avail; 612 struct tcp_sock *tp; 613 bool ret = false; 614 615 pr_debug("msk=%p ssk=%p\n", msk, ssk); 616 tp = tcp_sk(ssk); 617 do { 618 u32 map_remaining, offset; 619 u32 seq = tp->copied_seq; 620 struct sk_buff *skb; 621 bool fin; 622 623 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 624 break; 625 626 /* try to move as much data as available */ 627 map_remaining = subflow->map_data_len - 628 mptcp_subflow_get_map_offset(subflow); 629 630 skb = skb_peek(&ssk->sk_receive_queue); 631 if (unlikely(!skb)) 632 break; 633 634 if (__mptcp_check_fallback(msk)) { 635 /* Under fallback skbs have no MPTCP extension and TCP could 636 * collapse them between the dummy map creation and the 637 * current dequeue. Be sure to adjust the map size. 638 */ 639 map_remaining = skb->len; 640 subflow->map_data_len = skb->len; 641 } 642 643 offset = seq - TCP_SKB_CB(skb)->seq; 644 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 645 if (fin) 646 seq++; 647 648 if (offset < skb->len) { 649 size_t len = skb->len - offset; 650 651 ret = __mptcp_move_skb(msk, ssk, skb, offset, len) || ret; 652 seq += len; 653 654 if (unlikely(map_remaining < len)) { 655 DEBUG_NET_WARN_ON_ONCE(1); 656 mptcp_dss_corruption(msk, ssk); 657 } 658 } else { 659 if (unlikely(!fin)) { 660 DEBUG_NET_WARN_ON_ONCE(1); 661 mptcp_dss_corruption(msk, ssk); 662 } 663 664 sk_eat_skb(ssk, skb); 665 } 666 667 WRITE_ONCE(tp->copied_seq, seq); 668 more_data_avail = mptcp_subflow_data_available(ssk); 669 670 } while (more_data_avail); 671 672 if (ret) 673 msk->last_data_recv = tcp_jiffies32; 674 return ret; 675 } 676 677 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 678 { 679 struct sock *sk = (struct sock *)msk; 680 struct sk_buff *skb, *tail; 681 bool moved = false; 682 struct rb_node *p; 683 u64 end_seq; 684 685 p = rb_first(&msk->out_of_order_queue); 686 pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 687 while (p) { 688 skb = rb_to_skb(p); 689 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 690 break; 691 692 p = rb_next(p); 693 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 694 695 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 696 msk->ack_seq))) { 697 mptcp_drop(sk, skb); 698 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 699 continue; 700 } 701 702 end_seq = MPTCP_SKB_CB(skb)->end_seq; 703 tail = skb_peek_tail(&sk->sk_receive_queue); 704 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 705 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 706 707 /* skip overlapping data, if any */ 708 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n", 709 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 710 delta); 711 MPTCP_SKB_CB(skb)->offset += delta; 712 MPTCP_SKB_CB(skb)->map_seq += delta; 713 __skb_queue_tail(&sk->sk_receive_queue, skb); 714 } 715 msk->bytes_received += end_seq - msk->ack_seq; 716 WRITE_ONCE(msk->ack_seq, end_seq); 717 moved = true; 718 } 719 return moved; 720 } 721 722 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk) 723 { 724 int err = sock_error(ssk); 725 int ssk_state; 726 727 if (!err) 728 return false; 729 730 /* only propagate errors on fallen-back sockets or 731 * on MPC connect 732 */ 733 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk))) 734 return false; 735 736 /* We need to propagate only transition to CLOSE state. 737 * Orphaned socket will see such state change via 738 * subflow_sched_work_if_closed() and that path will properly 739 * destroy the msk as needed. 740 */ 741 ssk_state = inet_sk_state_load(ssk); 742 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD)) 743 mptcp_set_state(sk, ssk_state); 744 WRITE_ONCE(sk->sk_err, -err); 745 746 /* This barrier is coupled with smp_rmb() in mptcp_poll() */ 747 smp_wmb(); 748 sk_error_report(sk); 749 return true; 750 } 751 752 void __mptcp_error_report(struct sock *sk) 753 { 754 struct mptcp_subflow_context *subflow; 755 struct mptcp_sock *msk = mptcp_sk(sk); 756 757 mptcp_for_each_subflow(msk, subflow) 758 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow))) 759 break; 760 } 761 762 /* In most cases we will be able to lock the mptcp socket. If its already 763 * owned, we need to defer to the work queue to avoid ABBA deadlock. 764 */ 765 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 766 { 767 struct sock *sk = (struct sock *)msk; 768 bool moved; 769 770 moved = __mptcp_move_skbs_from_subflow(msk, ssk); 771 __mptcp_ofo_queue(msk); 772 if (unlikely(ssk->sk_err)) { 773 if (!sock_owned_by_user(sk)) 774 __mptcp_error_report(sk); 775 else 776 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags); 777 } 778 779 /* If the moves have caught up with the DATA_FIN sequence number 780 * it's time to ack the DATA_FIN and change socket state, but 781 * this is not a good place to change state. Let the workqueue 782 * do it. 783 */ 784 if (mptcp_pending_data_fin(sk, NULL)) 785 mptcp_schedule_work(sk); 786 return moved; 787 } 788 789 static void __mptcp_rcvbuf_update(struct sock *sk, struct sock *ssk) 790 { 791 if (unlikely(ssk->sk_rcvbuf > sk->sk_rcvbuf)) 792 WRITE_ONCE(sk->sk_rcvbuf, ssk->sk_rcvbuf); 793 } 794 795 static void __mptcp_data_ready(struct sock *sk, struct sock *ssk) 796 { 797 struct mptcp_sock *msk = mptcp_sk(sk); 798 799 __mptcp_rcvbuf_update(sk, ssk); 800 801 /* Wake-up the reader only for in-sequence data */ 802 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk)) 803 sk->sk_data_ready(sk); 804 } 805 806 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 807 { 808 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 809 810 /* The peer can send data while we are shutting down this 811 * subflow at msk destruction time, but we must avoid enqueuing 812 * more data to the msk receive queue 813 */ 814 if (unlikely(subflow->disposable)) 815 return; 816 817 mptcp_data_lock(sk); 818 if (!sock_owned_by_user(sk)) 819 __mptcp_data_ready(sk, ssk); 820 else 821 __set_bit(MPTCP_DEQUEUE, &mptcp_sk(sk)->cb_flags); 822 mptcp_data_unlock(sk); 823 } 824 825 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk) 826 { 827 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq); 828 msk->allow_infinite_fallback = false; 829 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 830 } 831 832 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk) 833 { 834 struct sock *sk = (struct sock *)msk; 835 836 if (sk->sk_state != TCP_ESTABLISHED) 837 return false; 838 839 spin_lock_bh(&msk->fallback_lock); 840 if (!msk->allow_subflows) { 841 spin_unlock_bh(&msk->fallback_lock); 842 return false; 843 } 844 mptcp_subflow_joined(msk, ssk); 845 spin_unlock_bh(&msk->fallback_lock); 846 847 /* attach to msk socket only after we are sure we will deal with it 848 * at close time 849 */ 850 if (sk->sk_socket && !ssk->sk_socket) 851 mptcp_sock_graft(ssk, sk->sk_socket); 852 853 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++; 854 mptcp_sockopt_sync_locked(msk, ssk); 855 mptcp_stop_tout_timer(sk); 856 __mptcp_propagate_sndbuf(sk, ssk); 857 return true; 858 } 859 860 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list) 861 { 862 struct mptcp_subflow_context *tmp, *subflow; 863 struct mptcp_sock *msk = mptcp_sk(sk); 864 865 list_for_each_entry_safe(subflow, tmp, join_list, node) { 866 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 867 bool slow = lock_sock_fast(ssk); 868 869 list_move_tail(&subflow->node, &msk->conn_list); 870 if (!__mptcp_finish_join(msk, ssk)) 871 mptcp_subflow_reset(ssk); 872 unlock_sock_fast(ssk, slow); 873 } 874 } 875 876 static bool mptcp_rtx_timer_pending(struct sock *sk) 877 { 878 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 879 } 880 881 static void mptcp_reset_rtx_timer(struct sock *sk) 882 { 883 struct inet_connection_sock *icsk = inet_csk(sk); 884 unsigned long tout; 885 886 /* prevent rescheduling on close */ 887 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 888 return; 889 890 tout = mptcp_sk(sk)->timer_ival; 891 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 892 } 893 894 bool mptcp_schedule_work(struct sock *sk) 895 { 896 if (inet_sk_state_load(sk) != TCP_CLOSE && 897 schedule_work(&mptcp_sk(sk)->work)) { 898 /* each subflow already holds a reference to the sk, and the 899 * workqueue is invoked by a subflow, so sk can't go away here. 900 */ 901 sock_hold(sk); 902 return true; 903 } 904 return false; 905 } 906 907 static bool mptcp_skb_can_collapse_to(u64 write_seq, 908 const struct sk_buff *skb, 909 const struct mptcp_ext *mpext) 910 { 911 if (!tcp_skb_can_collapse_to(skb)) 912 return false; 913 914 /* can collapse only if MPTCP level sequence is in order and this 915 * mapping has not been xmitted yet 916 */ 917 return mpext && mpext->data_seq + mpext->data_len == write_seq && 918 !mpext->frozen; 919 } 920 921 /* we can append data to the given data frag if: 922 * - there is space available in the backing page_frag 923 * - the data frag tail matches the current page_frag free offset 924 * - the data frag end sequence number matches the current write seq 925 */ 926 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 927 const struct page_frag *pfrag, 928 const struct mptcp_data_frag *df) 929 { 930 return df && pfrag->page == df->page && 931 pfrag->size - pfrag->offset > 0 && 932 pfrag->offset == (df->offset + df->data_len) && 933 df->data_seq + df->data_len == msk->write_seq; 934 } 935 936 static void dfrag_uncharge(struct sock *sk, int len) 937 { 938 sk_mem_uncharge(sk, len); 939 sk_wmem_queued_add(sk, -len); 940 } 941 942 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 943 { 944 int len = dfrag->data_len + dfrag->overhead; 945 946 list_del(&dfrag->list); 947 dfrag_uncharge(sk, len); 948 put_page(dfrag->page); 949 } 950 951 /* called under both the msk socket lock and the data lock */ 952 static void __mptcp_clean_una(struct sock *sk) 953 { 954 struct mptcp_sock *msk = mptcp_sk(sk); 955 struct mptcp_data_frag *dtmp, *dfrag; 956 u64 snd_una; 957 958 snd_una = msk->snd_una; 959 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 960 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 961 break; 962 963 if (unlikely(dfrag == msk->first_pending)) { 964 /* in recovery mode can see ack after the current snd head */ 965 if (WARN_ON_ONCE(!msk->recovery)) 966 break; 967 968 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 969 } 970 971 dfrag_clear(sk, dfrag); 972 } 973 974 dfrag = mptcp_rtx_head(sk); 975 if (dfrag && after64(snd_una, dfrag->data_seq)) { 976 u64 delta = snd_una - dfrag->data_seq; 977 978 /* prevent wrap around in recovery mode */ 979 if (unlikely(delta > dfrag->already_sent)) { 980 if (WARN_ON_ONCE(!msk->recovery)) 981 goto out; 982 if (WARN_ON_ONCE(delta > dfrag->data_len)) 983 goto out; 984 dfrag->already_sent += delta - dfrag->already_sent; 985 } 986 987 dfrag->data_seq += delta; 988 dfrag->offset += delta; 989 dfrag->data_len -= delta; 990 dfrag->already_sent -= delta; 991 992 dfrag_uncharge(sk, delta); 993 } 994 995 /* all retransmitted data acked, recovery completed */ 996 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 997 msk->recovery = false; 998 999 out: 1000 if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) { 1001 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 1002 mptcp_stop_rtx_timer(sk); 1003 } else { 1004 mptcp_reset_rtx_timer(sk); 1005 } 1006 1007 if (mptcp_pending_data_fin_ack(sk)) 1008 mptcp_schedule_work(sk); 1009 } 1010 1011 static void __mptcp_clean_una_wakeup(struct sock *sk) 1012 { 1013 lockdep_assert_held_once(&sk->sk_lock.slock); 1014 1015 __mptcp_clean_una(sk); 1016 mptcp_write_space(sk); 1017 } 1018 1019 static void mptcp_clean_una_wakeup(struct sock *sk) 1020 { 1021 mptcp_data_lock(sk); 1022 __mptcp_clean_una_wakeup(sk); 1023 mptcp_data_unlock(sk); 1024 } 1025 1026 static void mptcp_enter_memory_pressure(struct sock *sk) 1027 { 1028 struct mptcp_subflow_context *subflow; 1029 struct mptcp_sock *msk = mptcp_sk(sk); 1030 bool first = true; 1031 1032 mptcp_for_each_subflow(msk, subflow) { 1033 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1034 1035 if (first) 1036 tcp_enter_memory_pressure(ssk); 1037 sk_stream_moderate_sndbuf(ssk); 1038 1039 first = false; 1040 } 1041 __mptcp_sync_sndbuf(sk); 1042 } 1043 1044 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1045 * data 1046 */ 1047 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1048 { 1049 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1050 pfrag, sk->sk_allocation))) 1051 return true; 1052 1053 mptcp_enter_memory_pressure(sk); 1054 return false; 1055 } 1056 1057 static struct mptcp_data_frag * 1058 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1059 int orig_offset) 1060 { 1061 int offset = ALIGN(orig_offset, sizeof(long)); 1062 struct mptcp_data_frag *dfrag; 1063 1064 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1065 dfrag->data_len = 0; 1066 dfrag->data_seq = msk->write_seq; 1067 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1068 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1069 dfrag->already_sent = 0; 1070 dfrag->page = pfrag->page; 1071 1072 return dfrag; 1073 } 1074 1075 struct mptcp_sendmsg_info { 1076 int mss_now; 1077 int size_goal; 1078 u16 limit; 1079 u16 sent; 1080 unsigned int flags; 1081 bool data_lock_held; 1082 }; 1083 1084 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk, 1085 u64 data_seq, int avail_size) 1086 { 1087 u64 window_end = mptcp_wnd_end(msk); 1088 u64 mptcp_snd_wnd; 1089 1090 if (__mptcp_check_fallback(msk)) 1091 return avail_size; 1092 1093 mptcp_snd_wnd = window_end - data_seq; 1094 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size); 1095 1096 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) { 1097 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd); 1098 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED); 1099 } 1100 1101 return avail_size; 1102 } 1103 1104 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1105 { 1106 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1107 1108 if (!mpext) 1109 return false; 1110 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1111 return true; 1112 } 1113 1114 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1115 { 1116 struct sk_buff *skb; 1117 1118 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1119 if (likely(skb)) { 1120 if (likely(__mptcp_add_ext(skb, gfp))) { 1121 skb_reserve(skb, MAX_TCP_HEADER); 1122 skb->ip_summed = CHECKSUM_PARTIAL; 1123 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1124 return skb; 1125 } 1126 __kfree_skb(skb); 1127 } else { 1128 mptcp_enter_memory_pressure(sk); 1129 } 1130 return NULL; 1131 } 1132 1133 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1134 { 1135 struct sk_buff *skb; 1136 1137 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1138 if (!skb) 1139 return NULL; 1140 1141 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1142 tcp_skb_entail(ssk, skb); 1143 return skb; 1144 } 1145 tcp_skb_tsorted_anchor_cleanup(skb); 1146 kfree_skb(skb); 1147 return NULL; 1148 } 1149 1150 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1151 { 1152 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1153 1154 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1155 } 1156 1157 /* note: this always recompute the csum on the whole skb, even 1158 * if we just appended a single frag. More status info needed 1159 */ 1160 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1161 { 1162 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1163 __wsum csum = ~csum_unfold(mpext->csum); 1164 int offset = skb->len - added; 1165 1166 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1167 } 1168 1169 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1170 struct sock *ssk, 1171 struct mptcp_ext *mpext) 1172 { 1173 if (!mpext) 1174 return; 1175 1176 mpext->infinite_map = 1; 1177 mpext->data_len = 0; 1178 1179 if (!mptcp_try_fallback(ssk, MPTCP_MIB_INFINITEMAPTX)) { 1180 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED); 1181 mptcp_subflow_reset(ssk); 1182 return; 1183 } 1184 1185 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1186 } 1187 1188 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1)) 1189 1190 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1191 struct mptcp_data_frag *dfrag, 1192 struct mptcp_sendmsg_info *info) 1193 { 1194 u64 data_seq = dfrag->data_seq + info->sent; 1195 int offset = dfrag->offset + info->sent; 1196 struct mptcp_sock *msk = mptcp_sk(sk); 1197 bool zero_window_probe = false; 1198 struct mptcp_ext *mpext = NULL; 1199 bool can_coalesce = false; 1200 bool reuse_skb = true; 1201 struct sk_buff *skb; 1202 size_t copy; 1203 int i; 1204 1205 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n", 1206 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1207 1208 if (WARN_ON_ONCE(info->sent > info->limit || 1209 info->limit > dfrag->data_len)) 1210 return 0; 1211 1212 if (unlikely(!__tcp_can_send(ssk))) 1213 return -EAGAIN; 1214 1215 /* compute send limit */ 1216 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE)) 1217 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE; 1218 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1219 copy = info->size_goal; 1220 1221 skb = tcp_write_queue_tail(ssk); 1222 if (skb && copy > skb->len) { 1223 /* Limit the write to the size available in the 1224 * current skb, if any, so that we create at most a new skb. 1225 * Explicitly tells TCP internals to avoid collapsing on later 1226 * queue management operation, to avoid breaking the ext <-> 1227 * SSN association set here 1228 */ 1229 mpext = mptcp_get_ext(skb); 1230 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1231 TCP_SKB_CB(skb)->eor = 1; 1232 tcp_mark_push(tcp_sk(ssk), skb); 1233 goto alloc_skb; 1234 } 1235 1236 i = skb_shinfo(skb)->nr_frags; 1237 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1238 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) { 1239 tcp_mark_push(tcp_sk(ssk), skb); 1240 goto alloc_skb; 1241 } 1242 1243 copy -= skb->len; 1244 } else { 1245 alloc_skb: 1246 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1247 if (!skb) 1248 return -ENOMEM; 1249 1250 i = skb_shinfo(skb)->nr_frags; 1251 reuse_skb = false; 1252 mpext = mptcp_get_ext(skb); 1253 } 1254 1255 /* Zero window and all data acked? Probe. */ 1256 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy); 1257 if (copy == 0) { 1258 u64 snd_una = READ_ONCE(msk->snd_una); 1259 1260 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) { 1261 tcp_remove_empty_skb(ssk); 1262 return 0; 1263 } 1264 1265 zero_window_probe = true; 1266 data_seq = snd_una - 1; 1267 copy = 1; 1268 } 1269 1270 copy = min_t(size_t, copy, info->limit - info->sent); 1271 if (!sk_wmem_schedule(ssk, copy)) { 1272 tcp_remove_empty_skb(ssk); 1273 return -ENOMEM; 1274 } 1275 1276 if (can_coalesce) { 1277 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1278 } else { 1279 get_page(dfrag->page); 1280 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1281 } 1282 1283 skb->len += copy; 1284 skb->data_len += copy; 1285 skb->truesize += copy; 1286 sk_wmem_queued_add(ssk, copy); 1287 sk_mem_charge(ssk, copy); 1288 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1289 TCP_SKB_CB(skb)->end_seq += copy; 1290 tcp_skb_pcount_set(skb, 0); 1291 1292 /* on skb reuse we just need to update the DSS len */ 1293 if (reuse_skb) { 1294 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1295 mpext->data_len += copy; 1296 goto out; 1297 } 1298 1299 memset(mpext, 0, sizeof(*mpext)); 1300 mpext->data_seq = data_seq; 1301 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1302 mpext->data_len = copy; 1303 mpext->use_map = 1; 1304 mpext->dsn64 = 1; 1305 1306 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n", 1307 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1308 mpext->dsn64); 1309 1310 if (zero_window_probe) { 1311 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1312 mpext->frozen = 1; 1313 if (READ_ONCE(msk->csum_enabled)) 1314 mptcp_update_data_checksum(skb, copy); 1315 tcp_push_pending_frames(ssk); 1316 return 0; 1317 } 1318 out: 1319 if (READ_ONCE(msk->csum_enabled)) 1320 mptcp_update_data_checksum(skb, copy); 1321 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1322 mptcp_update_infinite_map(msk, ssk, mpext); 1323 trace_mptcp_sendmsg_frag(mpext); 1324 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1325 return copy; 1326 } 1327 1328 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1329 sizeof(struct tcphdr) - \ 1330 MAX_TCP_OPTION_SPACE - \ 1331 sizeof(struct ipv6hdr) - \ 1332 sizeof(struct frag_hdr)) 1333 1334 struct subflow_send_info { 1335 struct sock *ssk; 1336 u64 linger_time; 1337 }; 1338 1339 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1340 { 1341 if (!subflow->stale) 1342 return; 1343 1344 subflow->stale = 0; 1345 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1346 } 1347 1348 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1349 { 1350 if (unlikely(subflow->stale)) { 1351 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1352 1353 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1354 return false; 1355 1356 mptcp_subflow_set_active(subflow); 1357 } 1358 return __mptcp_subflow_active(subflow); 1359 } 1360 1361 #define SSK_MODE_ACTIVE 0 1362 #define SSK_MODE_BACKUP 1 1363 #define SSK_MODE_MAX 2 1364 1365 /* implement the mptcp packet scheduler; 1366 * returns the subflow that will transmit the next DSS 1367 * additionally updates the rtx timeout 1368 */ 1369 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1370 { 1371 struct subflow_send_info send_info[SSK_MODE_MAX]; 1372 struct mptcp_subflow_context *subflow; 1373 struct sock *sk = (struct sock *)msk; 1374 u32 pace, burst, wmem; 1375 int i, nr_active = 0; 1376 struct sock *ssk; 1377 u64 linger_time; 1378 long tout = 0; 1379 1380 /* pick the subflow with the lower wmem/wspace ratio */ 1381 for (i = 0; i < SSK_MODE_MAX; ++i) { 1382 send_info[i].ssk = NULL; 1383 send_info[i].linger_time = -1; 1384 } 1385 1386 mptcp_for_each_subflow(msk, subflow) { 1387 bool backup = subflow->backup || subflow->request_bkup; 1388 1389 trace_mptcp_subflow_get_send(subflow); 1390 ssk = mptcp_subflow_tcp_sock(subflow); 1391 if (!mptcp_subflow_active(subflow)) 1392 continue; 1393 1394 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1395 nr_active += !backup; 1396 pace = subflow->avg_pacing_rate; 1397 if (unlikely(!pace)) { 1398 /* init pacing rate from socket */ 1399 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1400 pace = subflow->avg_pacing_rate; 1401 if (!pace) 1402 continue; 1403 } 1404 1405 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1406 if (linger_time < send_info[backup].linger_time) { 1407 send_info[backup].ssk = ssk; 1408 send_info[backup].linger_time = linger_time; 1409 } 1410 } 1411 __mptcp_set_timeout(sk, tout); 1412 1413 /* pick the best backup if no other subflow is active */ 1414 if (!nr_active) 1415 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1416 1417 /* According to the blest algorithm, to avoid HoL blocking for the 1418 * faster flow, we need to: 1419 * - estimate the faster flow linger time 1420 * - use the above to estimate the amount of byte transferred 1421 * by the faster flow 1422 * - check that the amount of queued data is greater than the above, 1423 * otherwise do not use the picked, slower, subflow 1424 * We select the subflow with the shorter estimated time to flush 1425 * the queued mem, which basically ensure the above. We just need 1426 * to check that subflow has a non empty cwin. 1427 */ 1428 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1429 if (!ssk || !sk_stream_memory_free(ssk)) 1430 return NULL; 1431 1432 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt); 1433 wmem = READ_ONCE(ssk->sk_wmem_queued); 1434 if (!burst) 1435 return ssk; 1436 1437 subflow = mptcp_subflow_ctx(ssk); 1438 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1439 READ_ONCE(ssk->sk_pacing_rate) * burst, 1440 burst + wmem); 1441 msk->snd_burst = burst; 1442 return ssk; 1443 } 1444 1445 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1446 { 1447 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1448 release_sock(ssk); 1449 } 1450 1451 static void mptcp_update_post_push(struct mptcp_sock *msk, 1452 struct mptcp_data_frag *dfrag, 1453 u32 sent) 1454 { 1455 u64 snd_nxt_new = dfrag->data_seq; 1456 1457 dfrag->already_sent += sent; 1458 1459 msk->snd_burst -= sent; 1460 1461 snd_nxt_new += dfrag->already_sent; 1462 1463 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1464 * is recovering after a failover. In that event, this re-sends 1465 * old segments. 1466 * 1467 * Thus compute snd_nxt_new candidate based on 1468 * the dfrag->data_seq that was sent and the data 1469 * that has been handed to the subflow for transmission 1470 * and skip update in case it was old dfrag. 1471 */ 1472 if (likely(after64(snd_nxt_new, msk->snd_nxt))) { 1473 msk->bytes_sent += snd_nxt_new - msk->snd_nxt; 1474 WRITE_ONCE(msk->snd_nxt, snd_nxt_new); 1475 } 1476 } 1477 1478 void mptcp_check_and_set_pending(struct sock *sk) 1479 { 1480 if (mptcp_send_head(sk)) { 1481 mptcp_data_lock(sk); 1482 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING); 1483 mptcp_data_unlock(sk); 1484 } 1485 } 1486 1487 static int __subflow_push_pending(struct sock *sk, struct sock *ssk, 1488 struct mptcp_sendmsg_info *info) 1489 { 1490 struct mptcp_sock *msk = mptcp_sk(sk); 1491 struct mptcp_data_frag *dfrag; 1492 int len, copied = 0, err = 0; 1493 1494 while ((dfrag = mptcp_send_head(sk))) { 1495 info->sent = dfrag->already_sent; 1496 info->limit = dfrag->data_len; 1497 len = dfrag->data_len - dfrag->already_sent; 1498 while (len > 0) { 1499 int ret = 0; 1500 1501 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info); 1502 if (ret <= 0) { 1503 err = copied ? : ret; 1504 goto out; 1505 } 1506 1507 info->sent += ret; 1508 copied += ret; 1509 len -= ret; 1510 1511 mptcp_update_post_push(msk, dfrag, ret); 1512 } 1513 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1514 1515 if (msk->snd_burst <= 0 || 1516 !sk_stream_memory_free(ssk) || 1517 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) { 1518 err = copied; 1519 goto out; 1520 } 1521 mptcp_set_timeout(sk); 1522 } 1523 err = copied; 1524 1525 out: 1526 if (err > 0) 1527 msk->last_data_sent = tcp_jiffies32; 1528 return err; 1529 } 1530 1531 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1532 { 1533 struct sock *prev_ssk = NULL, *ssk = NULL; 1534 struct mptcp_sock *msk = mptcp_sk(sk); 1535 struct mptcp_sendmsg_info info = { 1536 .flags = flags, 1537 }; 1538 bool do_check_data_fin = false; 1539 int push_count = 1; 1540 1541 while (mptcp_send_head(sk) && (push_count > 0)) { 1542 struct mptcp_subflow_context *subflow; 1543 int ret = 0; 1544 1545 if (mptcp_sched_get_send(msk)) 1546 break; 1547 1548 push_count = 0; 1549 1550 mptcp_for_each_subflow(msk, subflow) { 1551 if (READ_ONCE(subflow->scheduled)) { 1552 mptcp_subflow_set_scheduled(subflow, false); 1553 1554 prev_ssk = ssk; 1555 ssk = mptcp_subflow_tcp_sock(subflow); 1556 if (ssk != prev_ssk) { 1557 /* First check. If the ssk has changed since 1558 * the last round, release prev_ssk 1559 */ 1560 if (prev_ssk) 1561 mptcp_push_release(prev_ssk, &info); 1562 1563 /* Need to lock the new subflow only if different 1564 * from the previous one, otherwise we are still 1565 * helding the relevant lock 1566 */ 1567 lock_sock(ssk); 1568 } 1569 1570 push_count++; 1571 1572 ret = __subflow_push_pending(sk, ssk, &info); 1573 if (ret <= 0) { 1574 if (ret != -EAGAIN || 1575 (1 << ssk->sk_state) & 1576 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE)) 1577 push_count--; 1578 continue; 1579 } 1580 do_check_data_fin = true; 1581 } 1582 } 1583 } 1584 1585 /* at this point we held the socket lock for the last subflow we used */ 1586 if (ssk) 1587 mptcp_push_release(ssk, &info); 1588 1589 /* ensure the rtx timer is running */ 1590 if (!mptcp_rtx_timer_pending(sk)) 1591 mptcp_reset_rtx_timer(sk); 1592 if (do_check_data_fin) 1593 mptcp_check_send_data_fin(sk); 1594 } 1595 1596 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first) 1597 { 1598 struct mptcp_sock *msk = mptcp_sk(sk); 1599 struct mptcp_sendmsg_info info = { 1600 .data_lock_held = true, 1601 }; 1602 bool keep_pushing = true; 1603 struct sock *xmit_ssk; 1604 int copied = 0; 1605 1606 info.flags = 0; 1607 while (mptcp_send_head(sk) && keep_pushing) { 1608 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1609 int ret = 0; 1610 1611 /* check for a different subflow usage only after 1612 * spooling the first chunk of data 1613 */ 1614 if (first) { 1615 mptcp_subflow_set_scheduled(subflow, false); 1616 ret = __subflow_push_pending(sk, ssk, &info); 1617 first = false; 1618 if (ret <= 0) 1619 break; 1620 copied += ret; 1621 continue; 1622 } 1623 1624 if (mptcp_sched_get_send(msk)) 1625 goto out; 1626 1627 if (READ_ONCE(subflow->scheduled)) { 1628 mptcp_subflow_set_scheduled(subflow, false); 1629 ret = __subflow_push_pending(sk, ssk, &info); 1630 if (ret <= 0) 1631 keep_pushing = false; 1632 copied += ret; 1633 } 1634 1635 mptcp_for_each_subflow(msk, subflow) { 1636 if (READ_ONCE(subflow->scheduled)) { 1637 xmit_ssk = mptcp_subflow_tcp_sock(subflow); 1638 if (xmit_ssk != ssk) { 1639 mptcp_subflow_delegate(subflow, 1640 MPTCP_DELEGATE_SEND); 1641 keep_pushing = false; 1642 } 1643 } 1644 } 1645 } 1646 1647 out: 1648 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1649 * not going to flush it via release_sock() 1650 */ 1651 if (copied) { 1652 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1653 info.size_goal); 1654 if (!mptcp_rtx_timer_pending(sk)) 1655 mptcp_reset_rtx_timer(sk); 1656 1657 if (msk->snd_data_fin_enable && 1658 msk->snd_nxt + 1 == msk->write_seq) 1659 mptcp_schedule_work(sk); 1660 } 1661 } 1662 1663 static int mptcp_disconnect(struct sock *sk, int flags); 1664 1665 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1666 size_t len, int *copied_syn) 1667 { 1668 unsigned int saved_flags = msg->msg_flags; 1669 struct mptcp_sock *msk = mptcp_sk(sk); 1670 struct sock *ssk; 1671 int ret; 1672 1673 /* on flags based fastopen the mptcp is supposed to create the 1674 * first subflow right now. Otherwise we are in the defer_connect 1675 * path, and the first subflow must be already present. 1676 * Since the defer_connect flag is cleared after the first succsful 1677 * fastopen attempt, no need to check for additional subflow status. 1678 */ 1679 if (msg->msg_flags & MSG_FASTOPEN) { 1680 ssk = __mptcp_nmpc_sk(msk); 1681 if (IS_ERR(ssk)) 1682 return PTR_ERR(ssk); 1683 } 1684 if (!msk->first) 1685 return -EINVAL; 1686 1687 ssk = msk->first; 1688 1689 lock_sock(ssk); 1690 msg->msg_flags |= MSG_DONTWAIT; 1691 msk->fastopening = 1; 1692 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL); 1693 msk->fastopening = 0; 1694 msg->msg_flags = saved_flags; 1695 release_sock(ssk); 1696 1697 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */ 1698 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) { 1699 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1700 msg->msg_namelen, msg->msg_flags, 1); 1701 1702 /* Keep the same behaviour of plain TCP: zero the copied bytes in 1703 * case of any error, except timeout or signal 1704 */ 1705 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR) 1706 *copied_syn = 0; 1707 } else if (ret && ret != -EINPROGRESS) { 1708 /* The disconnect() op called by tcp_sendmsg_fastopen()/ 1709 * __inet_stream_connect() can fail, due to looking check, 1710 * see mptcp_disconnect(). 1711 * Attempt it again outside the problematic scope. 1712 */ 1713 if (!mptcp_disconnect(sk, 0)) { 1714 sk->sk_disconnects++; 1715 sk->sk_socket->state = SS_UNCONNECTED; 1716 } 1717 } 1718 inet_clear_bit(DEFER_CONNECT, sk); 1719 1720 return ret; 1721 } 1722 1723 static int do_copy_data_nocache(struct sock *sk, int copy, 1724 struct iov_iter *from, char *to) 1725 { 1726 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) { 1727 if (!copy_from_iter_full_nocache(to, copy, from)) 1728 return -EFAULT; 1729 } else if (!copy_from_iter_full(to, copy, from)) { 1730 return -EFAULT; 1731 } 1732 return 0; 1733 } 1734 1735 /* open-code sk_stream_memory_free() plus sent limit computation to 1736 * avoid indirect calls in fast-path. 1737 * Called under the msk socket lock, so we can avoid a bunch of ONCE 1738 * annotations. 1739 */ 1740 static u32 mptcp_send_limit(const struct sock *sk) 1741 { 1742 const struct mptcp_sock *msk = mptcp_sk(sk); 1743 u32 limit, not_sent; 1744 1745 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf)) 1746 return 0; 1747 1748 limit = mptcp_notsent_lowat(sk); 1749 if (limit == UINT_MAX) 1750 return UINT_MAX; 1751 1752 not_sent = msk->write_seq - msk->snd_nxt; 1753 if (not_sent >= limit) 1754 return 0; 1755 1756 return limit - not_sent; 1757 } 1758 1759 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1760 { 1761 struct mptcp_sock *msk = mptcp_sk(sk); 1762 struct page_frag *pfrag; 1763 size_t copied = 0; 1764 int ret = 0; 1765 long timeo; 1766 1767 /* silently ignore everything else */ 1768 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN; 1769 1770 lock_sock(sk); 1771 1772 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) || 1773 msg->msg_flags & MSG_FASTOPEN)) { 1774 int copied_syn = 0; 1775 1776 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn); 1777 copied += copied_syn; 1778 if (ret == -EINPROGRESS && copied_syn > 0) 1779 goto out; 1780 else if (ret) 1781 goto do_error; 1782 } 1783 1784 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1785 1786 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1787 ret = sk_stream_wait_connect(sk, &timeo); 1788 if (ret) 1789 goto do_error; 1790 } 1791 1792 ret = -EPIPE; 1793 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))) 1794 goto do_error; 1795 1796 pfrag = sk_page_frag(sk); 1797 1798 while (msg_data_left(msg)) { 1799 int total_ts, frag_truesize = 0; 1800 struct mptcp_data_frag *dfrag; 1801 bool dfrag_collapsed; 1802 size_t psize, offset; 1803 u32 copy_limit; 1804 1805 /* ensure fitting the notsent_lowat() constraint */ 1806 copy_limit = mptcp_send_limit(sk); 1807 if (!copy_limit) 1808 goto wait_for_memory; 1809 1810 /* reuse tail pfrag, if possible, or carve a new one from the 1811 * page allocator 1812 */ 1813 dfrag = mptcp_pending_tail(sk); 1814 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1815 if (!dfrag_collapsed) { 1816 if (!mptcp_page_frag_refill(sk, pfrag)) 1817 goto wait_for_memory; 1818 1819 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1820 frag_truesize = dfrag->overhead; 1821 } 1822 1823 /* we do not bound vs wspace, to allow a single packet. 1824 * memory accounting will prevent execessive memory usage 1825 * anyway 1826 */ 1827 offset = dfrag->offset + dfrag->data_len; 1828 psize = pfrag->size - offset; 1829 psize = min_t(size_t, psize, msg_data_left(msg)); 1830 psize = min_t(size_t, psize, copy_limit); 1831 total_ts = psize + frag_truesize; 1832 1833 if (!sk_wmem_schedule(sk, total_ts)) 1834 goto wait_for_memory; 1835 1836 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter, 1837 page_address(dfrag->page) + offset); 1838 if (ret) 1839 goto do_error; 1840 1841 /* data successfully copied into the write queue */ 1842 sk_forward_alloc_add(sk, -total_ts); 1843 copied += psize; 1844 dfrag->data_len += psize; 1845 frag_truesize += psize; 1846 pfrag->offset += frag_truesize; 1847 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1848 1849 /* charge data on mptcp pending queue to the msk socket 1850 * Note: we charge such data both to sk and ssk 1851 */ 1852 sk_wmem_queued_add(sk, frag_truesize); 1853 if (!dfrag_collapsed) { 1854 get_page(dfrag->page); 1855 list_add_tail(&dfrag->list, &msk->rtx_queue); 1856 if (!msk->first_pending) 1857 WRITE_ONCE(msk->first_pending, dfrag); 1858 } 1859 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk, 1860 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1861 !dfrag_collapsed); 1862 1863 continue; 1864 1865 wait_for_memory: 1866 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1867 __mptcp_push_pending(sk, msg->msg_flags); 1868 ret = sk_stream_wait_memory(sk, &timeo); 1869 if (ret) 1870 goto do_error; 1871 } 1872 1873 if (copied) 1874 __mptcp_push_pending(sk, msg->msg_flags); 1875 1876 out: 1877 release_sock(sk); 1878 return copied; 1879 1880 do_error: 1881 if (copied) 1882 goto out; 1883 1884 copied = sk_stream_error(sk, msg->msg_flags, ret); 1885 goto out; 1886 } 1887 1888 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied); 1889 1890 static int __mptcp_recvmsg_mskq(struct sock *sk, 1891 struct msghdr *msg, 1892 size_t len, int flags, 1893 struct scm_timestamping_internal *tss, 1894 int *cmsg_flags) 1895 { 1896 struct mptcp_sock *msk = mptcp_sk(sk); 1897 struct sk_buff *skb, *tmp; 1898 int copied = 0; 1899 1900 skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) { 1901 u32 offset = MPTCP_SKB_CB(skb)->offset; 1902 u32 data_len = skb->len - offset; 1903 u32 count = min_t(size_t, len - copied, data_len); 1904 int err; 1905 1906 if (!(flags & MSG_TRUNC)) { 1907 err = skb_copy_datagram_msg(skb, offset, msg, count); 1908 if (unlikely(err < 0)) { 1909 if (!copied) 1910 return err; 1911 break; 1912 } 1913 } 1914 1915 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 1916 tcp_update_recv_tstamps(skb, tss); 1917 *cmsg_flags |= MPTCP_CMSG_TS; 1918 } 1919 1920 copied += count; 1921 1922 if (count < data_len) { 1923 if (!(flags & MSG_PEEK)) { 1924 MPTCP_SKB_CB(skb)->offset += count; 1925 MPTCP_SKB_CB(skb)->map_seq += count; 1926 msk->bytes_consumed += count; 1927 } 1928 break; 1929 } 1930 1931 if (!(flags & MSG_PEEK)) { 1932 /* avoid the indirect call, we know the destructor is sock_wfree */ 1933 skb->destructor = NULL; 1934 atomic_sub(skb->truesize, &sk->sk_rmem_alloc); 1935 sk_mem_uncharge(sk, skb->truesize); 1936 __skb_unlink(skb, &sk->sk_receive_queue); 1937 __kfree_skb(skb); 1938 msk->bytes_consumed += count; 1939 } 1940 1941 if (copied >= len) 1942 break; 1943 } 1944 1945 mptcp_rcv_space_adjust(msk, copied); 1946 return copied; 1947 } 1948 1949 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1950 * 1951 * Only difference: Use highest rtt estimate of the subflows in use. 1952 */ 1953 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1954 { 1955 struct mptcp_subflow_context *subflow; 1956 struct sock *sk = (struct sock *)msk; 1957 u8 scaling_ratio = U8_MAX; 1958 u32 time, advmss = 1; 1959 u64 rtt_us, mstamp; 1960 1961 msk_owned_by_me(msk); 1962 1963 if (copied <= 0) 1964 return; 1965 1966 if (!msk->rcvspace_init) 1967 mptcp_rcv_space_init(msk, msk->first); 1968 1969 msk->rcvq_space.copied += copied; 1970 1971 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1972 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1973 1974 rtt_us = msk->rcvq_space.rtt_us; 1975 if (rtt_us && time < (rtt_us >> 3)) 1976 return; 1977 1978 rtt_us = 0; 1979 mptcp_for_each_subflow(msk, subflow) { 1980 const struct tcp_sock *tp; 1981 u64 sf_rtt_us; 1982 u32 sf_advmss; 1983 1984 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1985 1986 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1987 sf_advmss = READ_ONCE(tp->advmss); 1988 1989 rtt_us = max(sf_rtt_us, rtt_us); 1990 advmss = max(sf_advmss, advmss); 1991 scaling_ratio = min(tp->scaling_ratio, scaling_ratio); 1992 } 1993 1994 msk->rcvq_space.rtt_us = rtt_us; 1995 msk->scaling_ratio = scaling_ratio; 1996 if (time < (rtt_us >> 3) || rtt_us == 0) 1997 return; 1998 1999 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 2000 goto new_measure; 2001 2002 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 2003 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 2004 u64 rcvwin, grow; 2005 int rcvbuf; 2006 2007 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 2008 2009 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 2010 2011 do_div(grow, msk->rcvq_space.space); 2012 rcvwin += (grow << 1); 2013 2014 rcvbuf = min_t(u64, mptcp_space_from_win(sk, rcvwin), 2015 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2016 2017 if (rcvbuf > sk->sk_rcvbuf) { 2018 u32 window_clamp; 2019 2020 window_clamp = mptcp_win_from_space(sk, rcvbuf); 2021 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 2022 2023 /* Make subflows follow along. If we do not do this, we 2024 * get drops at subflow level if skbs can't be moved to 2025 * the mptcp rx queue fast enough (announced rcv_win can 2026 * exceed ssk->sk_rcvbuf). 2027 */ 2028 mptcp_for_each_subflow(msk, subflow) { 2029 struct sock *ssk; 2030 bool slow; 2031 2032 ssk = mptcp_subflow_tcp_sock(subflow); 2033 slow = lock_sock_fast(ssk); 2034 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 2035 WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp); 2036 if (tcp_can_send_ack(ssk)) 2037 tcp_cleanup_rbuf(ssk, 1); 2038 unlock_sock_fast(ssk, slow); 2039 } 2040 } 2041 } 2042 2043 msk->rcvq_space.space = msk->rcvq_space.copied; 2044 new_measure: 2045 msk->rcvq_space.copied = 0; 2046 msk->rcvq_space.time = mstamp; 2047 } 2048 2049 static struct mptcp_subflow_context * 2050 __mptcp_first_ready_from(struct mptcp_sock *msk, 2051 struct mptcp_subflow_context *subflow) 2052 { 2053 struct mptcp_subflow_context *start_subflow = subflow; 2054 2055 while (!READ_ONCE(subflow->data_avail)) { 2056 subflow = mptcp_next_subflow(msk, subflow); 2057 if (subflow == start_subflow) 2058 return NULL; 2059 } 2060 return subflow; 2061 } 2062 2063 static bool __mptcp_move_skbs(struct sock *sk) 2064 { 2065 struct mptcp_subflow_context *subflow; 2066 struct mptcp_sock *msk = mptcp_sk(sk); 2067 bool ret = false; 2068 2069 if (list_empty(&msk->conn_list)) 2070 return false; 2071 2072 /* verify we can move any data from the subflow, eventually updating */ 2073 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 2074 mptcp_for_each_subflow(msk, subflow) 2075 __mptcp_rcvbuf_update(sk, subflow->tcp_sock); 2076 2077 subflow = list_first_entry(&msk->conn_list, 2078 struct mptcp_subflow_context, node); 2079 for (;;) { 2080 struct sock *ssk; 2081 bool slowpath; 2082 2083 /* 2084 * As an optimization avoid traversing the subflows list 2085 * and ev. acquiring the subflow socket lock before baling out 2086 */ 2087 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 2088 break; 2089 2090 subflow = __mptcp_first_ready_from(msk, subflow); 2091 if (!subflow) 2092 break; 2093 2094 ssk = mptcp_subflow_tcp_sock(subflow); 2095 slowpath = lock_sock_fast(ssk); 2096 ret = __mptcp_move_skbs_from_subflow(msk, ssk) || ret; 2097 if (unlikely(ssk->sk_err)) 2098 __mptcp_error_report(sk); 2099 unlock_sock_fast(ssk, slowpath); 2100 2101 subflow = mptcp_next_subflow(msk, subflow); 2102 } 2103 2104 __mptcp_ofo_queue(msk); 2105 if (ret) 2106 mptcp_check_data_fin((struct sock *)msk); 2107 return ret; 2108 } 2109 2110 static unsigned int mptcp_inq_hint(const struct sock *sk) 2111 { 2112 const struct mptcp_sock *msk = mptcp_sk(sk); 2113 const struct sk_buff *skb; 2114 2115 skb = skb_peek(&sk->sk_receive_queue); 2116 if (skb) { 2117 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq; 2118 2119 if (hint_val >= INT_MAX) 2120 return INT_MAX; 2121 2122 return (unsigned int)hint_val; 2123 } 2124 2125 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2126 return 1; 2127 2128 return 0; 2129 } 2130 2131 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2132 int flags, int *addr_len) 2133 { 2134 struct mptcp_sock *msk = mptcp_sk(sk); 2135 struct scm_timestamping_internal tss; 2136 int copied = 0, cmsg_flags = 0; 2137 int target; 2138 long timeo; 2139 2140 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2141 if (unlikely(flags & MSG_ERRQUEUE)) 2142 return inet_recv_error(sk, msg, len, addr_len); 2143 2144 lock_sock(sk); 2145 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2146 copied = -ENOTCONN; 2147 goto out_err; 2148 } 2149 2150 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2151 2152 len = min_t(size_t, len, INT_MAX); 2153 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2154 2155 if (unlikely(msk->recvmsg_inq)) 2156 cmsg_flags = MPTCP_CMSG_INQ; 2157 2158 while (copied < len) { 2159 int err, bytes_read; 2160 2161 bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags, &tss, &cmsg_flags); 2162 if (unlikely(bytes_read < 0)) { 2163 if (!copied) 2164 copied = bytes_read; 2165 goto out_err; 2166 } 2167 2168 copied += bytes_read; 2169 2170 if (skb_queue_empty(&sk->sk_receive_queue) && __mptcp_move_skbs(sk)) 2171 continue; 2172 2173 /* only the MPTCP socket status is relevant here. The exit 2174 * conditions mirror closely tcp_recvmsg() 2175 */ 2176 if (copied >= target) 2177 break; 2178 2179 if (copied) { 2180 if (sk->sk_err || 2181 sk->sk_state == TCP_CLOSE || 2182 (sk->sk_shutdown & RCV_SHUTDOWN) || 2183 !timeo || 2184 signal_pending(current)) 2185 break; 2186 } else { 2187 if (sk->sk_err) { 2188 copied = sock_error(sk); 2189 break; 2190 } 2191 2192 if (sk->sk_shutdown & RCV_SHUTDOWN) { 2193 /* race breaker: the shutdown could be after the 2194 * previous receive queue check 2195 */ 2196 if (__mptcp_move_skbs(sk)) 2197 continue; 2198 break; 2199 } 2200 2201 if (sk->sk_state == TCP_CLOSE) { 2202 copied = -ENOTCONN; 2203 break; 2204 } 2205 2206 if (!timeo) { 2207 copied = -EAGAIN; 2208 break; 2209 } 2210 2211 if (signal_pending(current)) { 2212 copied = sock_intr_errno(timeo); 2213 break; 2214 } 2215 } 2216 2217 pr_debug("block timeout %ld\n", timeo); 2218 mptcp_cleanup_rbuf(msk, copied); 2219 err = sk_wait_data(sk, &timeo, NULL); 2220 if (err < 0) { 2221 err = copied ? : err; 2222 goto out_err; 2223 } 2224 } 2225 2226 mptcp_cleanup_rbuf(msk, copied); 2227 2228 out_err: 2229 if (cmsg_flags && copied >= 0) { 2230 if (cmsg_flags & MPTCP_CMSG_TS) 2231 tcp_recv_timestamp(msg, sk, &tss); 2232 2233 if (cmsg_flags & MPTCP_CMSG_INQ) { 2234 unsigned int inq = mptcp_inq_hint(sk); 2235 2236 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2237 } 2238 } 2239 2240 pr_debug("msk=%p rx queue empty=%d copied=%d\n", 2241 msk, skb_queue_empty(&sk->sk_receive_queue), copied); 2242 2243 release_sock(sk); 2244 return copied; 2245 } 2246 2247 static void mptcp_retransmit_timer(struct timer_list *t) 2248 { 2249 struct inet_connection_sock *icsk = timer_container_of(icsk, t, 2250 icsk_retransmit_timer); 2251 struct sock *sk = &icsk->icsk_inet.sk; 2252 struct mptcp_sock *msk = mptcp_sk(sk); 2253 2254 bh_lock_sock(sk); 2255 if (!sock_owned_by_user(sk)) { 2256 /* we need a process context to retransmit */ 2257 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2258 mptcp_schedule_work(sk); 2259 } else { 2260 /* delegate our work to tcp_release_cb() */ 2261 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2262 } 2263 bh_unlock_sock(sk); 2264 sock_put(sk); 2265 } 2266 2267 static void mptcp_tout_timer(struct timer_list *t) 2268 { 2269 struct sock *sk = timer_container_of(sk, t, sk_timer); 2270 2271 mptcp_schedule_work(sk); 2272 sock_put(sk); 2273 } 2274 2275 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2276 * level. 2277 * 2278 * A backup subflow is returned only if that is the only kind available. 2279 */ 2280 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2281 { 2282 struct sock *backup = NULL, *pick = NULL; 2283 struct mptcp_subflow_context *subflow; 2284 int min_stale_count = INT_MAX; 2285 2286 mptcp_for_each_subflow(msk, subflow) { 2287 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2288 2289 if (!__mptcp_subflow_active(subflow)) 2290 continue; 2291 2292 /* still data outstanding at TCP level? skip this */ 2293 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2294 mptcp_pm_subflow_chk_stale(msk, ssk); 2295 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2296 continue; 2297 } 2298 2299 if (subflow->backup || subflow->request_bkup) { 2300 if (!backup) 2301 backup = ssk; 2302 continue; 2303 } 2304 2305 if (!pick) 2306 pick = ssk; 2307 } 2308 2309 if (pick) 2310 return pick; 2311 2312 /* use backup only if there are no progresses anywhere */ 2313 return min_stale_count > 1 ? backup : NULL; 2314 } 2315 2316 bool __mptcp_retransmit_pending_data(struct sock *sk) 2317 { 2318 struct mptcp_data_frag *cur, *rtx_head; 2319 struct mptcp_sock *msk = mptcp_sk(sk); 2320 2321 if (__mptcp_check_fallback(msk)) 2322 return false; 2323 2324 /* the closing socket has some data untransmitted and/or unacked: 2325 * some data in the mptcp rtx queue has not really xmitted yet. 2326 * keep it simple and re-inject the whole mptcp level rtx queue 2327 */ 2328 mptcp_data_lock(sk); 2329 __mptcp_clean_una_wakeup(sk); 2330 rtx_head = mptcp_rtx_head(sk); 2331 if (!rtx_head) { 2332 mptcp_data_unlock(sk); 2333 return false; 2334 } 2335 2336 msk->recovery_snd_nxt = msk->snd_nxt; 2337 msk->recovery = true; 2338 mptcp_data_unlock(sk); 2339 2340 msk->first_pending = rtx_head; 2341 msk->snd_burst = 0; 2342 2343 /* be sure to clear the "sent status" on all re-injected fragments */ 2344 list_for_each_entry(cur, &msk->rtx_queue, list) { 2345 if (!cur->already_sent) 2346 break; 2347 cur->already_sent = 0; 2348 } 2349 2350 return true; 2351 } 2352 2353 /* flags for __mptcp_close_ssk() */ 2354 #define MPTCP_CF_PUSH BIT(1) 2355 #define MPTCP_CF_FASTCLOSE BIT(2) 2356 2357 /* be sure to send a reset only if the caller asked for it, also 2358 * clean completely the subflow status when the subflow reaches 2359 * TCP_CLOSE state 2360 */ 2361 static void __mptcp_subflow_disconnect(struct sock *ssk, 2362 struct mptcp_subflow_context *subflow, 2363 unsigned int flags) 2364 { 2365 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 2366 (flags & MPTCP_CF_FASTCLOSE)) { 2367 /* The MPTCP code never wait on the subflow sockets, TCP-level 2368 * disconnect should never fail 2369 */ 2370 WARN_ON_ONCE(tcp_disconnect(ssk, 0)); 2371 mptcp_subflow_ctx_reset(subflow); 2372 } else { 2373 tcp_shutdown(ssk, SEND_SHUTDOWN); 2374 } 2375 } 2376 2377 /* subflow sockets can be either outgoing (connect) or incoming 2378 * (accept). 2379 * 2380 * Outgoing subflows use in-kernel sockets. 2381 * Incoming subflows do not have their own 'struct socket' allocated, 2382 * so we need to use tcp_close() after detaching them from the mptcp 2383 * parent socket. 2384 */ 2385 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2386 struct mptcp_subflow_context *subflow, 2387 unsigned int flags) 2388 { 2389 struct mptcp_sock *msk = mptcp_sk(sk); 2390 bool dispose_it, need_push = false; 2391 2392 /* If the first subflow moved to a close state before accept, e.g. due 2393 * to an incoming reset or listener shutdown, the subflow socket is 2394 * already deleted by inet_child_forget() and the mptcp socket can't 2395 * survive too. 2396 */ 2397 if (msk->in_accept_queue && msk->first == ssk && 2398 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) { 2399 /* ensure later check in mptcp_worker() will dispose the msk */ 2400 sock_set_flag(sk, SOCK_DEAD); 2401 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1)); 2402 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2403 mptcp_subflow_drop_ctx(ssk); 2404 goto out_release; 2405 } 2406 2407 dispose_it = msk->free_first || ssk != msk->first; 2408 if (dispose_it) 2409 list_del(&subflow->node); 2410 2411 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2412 2413 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) { 2414 /* be sure to force the tcp_close path 2415 * to generate the egress reset 2416 */ 2417 ssk->sk_lingertime = 0; 2418 sock_set_flag(ssk, SOCK_LINGER); 2419 subflow->send_fastclose = 1; 2420 } 2421 2422 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2423 if (!dispose_it) { 2424 __mptcp_subflow_disconnect(ssk, subflow, flags); 2425 release_sock(ssk); 2426 2427 goto out; 2428 } 2429 2430 subflow->disposable = 1; 2431 2432 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2433 * the ssk has been already destroyed, we just need to release the 2434 * reference owned by msk; 2435 */ 2436 if (!inet_csk(ssk)->icsk_ulp_ops) { 2437 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD)); 2438 kfree_rcu(subflow, rcu); 2439 } else { 2440 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2441 __tcp_close(ssk, 0); 2442 2443 /* close acquired an extra ref */ 2444 __sock_put(ssk); 2445 } 2446 2447 out_release: 2448 __mptcp_subflow_error_report(sk, ssk); 2449 release_sock(ssk); 2450 2451 sock_put(ssk); 2452 2453 if (ssk == msk->first) 2454 WRITE_ONCE(msk->first, NULL); 2455 2456 out: 2457 __mptcp_sync_sndbuf(sk); 2458 if (need_push) 2459 __mptcp_push_pending(sk, 0); 2460 2461 /* Catch every 'all subflows closed' scenario, including peers silently 2462 * closing them, e.g. due to timeout. 2463 * For established sockets, allow an additional timeout before closing, 2464 * as the protocol can still create more subflows. 2465 */ 2466 if (list_is_singular(&msk->conn_list) && msk->first && 2467 inet_sk_state_load(msk->first) == TCP_CLOSE) { 2468 if (sk->sk_state != TCP_ESTABLISHED || 2469 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) { 2470 mptcp_set_state(sk, TCP_CLOSE); 2471 mptcp_close_wake_up(sk); 2472 } else { 2473 mptcp_start_tout_timer(sk); 2474 } 2475 } 2476 } 2477 2478 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2479 struct mptcp_subflow_context *subflow) 2480 { 2481 /* The first subflow can already be closed and still in the list */ 2482 if (subflow->close_event_done) 2483 return; 2484 2485 subflow->close_event_done = true; 2486 2487 if (sk->sk_state == TCP_ESTABLISHED) 2488 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2489 2490 /* subflow aborted before reaching the fully_established status 2491 * attempt the creation of the next subflow 2492 */ 2493 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow); 2494 2495 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2496 } 2497 2498 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2499 { 2500 return 0; 2501 } 2502 2503 static void __mptcp_close_subflow(struct sock *sk) 2504 { 2505 struct mptcp_subflow_context *subflow, *tmp; 2506 struct mptcp_sock *msk = mptcp_sk(sk); 2507 2508 might_sleep(); 2509 2510 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2511 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2512 int ssk_state = inet_sk_state_load(ssk); 2513 2514 if (ssk_state != TCP_CLOSE && 2515 (ssk_state != TCP_CLOSE_WAIT || 2516 inet_sk_state_load(sk) != TCP_ESTABLISHED)) 2517 continue; 2518 2519 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2520 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2521 continue; 2522 2523 mptcp_close_ssk(sk, ssk, subflow); 2524 } 2525 2526 } 2527 2528 static bool mptcp_close_tout_expired(const struct sock *sk) 2529 { 2530 if (!inet_csk(sk)->icsk_mtup.probe_timestamp || 2531 sk->sk_state == TCP_CLOSE) 2532 return false; 2533 2534 return time_after32(tcp_jiffies32, 2535 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk)); 2536 } 2537 2538 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2539 { 2540 struct mptcp_subflow_context *subflow, *tmp; 2541 struct sock *sk = (struct sock *)msk; 2542 2543 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2544 return; 2545 2546 mptcp_token_destroy(msk); 2547 2548 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2549 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2550 bool slow; 2551 2552 slow = lock_sock_fast(tcp_sk); 2553 if (tcp_sk->sk_state != TCP_CLOSE) { 2554 mptcp_send_active_reset_reason(tcp_sk); 2555 tcp_set_state(tcp_sk, TCP_CLOSE); 2556 } 2557 unlock_sock_fast(tcp_sk, slow); 2558 } 2559 2560 /* Mirror the tcp_reset() error propagation */ 2561 switch (sk->sk_state) { 2562 case TCP_SYN_SENT: 2563 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 2564 break; 2565 case TCP_CLOSE_WAIT: 2566 WRITE_ONCE(sk->sk_err, EPIPE); 2567 break; 2568 case TCP_CLOSE: 2569 return; 2570 default: 2571 WRITE_ONCE(sk->sk_err, ECONNRESET); 2572 } 2573 2574 mptcp_set_state(sk, TCP_CLOSE); 2575 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2576 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2577 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2578 2579 /* the calling mptcp_worker will properly destroy the socket */ 2580 if (sock_flag(sk, SOCK_DEAD)) 2581 return; 2582 2583 sk->sk_state_change(sk); 2584 sk_error_report(sk); 2585 } 2586 2587 static void __mptcp_retrans(struct sock *sk) 2588 { 2589 struct mptcp_sendmsg_info info = { .data_lock_held = true, }; 2590 struct mptcp_sock *msk = mptcp_sk(sk); 2591 struct mptcp_subflow_context *subflow; 2592 struct mptcp_data_frag *dfrag; 2593 struct sock *ssk; 2594 int ret, err; 2595 u16 len = 0; 2596 2597 mptcp_clean_una_wakeup(sk); 2598 2599 /* first check ssk: need to kick "stale" logic */ 2600 err = mptcp_sched_get_retrans(msk); 2601 dfrag = mptcp_rtx_head(sk); 2602 if (!dfrag) { 2603 if (mptcp_data_fin_enabled(msk)) { 2604 struct inet_connection_sock *icsk = inet_csk(sk); 2605 2606 icsk->icsk_retransmits++; 2607 mptcp_set_datafin_timeout(sk); 2608 mptcp_send_ack(msk); 2609 2610 goto reset_timer; 2611 } 2612 2613 if (!mptcp_send_head(sk)) 2614 return; 2615 2616 goto reset_timer; 2617 } 2618 2619 if (err) 2620 goto reset_timer; 2621 2622 mptcp_for_each_subflow(msk, subflow) { 2623 if (READ_ONCE(subflow->scheduled)) { 2624 u16 copied = 0; 2625 2626 mptcp_subflow_set_scheduled(subflow, false); 2627 2628 ssk = mptcp_subflow_tcp_sock(subflow); 2629 2630 lock_sock(ssk); 2631 2632 /* limit retransmission to the bytes already sent on some subflows */ 2633 info.sent = 0; 2634 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : 2635 dfrag->already_sent; 2636 2637 /* 2638 * make the whole retrans decision, xmit, disallow 2639 * fallback atomic 2640 */ 2641 spin_lock_bh(&msk->fallback_lock); 2642 if (__mptcp_check_fallback(msk)) { 2643 spin_unlock_bh(&msk->fallback_lock); 2644 release_sock(ssk); 2645 return; 2646 } 2647 2648 while (info.sent < info.limit) { 2649 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2650 if (ret <= 0) 2651 break; 2652 2653 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2654 copied += ret; 2655 info.sent += ret; 2656 } 2657 if (copied) { 2658 len = max(copied, len); 2659 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2660 info.size_goal); 2661 msk->allow_infinite_fallback = false; 2662 } 2663 spin_unlock_bh(&msk->fallback_lock); 2664 2665 release_sock(ssk); 2666 } 2667 } 2668 2669 msk->bytes_retrans += len; 2670 dfrag->already_sent = max(dfrag->already_sent, len); 2671 2672 reset_timer: 2673 mptcp_check_and_set_pending(sk); 2674 2675 if (!mptcp_rtx_timer_pending(sk)) 2676 mptcp_reset_rtx_timer(sk); 2677 } 2678 2679 /* schedule the timeout timer for the relevant event: either close timeout 2680 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one 2681 */ 2682 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout) 2683 { 2684 struct sock *sk = (struct sock *)msk; 2685 unsigned long timeout, close_timeout; 2686 2687 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp) 2688 return; 2689 2690 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp - 2691 tcp_jiffies32 + jiffies + mptcp_close_timeout(sk); 2692 2693 /* the close timeout takes precedence on the fail one, and here at least one of 2694 * them is active 2695 */ 2696 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout; 2697 2698 sk_reset_timer(sk, &sk->sk_timer, timeout); 2699 } 2700 2701 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2702 { 2703 struct sock *ssk = msk->first; 2704 bool slow; 2705 2706 if (!ssk) 2707 return; 2708 2709 pr_debug("MP_FAIL doesn't respond, reset the subflow\n"); 2710 2711 slow = lock_sock_fast(ssk); 2712 mptcp_subflow_reset(ssk); 2713 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0); 2714 unlock_sock_fast(ssk, slow); 2715 } 2716 2717 static void mptcp_do_fastclose(struct sock *sk) 2718 { 2719 struct mptcp_subflow_context *subflow, *tmp; 2720 struct mptcp_sock *msk = mptcp_sk(sk); 2721 2722 mptcp_set_state(sk, TCP_CLOSE); 2723 mptcp_for_each_subflow_safe(msk, subflow, tmp) 2724 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), 2725 subflow, MPTCP_CF_FASTCLOSE); 2726 } 2727 2728 static void mptcp_worker(struct work_struct *work) 2729 { 2730 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2731 struct sock *sk = (struct sock *)msk; 2732 unsigned long fail_tout; 2733 int state; 2734 2735 lock_sock(sk); 2736 state = sk->sk_state; 2737 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN))) 2738 goto unlock; 2739 2740 mptcp_check_fastclose(msk); 2741 2742 mptcp_pm_worker(msk); 2743 2744 mptcp_check_send_data_fin(sk); 2745 mptcp_check_data_fin_ack(sk); 2746 mptcp_check_data_fin(sk); 2747 2748 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2749 __mptcp_close_subflow(sk); 2750 2751 if (mptcp_close_tout_expired(sk)) { 2752 mptcp_do_fastclose(sk); 2753 mptcp_close_wake_up(sk); 2754 } 2755 2756 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) { 2757 __mptcp_destroy_sock(sk); 2758 goto unlock; 2759 } 2760 2761 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2762 __mptcp_retrans(sk); 2763 2764 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0; 2765 if (fail_tout && time_after(jiffies, fail_tout)) 2766 mptcp_mp_fail_no_response(msk); 2767 2768 unlock: 2769 release_sock(sk); 2770 sock_put(sk); 2771 } 2772 2773 static void __mptcp_init_sock(struct sock *sk) 2774 { 2775 struct mptcp_sock *msk = mptcp_sk(sk); 2776 2777 INIT_LIST_HEAD(&msk->conn_list); 2778 INIT_LIST_HEAD(&msk->join_list); 2779 INIT_LIST_HEAD(&msk->rtx_queue); 2780 INIT_WORK(&msk->work, mptcp_worker); 2781 msk->out_of_order_queue = RB_ROOT; 2782 msk->first_pending = NULL; 2783 msk->timer_ival = TCP_RTO_MIN; 2784 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO; 2785 2786 WRITE_ONCE(msk->first, NULL); 2787 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2788 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2789 msk->allow_infinite_fallback = true; 2790 msk->allow_subflows = true; 2791 msk->recovery = false; 2792 msk->subflow_id = 1; 2793 msk->last_data_sent = tcp_jiffies32; 2794 msk->last_data_recv = tcp_jiffies32; 2795 msk->last_ack_recv = tcp_jiffies32; 2796 2797 mptcp_pm_data_init(msk); 2798 spin_lock_init(&msk->fallback_lock); 2799 2800 /* re-use the csk retrans timer for MPTCP-level retrans */ 2801 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2802 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0); 2803 } 2804 2805 static void mptcp_ca_reset(struct sock *sk) 2806 { 2807 struct inet_connection_sock *icsk = inet_csk(sk); 2808 2809 tcp_assign_congestion_control(sk); 2810 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name, 2811 sizeof(mptcp_sk(sk)->ca_name)); 2812 2813 /* no need to keep a reference to the ops, the name will suffice */ 2814 tcp_cleanup_congestion_control(sk); 2815 icsk->icsk_ca_ops = NULL; 2816 } 2817 2818 static int mptcp_init_sock(struct sock *sk) 2819 { 2820 struct net *net = sock_net(sk); 2821 int ret; 2822 2823 __mptcp_init_sock(sk); 2824 2825 if (!mptcp_is_enabled(net)) 2826 return -ENOPROTOOPT; 2827 2828 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2829 return -ENOMEM; 2830 2831 rcu_read_lock(); 2832 ret = mptcp_init_sched(mptcp_sk(sk), 2833 mptcp_sched_find(mptcp_get_scheduler(net))); 2834 rcu_read_unlock(); 2835 if (ret) 2836 return ret; 2837 2838 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); 2839 2840 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2841 * propagate the correct value 2842 */ 2843 mptcp_ca_reset(sk); 2844 2845 sk_sockets_allocated_inc(sk); 2846 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]); 2847 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]); 2848 2849 return 0; 2850 } 2851 2852 static void __mptcp_clear_xmit(struct sock *sk) 2853 { 2854 struct mptcp_sock *msk = mptcp_sk(sk); 2855 struct mptcp_data_frag *dtmp, *dfrag; 2856 2857 WRITE_ONCE(msk->first_pending, NULL); 2858 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2859 dfrag_clear(sk, dfrag); 2860 } 2861 2862 void mptcp_cancel_work(struct sock *sk) 2863 { 2864 struct mptcp_sock *msk = mptcp_sk(sk); 2865 2866 if (cancel_work_sync(&msk->work)) 2867 __sock_put(sk); 2868 } 2869 2870 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2871 { 2872 lock_sock(ssk); 2873 2874 switch (ssk->sk_state) { 2875 case TCP_LISTEN: 2876 if (!(how & RCV_SHUTDOWN)) 2877 break; 2878 fallthrough; 2879 case TCP_SYN_SENT: 2880 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK)); 2881 break; 2882 default: 2883 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2884 pr_debug("Fallback\n"); 2885 ssk->sk_shutdown |= how; 2886 tcp_shutdown(ssk, how); 2887 2888 /* simulate the data_fin ack reception to let the state 2889 * machine move forward 2890 */ 2891 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt); 2892 mptcp_schedule_work(sk); 2893 } else { 2894 pr_debug("Sending DATA_FIN on subflow %p\n", ssk); 2895 tcp_send_ack(ssk); 2896 if (!mptcp_rtx_timer_pending(sk)) 2897 mptcp_reset_rtx_timer(sk); 2898 } 2899 break; 2900 } 2901 2902 release_sock(ssk); 2903 } 2904 2905 void mptcp_set_state(struct sock *sk, int state) 2906 { 2907 int oldstate = sk->sk_state; 2908 2909 switch (state) { 2910 case TCP_ESTABLISHED: 2911 if (oldstate != TCP_ESTABLISHED) 2912 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2913 break; 2914 case TCP_CLOSE_WAIT: 2915 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state: 2916 * MPTCP "accepted" sockets will be created later on. So no 2917 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT. 2918 */ 2919 break; 2920 default: 2921 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 2922 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2923 } 2924 2925 inet_sk_state_store(sk, state); 2926 } 2927 2928 static const unsigned char new_state[16] = { 2929 /* current state: new state: action: */ 2930 [0 /* (Invalid) */] = TCP_CLOSE, 2931 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2932 [TCP_SYN_SENT] = TCP_CLOSE, 2933 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2934 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2935 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2936 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2937 [TCP_CLOSE] = TCP_CLOSE, 2938 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2939 [TCP_LAST_ACK] = TCP_LAST_ACK, 2940 [TCP_LISTEN] = TCP_CLOSE, 2941 [TCP_CLOSING] = TCP_CLOSING, 2942 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2943 }; 2944 2945 static int mptcp_close_state(struct sock *sk) 2946 { 2947 int next = (int)new_state[sk->sk_state]; 2948 int ns = next & TCP_STATE_MASK; 2949 2950 mptcp_set_state(sk, ns); 2951 2952 return next & TCP_ACTION_FIN; 2953 } 2954 2955 static void mptcp_check_send_data_fin(struct sock *sk) 2956 { 2957 struct mptcp_subflow_context *subflow; 2958 struct mptcp_sock *msk = mptcp_sk(sk); 2959 2960 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n", 2961 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2962 msk->snd_nxt, msk->write_seq); 2963 2964 /* we still need to enqueue subflows or not really shutting down, 2965 * skip this 2966 */ 2967 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2968 mptcp_send_head(sk)) 2969 return; 2970 2971 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2972 2973 mptcp_for_each_subflow(msk, subflow) { 2974 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2975 2976 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2977 } 2978 } 2979 2980 static void __mptcp_wr_shutdown(struct sock *sk) 2981 { 2982 struct mptcp_sock *msk = mptcp_sk(sk); 2983 2984 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n", 2985 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2986 !!mptcp_send_head(sk)); 2987 2988 /* will be ignored by fallback sockets */ 2989 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2990 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2991 2992 mptcp_check_send_data_fin(sk); 2993 } 2994 2995 static void __mptcp_destroy_sock(struct sock *sk) 2996 { 2997 struct mptcp_sock *msk = mptcp_sk(sk); 2998 2999 pr_debug("msk=%p\n", msk); 3000 3001 might_sleep(); 3002 3003 mptcp_stop_rtx_timer(sk); 3004 sk_stop_timer(sk, &sk->sk_timer); 3005 msk->pm.status = 0; 3006 mptcp_release_sched(msk); 3007 3008 sk->sk_prot->destroy(sk); 3009 3010 sk_stream_kill_queues(sk); 3011 xfrm_sk_free_policy(sk); 3012 3013 sock_put(sk); 3014 } 3015 3016 void __mptcp_unaccepted_force_close(struct sock *sk) 3017 { 3018 sock_set_flag(sk, SOCK_DEAD); 3019 mptcp_do_fastclose(sk); 3020 __mptcp_destroy_sock(sk); 3021 } 3022 3023 static __poll_t mptcp_check_readable(struct sock *sk) 3024 { 3025 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0; 3026 } 3027 3028 static void mptcp_check_listen_stop(struct sock *sk) 3029 { 3030 struct sock *ssk; 3031 3032 if (inet_sk_state_load(sk) != TCP_LISTEN) 3033 return; 3034 3035 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 3036 ssk = mptcp_sk(sk)->first; 3037 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN)) 3038 return; 3039 3040 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 3041 tcp_set_state(ssk, TCP_CLOSE); 3042 mptcp_subflow_queue_clean(sk, ssk); 3043 inet_csk_listen_stop(ssk); 3044 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED); 3045 release_sock(ssk); 3046 } 3047 3048 bool __mptcp_close(struct sock *sk, long timeout) 3049 { 3050 struct mptcp_subflow_context *subflow; 3051 struct mptcp_sock *msk = mptcp_sk(sk); 3052 bool do_cancel_work = false; 3053 int subflows_alive = 0; 3054 3055 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3056 3057 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 3058 mptcp_check_listen_stop(sk); 3059 mptcp_set_state(sk, TCP_CLOSE); 3060 goto cleanup; 3061 } 3062 3063 if (mptcp_data_avail(msk) || timeout < 0) { 3064 /* If the msk has read data, or the caller explicitly ask it, 3065 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose 3066 */ 3067 mptcp_do_fastclose(sk); 3068 timeout = 0; 3069 } else if (mptcp_close_state(sk)) { 3070 __mptcp_wr_shutdown(sk); 3071 } 3072 3073 sk_stream_wait_close(sk, timeout); 3074 3075 cleanup: 3076 /* orphan all the subflows */ 3077 mptcp_for_each_subflow(msk, subflow) { 3078 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3079 bool slow = lock_sock_fast_nested(ssk); 3080 3081 subflows_alive += ssk->sk_state != TCP_CLOSE; 3082 3083 /* since the close timeout takes precedence on the fail one, 3084 * cancel the latter 3085 */ 3086 if (ssk == msk->first) 3087 subflow->fail_tout = 0; 3088 3089 /* detach from the parent socket, but allow data_ready to 3090 * push incoming data into the mptcp stack, to properly ack it 3091 */ 3092 ssk->sk_socket = NULL; 3093 ssk->sk_wq = NULL; 3094 unlock_sock_fast(ssk, slow); 3095 } 3096 sock_orphan(sk); 3097 3098 /* all the subflows are closed, only timeout can change the msk 3099 * state, let's not keep resources busy for no reasons 3100 */ 3101 if (subflows_alive == 0) 3102 mptcp_set_state(sk, TCP_CLOSE); 3103 3104 sock_hold(sk); 3105 pr_debug("msk=%p state=%d\n", sk, sk->sk_state); 3106 mptcp_pm_connection_closed(msk); 3107 3108 if (sk->sk_state == TCP_CLOSE) { 3109 __mptcp_destroy_sock(sk); 3110 do_cancel_work = true; 3111 } else { 3112 mptcp_start_tout_timer(sk); 3113 } 3114 3115 return do_cancel_work; 3116 } 3117 3118 static void mptcp_close(struct sock *sk, long timeout) 3119 { 3120 bool do_cancel_work; 3121 3122 lock_sock(sk); 3123 3124 do_cancel_work = __mptcp_close(sk, timeout); 3125 release_sock(sk); 3126 if (do_cancel_work) 3127 mptcp_cancel_work(sk); 3128 3129 sock_put(sk); 3130 } 3131 3132 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 3133 { 3134 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3135 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 3136 struct ipv6_pinfo *msk6 = inet6_sk(msk); 3137 3138 msk->sk_v6_daddr = ssk->sk_v6_daddr; 3139 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 3140 3141 if (msk6 && ssk6) { 3142 msk6->saddr = ssk6->saddr; 3143 msk6->flow_label = ssk6->flow_label; 3144 } 3145 #endif 3146 3147 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 3148 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 3149 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 3150 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 3151 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 3152 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 3153 } 3154 3155 static int mptcp_disconnect(struct sock *sk, int flags) 3156 { 3157 struct mptcp_sock *msk = mptcp_sk(sk); 3158 3159 /* We are on the fastopen error path. We can't call straight into the 3160 * subflows cleanup code due to lock nesting (we are already under 3161 * msk->firstsocket lock). 3162 */ 3163 if (msk->fastopening) 3164 return -EBUSY; 3165 3166 mptcp_check_listen_stop(sk); 3167 mptcp_set_state(sk, TCP_CLOSE); 3168 3169 mptcp_stop_rtx_timer(sk); 3170 mptcp_stop_tout_timer(sk); 3171 3172 mptcp_pm_connection_closed(msk); 3173 3174 /* msk->subflow is still intact, the following will not free the first 3175 * subflow 3176 */ 3177 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE); 3178 3179 /* The first subflow is already in TCP_CLOSE status, the following 3180 * can't overlap with a fallback anymore 3181 */ 3182 spin_lock_bh(&msk->fallback_lock); 3183 msk->allow_subflows = true; 3184 msk->allow_infinite_fallback = true; 3185 WRITE_ONCE(msk->flags, 0); 3186 spin_unlock_bh(&msk->fallback_lock); 3187 3188 msk->cb_flags = 0; 3189 msk->recovery = false; 3190 WRITE_ONCE(msk->can_ack, false); 3191 WRITE_ONCE(msk->fully_established, false); 3192 WRITE_ONCE(msk->rcv_data_fin, false); 3193 WRITE_ONCE(msk->snd_data_fin_enable, false); 3194 WRITE_ONCE(msk->rcv_fastclose, false); 3195 WRITE_ONCE(msk->use_64bit_ack, false); 3196 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 3197 mptcp_pm_data_reset(msk); 3198 mptcp_ca_reset(sk); 3199 msk->bytes_consumed = 0; 3200 msk->bytes_acked = 0; 3201 msk->bytes_received = 0; 3202 msk->bytes_sent = 0; 3203 msk->bytes_retrans = 0; 3204 msk->rcvspace_init = 0; 3205 3206 WRITE_ONCE(sk->sk_shutdown, 0); 3207 sk_error_report(sk); 3208 return 0; 3209 } 3210 3211 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3212 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 3213 { 3214 struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk); 3215 3216 return &msk6->np; 3217 } 3218 3219 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk) 3220 { 3221 const struct ipv6_pinfo *np = inet6_sk(sk); 3222 struct ipv6_txoptions *opt; 3223 struct ipv6_pinfo *newnp; 3224 3225 newnp = inet6_sk(newsk); 3226 3227 rcu_read_lock(); 3228 opt = rcu_dereference(np->opt); 3229 if (opt) { 3230 opt = ipv6_dup_options(newsk, opt); 3231 if (!opt) 3232 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__); 3233 } 3234 RCU_INIT_POINTER(newnp->opt, opt); 3235 rcu_read_unlock(); 3236 } 3237 #endif 3238 3239 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk) 3240 { 3241 struct ip_options_rcu *inet_opt, *newopt = NULL; 3242 const struct inet_sock *inet = inet_sk(sk); 3243 struct inet_sock *newinet; 3244 3245 newinet = inet_sk(newsk); 3246 3247 rcu_read_lock(); 3248 inet_opt = rcu_dereference(inet->inet_opt); 3249 if (inet_opt) { 3250 newopt = sock_kmemdup(newsk, inet_opt, sizeof(*inet_opt) + 3251 inet_opt->opt.optlen, GFP_ATOMIC); 3252 if (!newopt) 3253 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__); 3254 } 3255 RCU_INIT_POINTER(newinet->inet_opt, newopt); 3256 rcu_read_unlock(); 3257 } 3258 3259 struct sock *mptcp_sk_clone_init(const struct sock *sk, 3260 const struct mptcp_options_received *mp_opt, 3261 struct sock *ssk, 3262 struct request_sock *req) 3263 { 3264 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 3265 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 3266 struct mptcp_subflow_context *subflow; 3267 struct mptcp_sock *msk; 3268 3269 if (!nsk) 3270 return NULL; 3271 3272 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3273 if (nsk->sk_family == AF_INET6) 3274 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 3275 #endif 3276 3277 __mptcp_init_sock(nsk); 3278 3279 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3280 if (nsk->sk_family == AF_INET6) 3281 mptcp_copy_ip6_options(nsk, sk); 3282 else 3283 #endif 3284 mptcp_copy_ip_options(nsk, sk); 3285 3286 msk = mptcp_sk(nsk); 3287 WRITE_ONCE(msk->local_key, subflow_req->local_key); 3288 WRITE_ONCE(msk->token, subflow_req->token); 3289 msk->in_accept_queue = 1; 3290 WRITE_ONCE(msk->fully_established, false); 3291 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 3292 WRITE_ONCE(msk->csum_enabled, true); 3293 3294 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1); 3295 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3296 WRITE_ONCE(msk->snd_una, msk->write_seq); 3297 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 3298 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 3299 mptcp_init_sched(msk, mptcp_sk(sk)->sched); 3300 3301 /* passive msk is created after the first/MPC subflow */ 3302 msk->subflow_id = 2; 3303 3304 sock_reset_flag(nsk, SOCK_RCU_FREE); 3305 security_inet_csk_clone(nsk, req); 3306 3307 /* this can't race with mptcp_close(), as the msk is 3308 * not yet exposted to user-space 3309 */ 3310 mptcp_set_state(nsk, TCP_ESTABLISHED); 3311 3312 /* The msk maintain a ref to each subflow in the connections list */ 3313 WRITE_ONCE(msk->first, ssk); 3314 subflow = mptcp_subflow_ctx(ssk); 3315 list_add(&subflow->node, &msk->conn_list); 3316 sock_hold(ssk); 3317 3318 /* new mpc subflow takes ownership of the newly 3319 * created mptcp socket 3320 */ 3321 mptcp_token_accept(subflow_req, msk); 3322 3323 /* set msk addresses early to ensure mptcp_pm_get_local_id() 3324 * uses the correct data 3325 */ 3326 mptcp_copy_inaddrs(nsk, ssk); 3327 __mptcp_propagate_sndbuf(nsk, ssk); 3328 3329 mptcp_rcv_space_init(msk, ssk); 3330 3331 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 3332 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 3333 bh_unlock_sock(nsk); 3334 3335 /* note: the newly allocated socket refcount is 2 now */ 3336 return nsk; 3337 } 3338 3339 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3340 { 3341 const struct tcp_sock *tp = tcp_sk(ssk); 3342 3343 msk->rcvspace_init = 1; 3344 msk->rcvq_space.copied = 0; 3345 msk->rcvq_space.rtt_us = 0; 3346 3347 msk->rcvq_space.time = tp->tcp_mstamp; 3348 3349 /* initial rcv_space offering made to peer */ 3350 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3351 TCP_INIT_CWND * tp->advmss); 3352 if (msk->rcvq_space.space == 0) 3353 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3354 } 3355 3356 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags) 3357 { 3358 struct mptcp_subflow_context *subflow, *tmp; 3359 struct sock *sk = (struct sock *)msk; 3360 3361 __mptcp_clear_xmit(sk); 3362 3363 /* join list will be eventually flushed (with rst) at sock lock release time */ 3364 mptcp_for_each_subflow_safe(msk, subflow, tmp) 3365 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags); 3366 3367 __skb_queue_purge(&sk->sk_receive_queue); 3368 skb_rbtree_purge(&msk->out_of_order_queue); 3369 3370 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3371 * inet_sock_destruct() will dispose it 3372 */ 3373 mptcp_token_destroy(msk); 3374 mptcp_pm_destroy(msk); 3375 } 3376 3377 static void mptcp_destroy(struct sock *sk) 3378 { 3379 struct mptcp_sock *msk = mptcp_sk(sk); 3380 3381 /* allow the following to close even the initial subflow */ 3382 msk->free_first = 1; 3383 mptcp_destroy_common(msk, 0); 3384 sk_sockets_allocated_dec(sk); 3385 } 3386 3387 void __mptcp_data_acked(struct sock *sk) 3388 { 3389 if (!sock_owned_by_user(sk)) 3390 __mptcp_clean_una(sk); 3391 else 3392 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3393 } 3394 3395 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3396 { 3397 if (!mptcp_send_head(sk)) 3398 return; 3399 3400 if (!sock_owned_by_user(sk)) 3401 __mptcp_subflow_push_pending(sk, ssk, false); 3402 else 3403 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3404 } 3405 3406 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3407 BIT(MPTCP_RETRANSMIT) | \ 3408 BIT(MPTCP_FLUSH_JOIN_LIST) | \ 3409 BIT(MPTCP_DEQUEUE)) 3410 3411 /* processes deferred events and flush wmem */ 3412 static void mptcp_release_cb(struct sock *sk) 3413 __must_hold(&sk->sk_lock.slock) 3414 { 3415 struct mptcp_sock *msk = mptcp_sk(sk); 3416 3417 for (;;) { 3418 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED); 3419 struct list_head join_list; 3420 3421 if (!flags) 3422 break; 3423 3424 INIT_LIST_HEAD(&join_list); 3425 list_splice_init(&msk->join_list, &join_list); 3426 3427 /* the following actions acquire the subflow socket lock 3428 * 3429 * 1) can't be invoked in atomic scope 3430 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3431 * datapath acquires the msk socket spinlock while helding 3432 * the subflow socket lock 3433 */ 3434 msk->cb_flags &= ~flags; 3435 spin_unlock_bh(&sk->sk_lock.slock); 3436 3437 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3438 __mptcp_flush_join_list(sk, &join_list); 3439 if (flags & BIT(MPTCP_PUSH_PENDING)) 3440 __mptcp_push_pending(sk, 0); 3441 if (flags & BIT(MPTCP_RETRANSMIT)) 3442 __mptcp_retrans(sk); 3443 if ((flags & BIT(MPTCP_DEQUEUE)) && __mptcp_move_skbs(sk)) { 3444 /* notify ack seq update */ 3445 mptcp_cleanup_rbuf(msk, 0); 3446 sk->sk_data_ready(sk); 3447 } 3448 3449 cond_resched(); 3450 spin_lock_bh(&sk->sk_lock.slock); 3451 } 3452 3453 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3454 __mptcp_clean_una_wakeup(sk); 3455 if (unlikely(msk->cb_flags)) { 3456 /* be sure to sync the msk state before taking actions 3457 * depending on sk_state (MPTCP_ERROR_REPORT) 3458 * On sk release avoid actions depending on the first subflow 3459 */ 3460 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first) 3461 __mptcp_sync_state(sk, msk->pending_state); 3462 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3463 __mptcp_error_report(sk); 3464 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags)) 3465 __mptcp_sync_sndbuf(sk); 3466 } 3467 } 3468 3469 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3470 * TCP can't schedule delack timer before the subflow is fully established. 3471 * MPTCP uses the delack timer to do 3rd ack retransmissions 3472 */ 3473 static void schedule_3rdack_retransmission(struct sock *ssk) 3474 { 3475 struct inet_connection_sock *icsk = inet_csk(ssk); 3476 struct tcp_sock *tp = tcp_sk(ssk); 3477 unsigned long timeout; 3478 3479 if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established)) 3480 return; 3481 3482 /* reschedule with a timeout above RTT, as we must look only for drop */ 3483 if (tp->srtt_us) 3484 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3485 else 3486 timeout = TCP_TIMEOUT_INIT; 3487 timeout += jiffies; 3488 3489 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3490 smp_store_release(&icsk->icsk_ack.pending, 3491 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER); 3492 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3493 } 3494 3495 void mptcp_subflow_process_delegated(struct sock *ssk, long status) 3496 { 3497 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3498 struct sock *sk = subflow->conn; 3499 3500 if (status & BIT(MPTCP_DELEGATE_SEND)) { 3501 mptcp_data_lock(sk); 3502 if (!sock_owned_by_user(sk)) 3503 __mptcp_subflow_push_pending(sk, ssk, true); 3504 else 3505 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3506 mptcp_data_unlock(sk); 3507 } 3508 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) { 3509 mptcp_data_lock(sk); 3510 if (!sock_owned_by_user(sk)) 3511 __mptcp_sync_sndbuf(sk); 3512 else 3513 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags); 3514 mptcp_data_unlock(sk); 3515 } 3516 if (status & BIT(MPTCP_DELEGATE_ACK)) 3517 schedule_3rdack_retransmission(ssk); 3518 } 3519 3520 static int mptcp_hash(struct sock *sk) 3521 { 3522 /* should never be called, 3523 * we hash the TCP subflows not the MPTCP socket 3524 */ 3525 WARN_ON_ONCE(1); 3526 return 0; 3527 } 3528 3529 static void mptcp_unhash(struct sock *sk) 3530 { 3531 /* called from sk_common_release(), but nothing to do here */ 3532 } 3533 3534 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3535 { 3536 struct mptcp_sock *msk = mptcp_sk(sk); 3537 3538 pr_debug("msk=%p, ssk=%p\n", msk, msk->first); 3539 if (WARN_ON_ONCE(!msk->first)) 3540 return -EINVAL; 3541 3542 return inet_csk_get_port(msk->first, snum); 3543 } 3544 3545 void mptcp_finish_connect(struct sock *ssk) 3546 { 3547 struct mptcp_subflow_context *subflow; 3548 struct mptcp_sock *msk; 3549 struct sock *sk; 3550 3551 subflow = mptcp_subflow_ctx(ssk); 3552 sk = subflow->conn; 3553 msk = mptcp_sk(sk); 3554 3555 pr_debug("msk=%p, token=%u\n", sk, subflow->token); 3556 3557 subflow->map_seq = subflow->iasn; 3558 subflow->map_subflow_seq = 1; 3559 3560 /* the socket is not connected yet, no msk/subflow ops can access/race 3561 * accessing the field below 3562 */ 3563 WRITE_ONCE(msk->local_key, subflow->local_key); 3564 3565 mptcp_pm_new_connection(msk, ssk, 0); 3566 } 3567 3568 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3569 { 3570 write_lock_bh(&sk->sk_callback_lock); 3571 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3572 sk_set_socket(sk, parent); 3573 write_unlock_bh(&sk->sk_callback_lock); 3574 } 3575 3576 bool mptcp_finish_join(struct sock *ssk) 3577 { 3578 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3579 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3580 struct sock *parent = (void *)msk; 3581 bool ret = true; 3582 3583 pr_debug("msk=%p, subflow=%p\n", msk, subflow); 3584 3585 /* mptcp socket already closing? */ 3586 if (!mptcp_is_fully_established(parent)) { 3587 subflow->reset_reason = MPTCP_RST_EMPTCP; 3588 return false; 3589 } 3590 3591 /* active subflow, already present inside the conn_list */ 3592 if (!list_empty(&subflow->node)) { 3593 spin_lock_bh(&msk->fallback_lock); 3594 if (!msk->allow_subflows) { 3595 spin_unlock_bh(&msk->fallback_lock); 3596 return false; 3597 } 3598 mptcp_subflow_joined(msk, ssk); 3599 spin_unlock_bh(&msk->fallback_lock); 3600 mptcp_propagate_sndbuf(parent, ssk); 3601 return true; 3602 } 3603 3604 if (!mptcp_pm_allow_new_subflow(msk)) { 3605 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_JOINREJECTED); 3606 goto err_prohibited; 3607 } 3608 3609 /* If we can't acquire msk socket lock here, let the release callback 3610 * handle it 3611 */ 3612 mptcp_data_lock(parent); 3613 if (!sock_owned_by_user(parent)) { 3614 ret = __mptcp_finish_join(msk, ssk); 3615 if (ret) { 3616 sock_hold(ssk); 3617 list_add_tail(&subflow->node, &msk->conn_list); 3618 } 3619 } else { 3620 sock_hold(ssk); 3621 list_add_tail(&subflow->node, &msk->join_list); 3622 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3623 } 3624 mptcp_data_unlock(parent); 3625 3626 if (!ret) { 3627 err_prohibited: 3628 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3629 return false; 3630 } 3631 3632 return true; 3633 } 3634 3635 static void mptcp_shutdown(struct sock *sk, int how) 3636 { 3637 pr_debug("sk=%p, how=%d\n", sk, how); 3638 3639 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3640 __mptcp_wr_shutdown(sk); 3641 } 3642 3643 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3644 { 3645 const struct sock *sk = (void *)msk; 3646 u64 delta; 3647 3648 if (sk->sk_state == TCP_LISTEN) 3649 return -EINVAL; 3650 3651 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3652 return 0; 3653 3654 delta = msk->write_seq - v; 3655 if (__mptcp_check_fallback(msk) && msk->first) { 3656 struct tcp_sock *tp = tcp_sk(msk->first); 3657 3658 /* the first subflow is disconnected after close - see 3659 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3660 * so ignore that status, too. 3661 */ 3662 if (!((1 << msk->first->sk_state) & 3663 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3664 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3665 } 3666 if (delta > INT_MAX) 3667 delta = INT_MAX; 3668 3669 return (int)delta; 3670 } 3671 3672 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg) 3673 { 3674 struct mptcp_sock *msk = mptcp_sk(sk); 3675 bool slow; 3676 3677 switch (cmd) { 3678 case SIOCINQ: 3679 if (sk->sk_state == TCP_LISTEN) 3680 return -EINVAL; 3681 3682 lock_sock(sk); 3683 if (__mptcp_move_skbs(sk)) 3684 mptcp_cleanup_rbuf(msk, 0); 3685 *karg = mptcp_inq_hint(sk); 3686 release_sock(sk); 3687 break; 3688 case SIOCOUTQ: 3689 slow = lock_sock_fast(sk); 3690 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3691 unlock_sock_fast(sk, slow); 3692 break; 3693 case SIOCOUTQNSD: 3694 slow = lock_sock_fast(sk); 3695 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt); 3696 unlock_sock_fast(sk, slow); 3697 break; 3698 default: 3699 return -ENOIOCTLCMD; 3700 } 3701 3702 return 0; 3703 } 3704 3705 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) 3706 { 3707 struct mptcp_subflow_context *subflow; 3708 struct mptcp_sock *msk = mptcp_sk(sk); 3709 int err = -EINVAL; 3710 struct sock *ssk; 3711 3712 ssk = __mptcp_nmpc_sk(msk); 3713 if (IS_ERR(ssk)) 3714 return PTR_ERR(ssk); 3715 3716 mptcp_set_state(sk, TCP_SYN_SENT); 3717 subflow = mptcp_subflow_ctx(ssk); 3718 #ifdef CONFIG_TCP_MD5SIG 3719 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3720 * TCP option space. 3721 */ 3722 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info)) 3723 mptcp_early_fallback(msk, subflow, MPTCP_MIB_MD5SIGFALLBACK); 3724 #endif 3725 if (subflow->request_mptcp) { 3726 if (mptcp_active_should_disable(sk)) 3727 mptcp_early_fallback(msk, subflow, 3728 MPTCP_MIB_MPCAPABLEACTIVEDISABLED); 3729 else if (mptcp_token_new_connect(ssk) < 0) 3730 mptcp_early_fallback(msk, subflow, 3731 MPTCP_MIB_TOKENFALLBACKINIT); 3732 } 3733 3734 WRITE_ONCE(msk->write_seq, subflow->idsn); 3735 WRITE_ONCE(msk->snd_nxt, subflow->idsn); 3736 WRITE_ONCE(msk->snd_una, subflow->idsn); 3737 if (likely(!__mptcp_check_fallback(msk))) 3738 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE); 3739 3740 /* if reaching here via the fastopen/sendmsg path, the caller already 3741 * acquired the subflow socket lock, too. 3742 */ 3743 if (!msk->fastopening) 3744 lock_sock(ssk); 3745 3746 /* the following mirrors closely a very small chunk of code from 3747 * __inet_stream_connect() 3748 */ 3749 if (ssk->sk_state != TCP_CLOSE) 3750 goto out; 3751 3752 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) { 3753 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len); 3754 if (err) 3755 goto out; 3756 } 3757 3758 err = ssk->sk_prot->connect(ssk, uaddr, addr_len); 3759 if (err < 0) 3760 goto out; 3761 3762 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk)); 3763 3764 out: 3765 if (!msk->fastopening) 3766 release_sock(ssk); 3767 3768 /* on successful connect, the msk state will be moved to established by 3769 * subflow_finish_connect() 3770 */ 3771 if (unlikely(err)) { 3772 /* avoid leaving a dangling token in an unconnected socket */ 3773 mptcp_token_destroy(msk); 3774 mptcp_set_state(sk, TCP_CLOSE); 3775 return err; 3776 } 3777 3778 mptcp_copy_inaddrs(sk, ssk); 3779 return 0; 3780 } 3781 3782 static struct proto mptcp_prot = { 3783 .name = "MPTCP", 3784 .owner = THIS_MODULE, 3785 .init = mptcp_init_sock, 3786 .connect = mptcp_connect, 3787 .disconnect = mptcp_disconnect, 3788 .close = mptcp_close, 3789 .setsockopt = mptcp_setsockopt, 3790 .getsockopt = mptcp_getsockopt, 3791 .shutdown = mptcp_shutdown, 3792 .destroy = mptcp_destroy, 3793 .sendmsg = mptcp_sendmsg, 3794 .ioctl = mptcp_ioctl, 3795 .recvmsg = mptcp_recvmsg, 3796 .release_cb = mptcp_release_cb, 3797 .hash = mptcp_hash, 3798 .unhash = mptcp_unhash, 3799 .get_port = mptcp_get_port, 3800 .stream_memory_free = mptcp_stream_memory_free, 3801 .sockets_allocated = &mptcp_sockets_allocated, 3802 3803 .memory_allocated = &net_aligned_data.tcp_memory_allocated, 3804 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc, 3805 3806 .memory_pressure = &tcp_memory_pressure, 3807 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3808 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3809 .sysctl_mem = sysctl_tcp_mem, 3810 .obj_size = sizeof(struct mptcp_sock), 3811 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3812 .no_autobind = true, 3813 }; 3814 3815 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3816 { 3817 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3818 struct sock *ssk, *sk = sock->sk; 3819 int err = -EINVAL; 3820 3821 lock_sock(sk); 3822 ssk = __mptcp_nmpc_sk(msk); 3823 if (IS_ERR(ssk)) { 3824 err = PTR_ERR(ssk); 3825 goto unlock; 3826 } 3827 3828 if (sk->sk_family == AF_INET) 3829 err = inet_bind_sk(ssk, uaddr, addr_len); 3830 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3831 else if (sk->sk_family == AF_INET6) 3832 err = inet6_bind_sk(ssk, uaddr, addr_len); 3833 #endif 3834 if (!err) 3835 mptcp_copy_inaddrs(sk, ssk); 3836 3837 unlock: 3838 release_sock(sk); 3839 return err; 3840 } 3841 3842 static int mptcp_listen(struct socket *sock, int backlog) 3843 { 3844 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3845 struct sock *sk = sock->sk; 3846 struct sock *ssk; 3847 int err; 3848 3849 pr_debug("msk=%p\n", msk); 3850 3851 lock_sock(sk); 3852 3853 err = -EINVAL; 3854 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM) 3855 goto unlock; 3856 3857 ssk = __mptcp_nmpc_sk(msk); 3858 if (IS_ERR(ssk)) { 3859 err = PTR_ERR(ssk); 3860 goto unlock; 3861 } 3862 3863 mptcp_set_state(sk, TCP_LISTEN); 3864 sock_set_flag(sk, SOCK_RCU_FREE); 3865 3866 lock_sock(ssk); 3867 err = __inet_listen_sk(ssk, backlog); 3868 release_sock(ssk); 3869 mptcp_set_state(sk, inet_sk_state_load(ssk)); 3870 3871 if (!err) { 3872 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 3873 mptcp_copy_inaddrs(sk, ssk); 3874 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED); 3875 } 3876 3877 unlock: 3878 release_sock(sk); 3879 return err; 3880 } 3881 3882 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3883 struct proto_accept_arg *arg) 3884 { 3885 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3886 struct sock *ssk, *newsk; 3887 3888 pr_debug("msk=%p\n", msk); 3889 3890 /* Buggy applications can call accept on socket states other then LISTEN 3891 * but no need to allocate the first subflow just to error out. 3892 */ 3893 ssk = READ_ONCE(msk->first); 3894 if (!ssk) 3895 return -EINVAL; 3896 3897 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk)); 3898 newsk = inet_csk_accept(ssk, arg); 3899 if (!newsk) 3900 return arg->err; 3901 3902 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk)); 3903 if (sk_is_mptcp(newsk)) { 3904 struct mptcp_subflow_context *subflow; 3905 struct sock *new_mptcp_sock; 3906 3907 subflow = mptcp_subflow_ctx(newsk); 3908 new_mptcp_sock = subflow->conn; 3909 3910 /* is_mptcp should be false if subflow->conn is missing, see 3911 * subflow_syn_recv_sock() 3912 */ 3913 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3914 tcp_sk(newsk)->is_mptcp = 0; 3915 goto tcpfallback; 3916 } 3917 3918 newsk = new_mptcp_sock; 3919 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3920 3921 newsk->sk_kern_sock = arg->kern; 3922 lock_sock(newsk); 3923 __inet_accept(sock, newsock, newsk); 3924 3925 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags); 3926 msk = mptcp_sk(newsk); 3927 msk->in_accept_queue = 0; 3928 3929 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3930 * This is needed so NOSPACE flag can be set from tcp stack. 3931 */ 3932 mptcp_for_each_subflow(msk, subflow) { 3933 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3934 3935 if (!ssk->sk_socket) 3936 mptcp_sock_graft(ssk, newsock); 3937 } 3938 3939 /* Do late cleanup for the first subflow as necessary. Also 3940 * deal with bad peers not doing a complete shutdown. 3941 */ 3942 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) { 3943 __mptcp_close_ssk(newsk, msk->first, 3944 mptcp_subflow_ctx(msk->first), 0); 3945 if (unlikely(list_is_singular(&msk->conn_list))) 3946 mptcp_set_state(newsk, TCP_CLOSE); 3947 } 3948 } else { 3949 tcpfallback: 3950 newsk->sk_kern_sock = arg->kern; 3951 lock_sock(newsk); 3952 __inet_accept(sock, newsock, newsk); 3953 /* we are being invoked after accepting a non-mp-capable 3954 * flow: sk is a tcp_sk, not an mptcp one. 3955 * 3956 * Hand the socket over to tcp so all further socket ops 3957 * bypass mptcp. 3958 */ 3959 WRITE_ONCE(newsock->sk->sk_socket->ops, 3960 mptcp_fallback_tcp_ops(newsock->sk)); 3961 } 3962 release_sock(newsk); 3963 3964 return 0; 3965 } 3966 3967 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3968 { 3969 struct sock *sk = (struct sock *)msk; 3970 3971 if (__mptcp_stream_is_writeable(sk, 1)) 3972 return EPOLLOUT | EPOLLWRNORM; 3973 3974 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 3975 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */ 3976 if (__mptcp_stream_is_writeable(sk, 1)) 3977 return EPOLLOUT | EPOLLWRNORM; 3978 3979 return 0; 3980 } 3981 3982 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3983 struct poll_table_struct *wait) 3984 { 3985 struct sock *sk = sock->sk; 3986 struct mptcp_sock *msk; 3987 __poll_t mask = 0; 3988 u8 shutdown; 3989 int state; 3990 3991 msk = mptcp_sk(sk); 3992 sock_poll_wait(file, sock, wait); 3993 3994 state = inet_sk_state_load(sk); 3995 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags); 3996 if (state == TCP_LISTEN) { 3997 struct sock *ssk = READ_ONCE(msk->first); 3998 3999 if (WARN_ON_ONCE(!ssk)) 4000 return 0; 4001 4002 return inet_csk_listen_poll(ssk); 4003 } 4004 4005 shutdown = READ_ONCE(sk->sk_shutdown); 4006 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 4007 mask |= EPOLLHUP; 4008 if (shutdown & RCV_SHUTDOWN) 4009 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 4010 4011 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 4012 mask |= mptcp_check_readable(sk); 4013 if (shutdown & SEND_SHUTDOWN) 4014 mask |= EPOLLOUT | EPOLLWRNORM; 4015 else 4016 mask |= mptcp_check_writeable(msk); 4017 } else if (state == TCP_SYN_SENT && 4018 inet_test_bit(DEFER_CONNECT, sk)) { 4019 /* cf tcp_poll() note about TFO */ 4020 mask |= EPOLLOUT | EPOLLWRNORM; 4021 } 4022 4023 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */ 4024 smp_rmb(); 4025 if (READ_ONCE(sk->sk_err)) 4026 mask |= EPOLLERR; 4027 4028 return mask; 4029 } 4030 4031 static const struct proto_ops mptcp_stream_ops = { 4032 .family = PF_INET, 4033 .owner = THIS_MODULE, 4034 .release = inet_release, 4035 .bind = mptcp_bind, 4036 .connect = inet_stream_connect, 4037 .socketpair = sock_no_socketpair, 4038 .accept = mptcp_stream_accept, 4039 .getname = inet_getname, 4040 .poll = mptcp_poll, 4041 .ioctl = inet_ioctl, 4042 .gettstamp = sock_gettstamp, 4043 .listen = mptcp_listen, 4044 .shutdown = inet_shutdown, 4045 .setsockopt = sock_common_setsockopt, 4046 .getsockopt = sock_common_getsockopt, 4047 .sendmsg = inet_sendmsg, 4048 .recvmsg = inet_recvmsg, 4049 .mmap = sock_no_mmap, 4050 .set_rcvlowat = mptcp_set_rcvlowat, 4051 }; 4052 4053 static struct inet_protosw mptcp_protosw = { 4054 .type = SOCK_STREAM, 4055 .protocol = IPPROTO_MPTCP, 4056 .prot = &mptcp_prot, 4057 .ops = &mptcp_stream_ops, 4058 .flags = INET_PROTOSW_ICSK, 4059 }; 4060 4061 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 4062 { 4063 struct mptcp_delegated_action *delegated; 4064 struct mptcp_subflow_context *subflow; 4065 int work_done = 0; 4066 4067 delegated = container_of(napi, struct mptcp_delegated_action, napi); 4068 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 4069 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4070 4071 bh_lock_sock_nested(ssk); 4072 if (!sock_owned_by_user(ssk)) { 4073 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0)); 4074 } else { 4075 /* tcp_release_cb_override already processed 4076 * the action or will do at next release_sock(). 4077 * In both case must dequeue the subflow here - on the same 4078 * CPU that scheduled it. 4079 */ 4080 smp_wmb(); 4081 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status); 4082 } 4083 bh_unlock_sock(ssk); 4084 sock_put(ssk); 4085 4086 if (++work_done == budget) 4087 return budget; 4088 } 4089 4090 /* always provide a 0 'work_done' argument, so that napi_complete_done 4091 * will not try accessing the NULL napi->dev ptr 4092 */ 4093 napi_complete_done(napi, 0); 4094 return work_done; 4095 } 4096 4097 void __init mptcp_proto_init(void) 4098 { 4099 struct mptcp_delegated_action *delegated; 4100 int cpu; 4101 4102 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 4103 4104 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 4105 panic("Failed to allocate MPTCP pcpu counter\n"); 4106 4107 mptcp_napi_dev = alloc_netdev_dummy(0); 4108 if (!mptcp_napi_dev) 4109 panic("Failed to allocate MPTCP dummy netdev\n"); 4110 for_each_possible_cpu(cpu) { 4111 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 4112 INIT_LIST_HEAD(&delegated->head); 4113 netif_napi_add_tx(mptcp_napi_dev, &delegated->napi, 4114 mptcp_napi_poll); 4115 napi_enable(&delegated->napi); 4116 } 4117 4118 mptcp_subflow_init(); 4119 mptcp_pm_init(); 4120 mptcp_sched_init(); 4121 mptcp_token_init(); 4122 4123 if (proto_register(&mptcp_prot, 1) != 0) 4124 panic("Failed to register MPTCP proto.\n"); 4125 4126 inet_register_protosw(&mptcp_protosw); 4127 4128 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 4129 } 4130 4131 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 4132 static const struct proto_ops mptcp_v6_stream_ops = { 4133 .family = PF_INET6, 4134 .owner = THIS_MODULE, 4135 .release = inet6_release, 4136 .bind = mptcp_bind, 4137 .connect = inet_stream_connect, 4138 .socketpair = sock_no_socketpair, 4139 .accept = mptcp_stream_accept, 4140 .getname = inet6_getname, 4141 .poll = mptcp_poll, 4142 .ioctl = inet6_ioctl, 4143 .gettstamp = sock_gettstamp, 4144 .listen = mptcp_listen, 4145 .shutdown = inet_shutdown, 4146 .setsockopt = sock_common_setsockopt, 4147 .getsockopt = sock_common_getsockopt, 4148 .sendmsg = inet6_sendmsg, 4149 .recvmsg = inet6_recvmsg, 4150 .mmap = sock_no_mmap, 4151 #ifdef CONFIG_COMPAT 4152 .compat_ioctl = inet6_compat_ioctl, 4153 #endif 4154 .set_rcvlowat = mptcp_set_rcvlowat, 4155 }; 4156 4157 static struct proto mptcp_v6_prot; 4158 4159 static struct inet_protosw mptcp_v6_protosw = { 4160 .type = SOCK_STREAM, 4161 .protocol = IPPROTO_MPTCP, 4162 .prot = &mptcp_v6_prot, 4163 .ops = &mptcp_v6_stream_ops, 4164 .flags = INET_PROTOSW_ICSK, 4165 }; 4166 4167 int __init mptcp_proto_v6_init(void) 4168 { 4169 int err; 4170 4171 mptcp_v6_prot = mptcp_prot; 4172 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name)); 4173 mptcp_v6_prot.slab = NULL; 4174 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 4175 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np); 4176 4177 err = proto_register(&mptcp_v6_prot, 1); 4178 if (err) 4179 return err; 4180 4181 err = inet6_register_protosw(&mptcp_v6_protosw); 4182 if (err) 4183 proto_unregister(&mptcp_v6_prot); 4184 4185 return err; 4186 } 4187 #endif 4188