1 // SPDX-License-Identifier: GPL-2.0 2 /* Multipath TCP 3 * 4 * Copyright (c) 2017 - 2019, Intel Corporation. 5 */ 6 7 #define pr_fmt(fmt) "MPTCP: " fmt 8 9 #include <linux/kernel.h> 10 #include <linux/module.h> 11 #include <linux/netdevice.h> 12 #include <linux/sched/signal.h> 13 #include <linux/atomic.h> 14 #include <net/sock.h> 15 #include <net/inet_common.h> 16 #include <net/inet_hashtables.h> 17 #include <net/protocol.h> 18 #include <net/tcp_states.h> 19 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 20 #include <net/transp_v6.h> 21 #endif 22 #include <net/mptcp.h> 23 #include <net/hotdata.h> 24 #include <net/xfrm.h> 25 #include <asm/ioctls.h> 26 #include "protocol.h" 27 #include "mib.h" 28 29 #define CREATE_TRACE_POINTS 30 #include <trace/events/mptcp.h> 31 32 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 33 struct mptcp6_sock { 34 struct mptcp_sock msk; 35 struct ipv6_pinfo np; 36 }; 37 #endif 38 39 enum { 40 MPTCP_CMSG_TS = BIT(0), 41 MPTCP_CMSG_INQ = BIT(1), 42 }; 43 44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp; 45 46 static void __mptcp_destroy_sock(struct sock *sk); 47 static void mptcp_check_send_data_fin(struct sock *sk); 48 49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions); 50 static struct net_device *mptcp_napi_dev; 51 52 /* Returns end sequence number of the receiver's advertised window */ 53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 54 { 55 return READ_ONCE(msk->wnd_end); 56 } 57 58 static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk) 59 { 60 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 61 if (sk->sk_prot == &tcpv6_prot) 62 return &inet6_stream_ops; 63 #endif 64 WARN_ON_ONCE(sk->sk_prot != &tcp_prot); 65 return &inet_stream_ops; 66 } 67 68 static int __mptcp_socket_create(struct mptcp_sock *msk) 69 { 70 struct mptcp_subflow_context *subflow; 71 struct sock *sk = (struct sock *)msk; 72 struct socket *ssock; 73 int err; 74 75 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock); 76 if (err) 77 return err; 78 79 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio; 80 WRITE_ONCE(msk->first, ssock->sk); 81 subflow = mptcp_subflow_ctx(ssock->sk); 82 list_add(&subflow->node, &msk->conn_list); 83 sock_hold(ssock->sk); 84 subflow->request_mptcp = 1; 85 subflow->subflow_id = msk->subflow_id++; 86 87 /* This is the first subflow, always with id 0 */ 88 WRITE_ONCE(subflow->local_id, 0); 89 mptcp_sock_graft(msk->first, sk->sk_socket); 90 iput(SOCK_INODE(ssock)); 91 92 return 0; 93 } 94 95 /* If the MPC handshake is not started, returns the first subflow, 96 * eventually allocating it. 97 */ 98 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk) 99 { 100 struct sock *sk = (struct sock *)msk; 101 int ret; 102 103 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 104 return ERR_PTR(-EINVAL); 105 106 if (!msk->first) { 107 ret = __mptcp_socket_create(msk); 108 if (ret) 109 return ERR_PTR(ret); 110 } 111 112 return msk->first; 113 } 114 115 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 116 { 117 sk_drops_add(sk, skb); 118 __kfree_skb(skb); 119 } 120 121 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 122 struct sk_buff *from) 123 { 124 bool fragstolen; 125 int delta; 126 127 if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) || 128 MPTCP_SKB_CB(from)->offset || 129 ((to->len + from->len) > (sk->sk_rcvbuf >> 3)) || 130 !skb_try_coalesce(to, from, &fragstolen, &delta)) 131 return false; 132 133 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n", 134 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 135 to->len, MPTCP_SKB_CB(from)->end_seq); 136 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 137 138 /* note the fwd memory can reach a negative value after accounting 139 * for the delta, but the later skb free will restore a non 140 * negative one 141 */ 142 atomic_add(delta, &sk->sk_rmem_alloc); 143 sk_mem_charge(sk, delta); 144 kfree_skb_partial(from, fragstolen); 145 146 return true; 147 } 148 149 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 150 struct sk_buff *from) 151 { 152 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 153 return false; 154 155 return mptcp_try_coalesce((struct sock *)msk, to, from); 156 } 157 158 /* "inspired" by tcp_data_queue_ofo(), main differences: 159 * - use mptcp seqs 160 * - don't cope with sacks 161 */ 162 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 163 { 164 struct sock *sk = (struct sock *)msk; 165 struct rb_node **p, *parent; 166 u64 seq, end_seq, max_seq; 167 struct sk_buff *skb1; 168 169 seq = MPTCP_SKB_CB(skb)->map_seq; 170 end_seq = MPTCP_SKB_CB(skb)->end_seq; 171 max_seq = atomic64_read(&msk->rcv_wnd_sent); 172 173 pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq, 174 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 175 if (after64(end_seq, max_seq)) { 176 /* out of window */ 177 mptcp_drop(sk, skb); 178 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 179 (unsigned long long)end_seq - (unsigned long)max_seq, 180 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent)); 181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 182 return; 183 } 184 185 p = &msk->out_of_order_queue.rb_node; 186 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 187 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 188 rb_link_node(&skb->rbnode, NULL, p); 189 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 190 msk->ooo_last_skb = skb; 191 goto end; 192 } 193 194 /* with 2 subflows, adding at end of ooo queue is quite likely 195 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 196 */ 197 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 198 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 199 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 200 return; 201 } 202 203 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 204 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 205 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 206 parent = &msk->ooo_last_skb->rbnode; 207 p = &parent->rb_right; 208 goto insert; 209 } 210 211 /* Find place to insert this segment. Handle overlaps on the way. */ 212 parent = NULL; 213 while (*p) { 214 parent = *p; 215 skb1 = rb_to_skb(parent); 216 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 217 p = &parent->rb_left; 218 continue; 219 } 220 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 221 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 222 /* All the bits are present. Drop. */ 223 mptcp_drop(sk, skb); 224 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 225 return; 226 } 227 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 228 /* partial overlap: 229 * | skb | 230 * | skb1 | 231 * continue traversing 232 */ 233 } else { 234 /* skb's seq == skb1's seq and skb covers skb1. 235 * Replace skb1 with skb. 236 */ 237 rb_replace_node(&skb1->rbnode, &skb->rbnode, 238 &msk->out_of_order_queue); 239 mptcp_drop(sk, skb1); 240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 241 goto merge_right; 242 } 243 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 245 return; 246 } 247 p = &parent->rb_right; 248 } 249 250 insert: 251 /* Insert segment into RB tree. */ 252 rb_link_node(&skb->rbnode, parent, p); 253 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 254 255 merge_right: 256 /* Remove other segments covered by skb. */ 257 while ((skb1 = skb_rb_next(skb)) != NULL) { 258 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 259 break; 260 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 261 mptcp_drop(sk, skb1); 262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 263 } 264 /* If there is no skb after us, we are the last_skb ! */ 265 if (!skb1) 266 msk->ooo_last_skb = skb; 267 268 end: 269 skb_condense(skb); 270 skb_set_owner_r(skb, sk); 271 } 272 273 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk, 274 struct sk_buff *skb, unsigned int offset, 275 size_t copy_len) 276 { 277 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 278 struct sock *sk = (struct sock *)msk; 279 struct sk_buff *tail; 280 bool has_rxtstamp; 281 282 __skb_unlink(skb, &ssk->sk_receive_queue); 283 284 skb_ext_reset(skb); 285 skb_orphan(skb); 286 287 /* try to fetch required memory from subflow */ 288 if (!sk_rmem_schedule(sk, skb, skb->truesize)) { 289 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 290 goto drop; 291 } 292 293 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 294 295 /* the skb map_seq accounts for the skb offset: 296 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 297 * value 298 */ 299 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 300 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 301 MPTCP_SKB_CB(skb)->offset = offset; 302 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 303 MPTCP_SKB_CB(skb)->cant_coalesce = 0; 304 305 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 306 /* in sequence */ 307 msk->bytes_received += copy_len; 308 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 309 tail = skb_peek_tail(&sk->sk_receive_queue); 310 if (tail && mptcp_try_coalesce(sk, tail, skb)) 311 return true; 312 313 skb_set_owner_r(skb, sk); 314 __skb_queue_tail(&sk->sk_receive_queue, skb); 315 return true; 316 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 317 mptcp_data_queue_ofo(msk, skb); 318 return false; 319 } 320 321 /* old data, keep it simple and drop the whole pkt, sender 322 * will retransmit as needed, if needed. 323 */ 324 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 325 drop: 326 mptcp_drop(sk, skb); 327 return false; 328 } 329 330 static void mptcp_stop_rtx_timer(struct sock *sk) 331 { 332 struct inet_connection_sock *icsk = inet_csk(sk); 333 334 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 335 mptcp_sk(sk)->timer_ival = 0; 336 } 337 338 static void mptcp_close_wake_up(struct sock *sk) 339 { 340 if (sock_flag(sk, SOCK_DEAD)) 341 return; 342 343 sk->sk_state_change(sk); 344 if (sk->sk_shutdown == SHUTDOWN_MASK || 345 sk->sk_state == TCP_CLOSE) 346 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 347 else 348 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 349 } 350 351 /* called under the msk socket lock */ 352 static bool mptcp_pending_data_fin_ack(struct sock *sk) 353 { 354 struct mptcp_sock *msk = mptcp_sk(sk); 355 356 return ((1 << sk->sk_state) & 357 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 358 msk->write_seq == READ_ONCE(msk->snd_una); 359 } 360 361 static void mptcp_check_data_fin_ack(struct sock *sk) 362 { 363 struct mptcp_sock *msk = mptcp_sk(sk); 364 365 /* Look for an acknowledged DATA_FIN */ 366 if (mptcp_pending_data_fin_ack(sk)) { 367 WRITE_ONCE(msk->snd_data_fin_enable, 0); 368 369 switch (sk->sk_state) { 370 case TCP_FIN_WAIT1: 371 mptcp_set_state(sk, TCP_FIN_WAIT2); 372 break; 373 case TCP_CLOSING: 374 case TCP_LAST_ACK: 375 mptcp_set_state(sk, TCP_CLOSE); 376 break; 377 } 378 379 mptcp_close_wake_up(sk); 380 } 381 } 382 383 /* can be called with no lock acquired */ 384 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 385 { 386 struct mptcp_sock *msk = mptcp_sk(sk); 387 388 if (READ_ONCE(msk->rcv_data_fin) && 389 ((1 << inet_sk_state_load(sk)) & 390 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 391 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 392 393 if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) { 394 if (seq) 395 *seq = rcv_data_fin_seq; 396 397 return true; 398 } 399 } 400 401 return false; 402 } 403 404 static void mptcp_set_datafin_timeout(struct sock *sk) 405 { 406 struct inet_connection_sock *icsk = inet_csk(sk); 407 u32 retransmits; 408 409 retransmits = min_t(u32, icsk->icsk_retransmits, 410 ilog2(TCP_RTO_MAX / TCP_RTO_MIN)); 411 412 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits; 413 } 414 415 static void __mptcp_set_timeout(struct sock *sk, long tout) 416 { 417 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 418 } 419 420 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 421 { 422 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 423 424 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 425 inet_csk(ssk)->icsk_timeout - jiffies : 0; 426 } 427 428 static void mptcp_set_timeout(struct sock *sk) 429 { 430 struct mptcp_subflow_context *subflow; 431 long tout = 0; 432 433 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 434 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 435 __mptcp_set_timeout(sk, tout); 436 } 437 438 static inline bool tcp_can_send_ack(const struct sock *ssk) 439 { 440 return !((1 << inet_sk_state_load(ssk)) & 441 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 442 } 443 444 void __mptcp_subflow_send_ack(struct sock *ssk) 445 { 446 if (tcp_can_send_ack(ssk)) 447 tcp_send_ack(ssk); 448 } 449 450 static void mptcp_subflow_send_ack(struct sock *ssk) 451 { 452 bool slow; 453 454 slow = lock_sock_fast(ssk); 455 __mptcp_subflow_send_ack(ssk); 456 unlock_sock_fast(ssk, slow); 457 } 458 459 static void mptcp_send_ack(struct mptcp_sock *msk) 460 { 461 struct mptcp_subflow_context *subflow; 462 463 mptcp_for_each_subflow(msk, subflow) 464 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow)); 465 } 466 467 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied) 468 { 469 bool slow; 470 471 slow = lock_sock_fast(ssk); 472 if (tcp_can_send_ack(ssk)) 473 tcp_cleanup_rbuf(ssk, copied); 474 unlock_sock_fast(ssk, slow); 475 } 476 477 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 478 { 479 const struct inet_connection_sock *icsk = inet_csk(ssk); 480 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 481 const struct tcp_sock *tp = tcp_sk(ssk); 482 483 return (ack_pending & ICSK_ACK_SCHED) && 484 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 485 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 486 (rx_empty && ack_pending & 487 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 488 } 489 490 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied) 491 { 492 int old_space = READ_ONCE(msk->old_wspace); 493 struct mptcp_subflow_context *subflow; 494 struct sock *sk = (struct sock *)msk; 495 int space = __mptcp_space(sk); 496 bool cleanup, rx_empty; 497 498 cleanup = (space > 0) && (space >= (old_space << 1)) && copied; 499 rx_empty = !sk_rmem_alloc_get(sk) && copied; 500 501 mptcp_for_each_subflow(msk, subflow) { 502 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 503 504 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 505 mptcp_subflow_cleanup_rbuf(ssk, copied); 506 } 507 } 508 509 static bool mptcp_check_data_fin(struct sock *sk) 510 { 511 struct mptcp_sock *msk = mptcp_sk(sk); 512 u64 rcv_data_fin_seq; 513 bool ret = false; 514 515 /* Need to ack a DATA_FIN received from a peer while this side 516 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 517 * msk->rcv_data_fin was set when parsing the incoming options 518 * at the subflow level and the msk lock was not held, so this 519 * is the first opportunity to act on the DATA_FIN and change 520 * the msk state. 521 * 522 * If we are caught up to the sequence number of the incoming 523 * DATA_FIN, send the DATA_ACK now and do state transition. If 524 * not caught up, do nothing and let the recv code send DATA_ACK 525 * when catching up. 526 */ 527 528 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 529 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 530 WRITE_ONCE(msk->rcv_data_fin, 0); 531 532 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 533 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 534 535 switch (sk->sk_state) { 536 case TCP_ESTABLISHED: 537 mptcp_set_state(sk, TCP_CLOSE_WAIT); 538 break; 539 case TCP_FIN_WAIT1: 540 mptcp_set_state(sk, TCP_CLOSING); 541 break; 542 case TCP_FIN_WAIT2: 543 mptcp_set_state(sk, TCP_CLOSE); 544 break; 545 default: 546 /* Other states not expected */ 547 WARN_ON_ONCE(1); 548 break; 549 } 550 551 ret = true; 552 if (!__mptcp_check_fallback(msk)) 553 mptcp_send_ack(msk); 554 mptcp_close_wake_up(sk); 555 } 556 return ret; 557 } 558 559 static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk) 560 { 561 if (READ_ONCE(msk->allow_infinite_fallback)) { 562 MPTCP_INC_STATS(sock_net(ssk), 563 MPTCP_MIB_DSSCORRUPTIONFALLBACK); 564 mptcp_do_fallback(ssk); 565 } else { 566 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET); 567 mptcp_subflow_reset(ssk); 568 } 569 } 570 571 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 572 struct sock *ssk, 573 unsigned int *bytes) 574 { 575 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 576 struct sock *sk = (struct sock *)msk; 577 unsigned int moved = 0; 578 bool more_data_avail; 579 struct tcp_sock *tp; 580 bool done = false; 581 582 pr_debug("msk=%p ssk=%p\n", msk, ssk); 583 tp = tcp_sk(ssk); 584 do { 585 u32 map_remaining, offset; 586 u32 seq = tp->copied_seq; 587 struct sk_buff *skb; 588 bool fin; 589 590 /* try to move as much data as available */ 591 map_remaining = subflow->map_data_len - 592 mptcp_subflow_get_map_offset(subflow); 593 594 skb = skb_peek(&ssk->sk_receive_queue); 595 if (!skb) { 596 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(), 597 * a different CPU can have already processed the pending 598 * data, stop here or we can enter an infinite loop 599 */ 600 if (!moved) 601 done = true; 602 break; 603 } 604 605 if (__mptcp_check_fallback(msk)) { 606 /* Under fallback skbs have no MPTCP extension and TCP could 607 * collapse them between the dummy map creation and the 608 * current dequeue. Be sure to adjust the map size. 609 */ 610 map_remaining = skb->len; 611 subflow->map_data_len = skb->len; 612 } 613 614 offset = seq - TCP_SKB_CB(skb)->seq; 615 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 616 if (fin) { 617 done = true; 618 seq++; 619 } 620 621 if (offset < skb->len) { 622 size_t len = skb->len - offset; 623 624 if (tp->urg_data) 625 done = true; 626 627 if (__mptcp_move_skb(msk, ssk, skb, offset, len)) 628 moved += len; 629 seq += len; 630 631 if (unlikely(map_remaining < len)) { 632 DEBUG_NET_WARN_ON_ONCE(1); 633 mptcp_dss_corruption(msk, ssk); 634 } 635 } else { 636 if (unlikely(!fin)) { 637 DEBUG_NET_WARN_ON_ONCE(1); 638 mptcp_dss_corruption(msk, ssk); 639 } 640 641 sk_eat_skb(ssk, skb); 642 done = true; 643 } 644 645 WRITE_ONCE(tp->copied_seq, seq); 646 more_data_avail = mptcp_subflow_data_available(ssk); 647 648 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) { 649 done = true; 650 break; 651 } 652 } while (more_data_avail); 653 654 if (moved > 0) 655 msk->last_data_recv = tcp_jiffies32; 656 *bytes += moved; 657 return done; 658 } 659 660 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 661 { 662 struct sock *sk = (struct sock *)msk; 663 struct sk_buff *skb, *tail; 664 bool moved = false; 665 struct rb_node *p; 666 u64 end_seq; 667 668 p = rb_first(&msk->out_of_order_queue); 669 pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 670 while (p) { 671 skb = rb_to_skb(p); 672 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 673 break; 674 675 p = rb_next(p); 676 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 677 678 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 679 msk->ack_seq))) { 680 mptcp_drop(sk, skb); 681 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 682 continue; 683 } 684 685 end_seq = MPTCP_SKB_CB(skb)->end_seq; 686 tail = skb_peek_tail(&sk->sk_receive_queue); 687 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 688 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 689 690 /* skip overlapping data, if any */ 691 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n", 692 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 693 delta); 694 MPTCP_SKB_CB(skb)->offset += delta; 695 MPTCP_SKB_CB(skb)->map_seq += delta; 696 __skb_queue_tail(&sk->sk_receive_queue, skb); 697 } 698 msk->bytes_received += end_seq - msk->ack_seq; 699 WRITE_ONCE(msk->ack_seq, end_seq); 700 moved = true; 701 } 702 return moved; 703 } 704 705 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk) 706 { 707 int err = sock_error(ssk); 708 int ssk_state; 709 710 if (!err) 711 return false; 712 713 /* only propagate errors on fallen-back sockets or 714 * on MPC connect 715 */ 716 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk))) 717 return false; 718 719 /* We need to propagate only transition to CLOSE state. 720 * Orphaned socket will see such state change via 721 * subflow_sched_work_if_closed() and that path will properly 722 * destroy the msk as needed. 723 */ 724 ssk_state = inet_sk_state_load(ssk); 725 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD)) 726 mptcp_set_state(sk, ssk_state); 727 WRITE_ONCE(sk->sk_err, -err); 728 729 /* This barrier is coupled with smp_rmb() in mptcp_poll() */ 730 smp_wmb(); 731 sk_error_report(sk); 732 return true; 733 } 734 735 void __mptcp_error_report(struct sock *sk) 736 { 737 struct mptcp_subflow_context *subflow; 738 struct mptcp_sock *msk = mptcp_sk(sk); 739 740 mptcp_for_each_subflow(msk, subflow) 741 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow))) 742 break; 743 } 744 745 /* In most cases we will be able to lock the mptcp socket. If its already 746 * owned, we need to defer to the work queue to avoid ABBA deadlock. 747 */ 748 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 749 { 750 struct sock *sk = (struct sock *)msk; 751 unsigned int moved = 0; 752 753 __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 754 __mptcp_ofo_queue(msk); 755 if (unlikely(ssk->sk_err)) { 756 if (!sock_owned_by_user(sk)) 757 __mptcp_error_report(sk); 758 else 759 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags); 760 } 761 762 /* If the moves have caught up with the DATA_FIN sequence number 763 * it's time to ack the DATA_FIN and change socket state, but 764 * this is not a good place to change state. Let the workqueue 765 * do it. 766 */ 767 if (mptcp_pending_data_fin(sk, NULL)) 768 mptcp_schedule_work(sk); 769 return moved > 0; 770 } 771 772 static void __mptcp_rcvbuf_update(struct sock *sk, struct sock *ssk) 773 { 774 if (unlikely(ssk->sk_rcvbuf > sk->sk_rcvbuf)) 775 WRITE_ONCE(sk->sk_rcvbuf, ssk->sk_rcvbuf); 776 } 777 778 static void __mptcp_data_ready(struct sock *sk, struct sock *ssk) 779 { 780 struct mptcp_sock *msk = mptcp_sk(sk); 781 782 __mptcp_rcvbuf_update(sk, ssk); 783 784 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/ 785 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 786 return; 787 788 /* Wake-up the reader only for in-sequence data */ 789 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk)) 790 sk->sk_data_ready(sk); 791 } 792 793 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 794 { 795 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 796 797 /* The peer can send data while we are shutting down this 798 * subflow at msk destruction time, but we must avoid enqueuing 799 * more data to the msk receive queue 800 */ 801 if (unlikely(subflow->disposable)) 802 return; 803 804 mptcp_data_lock(sk); 805 if (!sock_owned_by_user(sk)) 806 __mptcp_data_ready(sk, ssk); 807 else 808 __set_bit(MPTCP_DEQUEUE, &mptcp_sk(sk)->cb_flags); 809 mptcp_data_unlock(sk); 810 } 811 812 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk) 813 { 814 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq); 815 WRITE_ONCE(msk->allow_infinite_fallback, false); 816 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 817 } 818 819 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk) 820 { 821 struct sock *sk = (struct sock *)msk; 822 823 if (sk->sk_state != TCP_ESTABLISHED) 824 return false; 825 826 /* attach to msk socket only after we are sure we will deal with it 827 * at close time 828 */ 829 if (sk->sk_socket && !ssk->sk_socket) 830 mptcp_sock_graft(ssk, sk->sk_socket); 831 832 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++; 833 mptcp_sockopt_sync_locked(msk, ssk); 834 mptcp_subflow_joined(msk, ssk); 835 mptcp_stop_tout_timer(sk); 836 __mptcp_propagate_sndbuf(sk, ssk); 837 return true; 838 } 839 840 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list) 841 { 842 struct mptcp_subflow_context *tmp, *subflow; 843 struct mptcp_sock *msk = mptcp_sk(sk); 844 845 list_for_each_entry_safe(subflow, tmp, join_list, node) { 846 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 847 bool slow = lock_sock_fast(ssk); 848 849 list_move_tail(&subflow->node, &msk->conn_list); 850 if (!__mptcp_finish_join(msk, ssk)) 851 mptcp_subflow_reset(ssk); 852 unlock_sock_fast(ssk, slow); 853 } 854 } 855 856 static bool mptcp_rtx_timer_pending(struct sock *sk) 857 { 858 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 859 } 860 861 static void mptcp_reset_rtx_timer(struct sock *sk) 862 { 863 struct inet_connection_sock *icsk = inet_csk(sk); 864 unsigned long tout; 865 866 /* prevent rescheduling on close */ 867 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 868 return; 869 870 tout = mptcp_sk(sk)->timer_ival; 871 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 872 } 873 874 bool mptcp_schedule_work(struct sock *sk) 875 { 876 if (inet_sk_state_load(sk) != TCP_CLOSE && 877 schedule_work(&mptcp_sk(sk)->work)) { 878 /* each subflow already holds a reference to the sk, and the 879 * workqueue is invoked by a subflow, so sk can't go away here. 880 */ 881 sock_hold(sk); 882 return true; 883 } 884 return false; 885 } 886 887 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk) 888 { 889 struct mptcp_subflow_context *subflow; 890 891 msk_owned_by_me(msk); 892 893 mptcp_for_each_subflow(msk, subflow) { 894 if (READ_ONCE(subflow->data_avail)) 895 return mptcp_subflow_tcp_sock(subflow); 896 } 897 898 return NULL; 899 } 900 901 static bool mptcp_skb_can_collapse_to(u64 write_seq, 902 const struct sk_buff *skb, 903 const struct mptcp_ext *mpext) 904 { 905 if (!tcp_skb_can_collapse_to(skb)) 906 return false; 907 908 /* can collapse only if MPTCP level sequence is in order and this 909 * mapping has not been xmitted yet 910 */ 911 return mpext && mpext->data_seq + mpext->data_len == write_seq && 912 !mpext->frozen; 913 } 914 915 /* we can append data to the given data frag if: 916 * - there is space available in the backing page_frag 917 * - the data frag tail matches the current page_frag free offset 918 * - the data frag end sequence number matches the current write seq 919 */ 920 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 921 const struct page_frag *pfrag, 922 const struct mptcp_data_frag *df) 923 { 924 return df && pfrag->page == df->page && 925 pfrag->size - pfrag->offset > 0 && 926 pfrag->offset == (df->offset + df->data_len) && 927 df->data_seq + df->data_len == msk->write_seq; 928 } 929 930 static void dfrag_uncharge(struct sock *sk, int len) 931 { 932 sk_mem_uncharge(sk, len); 933 sk_wmem_queued_add(sk, -len); 934 } 935 936 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 937 { 938 int len = dfrag->data_len + dfrag->overhead; 939 940 list_del(&dfrag->list); 941 dfrag_uncharge(sk, len); 942 put_page(dfrag->page); 943 } 944 945 /* called under both the msk socket lock and the data lock */ 946 static void __mptcp_clean_una(struct sock *sk) 947 { 948 struct mptcp_sock *msk = mptcp_sk(sk); 949 struct mptcp_data_frag *dtmp, *dfrag; 950 u64 snd_una; 951 952 snd_una = msk->snd_una; 953 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 954 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 955 break; 956 957 if (unlikely(dfrag == msk->first_pending)) { 958 /* in recovery mode can see ack after the current snd head */ 959 if (WARN_ON_ONCE(!msk->recovery)) 960 break; 961 962 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 963 } 964 965 dfrag_clear(sk, dfrag); 966 } 967 968 dfrag = mptcp_rtx_head(sk); 969 if (dfrag && after64(snd_una, dfrag->data_seq)) { 970 u64 delta = snd_una - dfrag->data_seq; 971 972 /* prevent wrap around in recovery mode */ 973 if (unlikely(delta > dfrag->already_sent)) { 974 if (WARN_ON_ONCE(!msk->recovery)) 975 goto out; 976 if (WARN_ON_ONCE(delta > dfrag->data_len)) 977 goto out; 978 dfrag->already_sent += delta - dfrag->already_sent; 979 } 980 981 dfrag->data_seq += delta; 982 dfrag->offset += delta; 983 dfrag->data_len -= delta; 984 dfrag->already_sent -= delta; 985 986 dfrag_uncharge(sk, delta); 987 } 988 989 /* all retransmitted data acked, recovery completed */ 990 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 991 msk->recovery = false; 992 993 out: 994 if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) { 995 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 996 mptcp_stop_rtx_timer(sk); 997 } else { 998 mptcp_reset_rtx_timer(sk); 999 } 1000 1001 if (mptcp_pending_data_fin_ack(sk)) 1002 mptcp_schedule_work(sk); 1003 } 1004 1005 static void __mptcp_clean_una_wakeup(struct sock *sk) 1006 { 1007 lockdep_assert_held_once(&sk->sk_lock.slock); 1008 1009 __mptcp_clean_una(sk); 1010 mptcp_write_space(sk); 1011 } 1012 1013 static void mptcp_clean_una_wakeup(struct sock *sk) 1014 { 1015 mptcp_data_lock(sk); 1016 __mptcp_clean_una_wakeup(sk); 1017 mptcp_data_unlock(sk); 1018 } 1019 1020 static void mptcp_enter_memory_pressure(struct sock *sk) 1021 { 1022 struct mptcp_subflow_context *subflow; 1023 struct mptcp_sock *msk = mptcp_sk(sk); 1024 bool first = true; 1025 1026 mptcp_for_each_subflow(msk, subflow) { 1027 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1028 1029 if (first) 1030 tcp_enter_memory_pressure(ssk); 1031 sk_stream_moderate_sndbuf(ssk); 1032 1033 first = false; 1034 } 1035 __mptcp_sync_sndbuf(sk); 1036 } 1037 1038 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1039 * data 1040 */ 1041 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1042 { 1043 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1044 pfrag, sk->sk_allocation))) 1045 return true; 1046 1047 mptcp_enter_memory_pressure(sk); 1048 return false; 1049 } 1050 1051 static struct mptcp_data_frag * 1052 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1053 int orig_offset) 1054 { 1055 int offset = ALIGN(orig_offset, sizeof(long)); 1056 struct mptcp_data_frag *dfrag; 1057 1058 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1059 dfrag->data_len = 0; 1060 dfrag->data_seq = msk->write_seq; 1061 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1062 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1063 dfrag->already_sent = 0; 1064 dfrag->page = pfrag->page; 1065 1066 return dfrag; 1067 } 1068 1069 struct mptcp_sendmsg_info { 1070 int mss_now; 1071 int size_goal; 1072 u16 limit; 1073 u16 sent; 1074 unsigned int flags; 1075 bool data_lock_held; 1076 }; 1077 1078 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk, 1079 u64 data_seq, int avail_size) 1080 { 1081 u64 window_end = mptcp_wnd_end(msk); 1082 u64 mptcp_snd_wnd; 1083 1084 if (__mptcp_check_fallback(msk)) 1085 return avail_size; 1086 1087 mptcp_snd_wnd = window_end - data_seq; 1088 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size); 1089 1090 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) { 1091 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd); 1092 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED); 1093 } 1094 1095 return avail_size; 1096 } 1097 1098 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1099 { 1100 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1101 1102 if (!mpext) 1103 return false; 1104 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1105 return true; 1106 } 1107 1108 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1109 { 1110 struct sk_buff *skb; 1111 1112 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1113 if (likely(skb)) { 1114 if (likely(__mptcp_add_ext(skb, gfp))) { 1115 skb_reserve(skb, MAX_TCP_HEADER); 1116 skb->ip_summed = CHECKSUM_PARTIAL; 1117 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1118 return skb; 1119 } 1120 __kfree_skb(skb); 1121 } else { 1122 mptcp_enter_memory_pressure(sk); 1123 } 1124 return NULL; 1125 } 1126 1127 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1128 { 1129 struct sk_buff *skb; 1130 1131 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1132 if (!skb) 1133 return NULL; 1134 1135 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1136 tcp_skb_entail(ssk, skb); 1137 return skb; 1138 } 1139 tcp_skb_tsorted_anchor_cleanup(skb); 1140 kfree_skb(skb); 1141 return NULL; 1142 } 1143 1144 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1145 { 1146 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1147 1148 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1149 } 1150 1151 /* note: this always recompute the csum on the whole skb, even 1152 * if we just appended a single frag. More status info needed 1153 */ 1154 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1155 { 1156 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1157 __wsum csum = ~csum_unfold(mpext->csum); 1158 int offset = skb->len - added; 1159 1160 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1161 } 1162 1163 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1164 struct sock *ssk, 1165 struct mptcp_ext *mpext) 1166 { 1167 if (!mpext) 1168 return; 1169 1170 mpext->infinite_map = 1; 1171 mpext->data_len = 0; 1172 1173 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX); 1174 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1175 pr_fallback(msk); 1176 mptcp_do_fallback(ssk); 1177 } 1178 1179 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1)) 1180 1181 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1182 struct mptcp_data_frag *dfrag, 1183 struct mptcp_sendmsg_info *info) 1184 { 1185 u64 data_seq = dfrag->data_seq + info->sent; 1186 int offset = dfrag->offset + info->sent; 1187 struct mptcp_sock *msk = mptcp_sk(sk); 1188 bool zero_window_probe = false; 1189 struct mptcp_ext *mpext = NULL; 1190 bool can_coalesce = false; 1191 bool reuse_skb = true; 1192 struct sk_buff *skb; 1193 size_t copy; 1194 int i; 1195 1196 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n", 1197 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1198 1199 if (WARN_ON_ONCE(info->sent > info->limit || 1200 info->limit > dfrag->data_len)) 1201 return 0; 1202 1203 if (unlikely(!__tcp_can_send(ssk))) 1204 return -EAGAIN; 1205 1206 /* compute send limit */ 1207 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE)) 1208 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE; 1209 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1210 copy = info->size_goal; 1211 1212 skb = tcp_write_queue_tail(ssk); 1213 if (skb && copy > skb->len) { 1214 /* Limit the write to the size available in the 1215 * current skb, if any, so that we create at most a new skb. 1216 * Explicitly tells TCP internals to avoid collapsing on later 1217 * queue management operation, to avoid breaking the ext <-> 1218 * SSN association set here 1219 */ 1220 mpext = mptcp_get_ext(skb); 1221 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1222 TCP_SKB_CB(skb)->eor = 1; 1223 tcp_mark_push(tcp_sk(ssk), skb); 1224 goto alloc_skb; 1225 } 1226 1227 i = skb_shinfo(skb)->nr_frags; 1228 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1229 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) { 1230 tcp_mark_push(tcp_sk(ssk), skb); 1231 goto alloc_skb; 1232 } 1233 1234 copy -= skb->len; 1235 } else { 1236 alloc_skb: 1237 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1238 if (!skb) 1239 return -ENOMEM; 1240 1241 i = skb_shinfo(skb)->nr_frags; 1242 reuse_skb = false; 1243 mpext = mptcp_get_ext(skb); 1244 } 1245 1246 /* Zero window and all data acked? Probe. */ 1247 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy); 1248 if (copy == 0) { 1249 u64 snd_una = READ_ONCE(msk->snd_una); 1250 1251 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) { 1252 tcp_remove_empty_skb(ssk); 1253 return 0; 1254 } 1255 1256 zero_window_probe = true; 1257 data_seq = snd_una - 1; 1258 copy = 1; 1259 } 1260 1261 copy = min_t(size_t, copy, info->limit - info->sent); 1262 if (!sk_wmem_schedule(ssk, copy)) { 1263 tcp_remove_empty_skb(ssk); 1264 return -ENOMEM; 1265 } 1266 1267 if (can_coalesce) { 1268 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1269 } else { 1270 get_page(dfrag->page); 1271 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1272 } 1273 1274 skb->len += copy; 1275 skb->data_len += copy; 1276 skb->truesize += copy; 1277 sk_wmem_queued_add(ssk, copy); 1278 sk_mem_charge(ssk, copy); 1279 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1280 TCP_SKB_CB(skb)->end_seq += copy; 1281 tcp_skb_pcount_set(skb, 0); 1282 1283 /* on skb reuse we just need to update the DSS len */ 1284 if (reuse_skb) { 1285 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1286 mpext->data_len += copy; 1287 goto out; 1288 } 1289 1290 memset(mpext, 0, sizeof(*mpext)); 1291 mpext->data_seq = data_seq; 1292 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1293 mpext->data_len = copy; 1294 mpext->use_map = 1; 1295 mpext->dsn64 = 1; 1296 1297 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n", 1298 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1299 mpext->dsn64); 1300 1301 if (zero_window_probe) { 1302 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1303 mpext->frozen = 1; 1304 if (READ_ONCE(msk->csum_enabled)) 1305 mptcp_update_data_checksum(skb, copy); 1306 tcp_push_pending_frames(ssk); 1307 return 0; 1308 } 1309 out: 1310 if (READ_ONCE(msk->csum_enabled)) 1311 mptcp_update_data_checksum(skb, copy); 1312 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1313 mptcp_update_infinite_map(msk, ssk, mpext); 1314 trace_mptcp_sendmsg_frag(mpext); 1315 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1316 return copy; 1317 } 1318 1319 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1320 sizeof(struct tcphdr) - \ 1321 MAX_TCP_OPTION_SPACE - \ 1322 sizeof(struct ipv6hdr) - \ 1323 sizeof(struct frag_hdr)) 1324 1325 struct subflow_send_info { 1326 struct sock *ssk; 1327 u64 linger_time; 1328 }; 1329 1330 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1331 { 1332 if (!subflow->stale) 1333 return; 1334 1335 subflow->stale = 0; 1336 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1337 } 1338 1339 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1340 { 1341 if (unlikely(subflow->stale)) { 1342 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1343 1344 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1345 return false; 1346 1347 mptcp_subflow_set_active(subflow); 1348 } 1349 return __mptcp_subflow_active(subflow); 1350 } 1351 1352 #define SSK_MODE_ACTIVE 0 1353 #define SSK_MODE_BACKUP 1 1354 #define SSK_MODE_MAX 2 1355 1356 /* implement the mptcp packet scheduler; 1357 * returns the subflow that will transmit the next DSS 1358 * additionally updates the rtx timeout 1359 */ 1360 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1361 { 1362 struct subflow_send_info send_info[SSK_MODE_MAX]; 1363 struct mptcp_subflow_context *subflow; 1364 struct sock *sk = (struct sock *)msk; 1365 u32 pace, burst, wmem; 1366 int i, nr_active = 0; 1367 struct sock *ssk; 1368 u64 linger_time; 1369 long tout = 0; 1370 1371 /* pick the subflow with the lower wmem/wspace ratio */ 1372 for (i = 0; i < SSK_MODE_MAX; ++i) { 1373 send_info[i].ssk = NULL; 1374 send_info[i].linger_time = -1; 1375 } 1376 1377 mptcp_for_each_subflow(msk, subflow) { 1378 bool backup = subflow->backup || subflow->request_bkup; 1379 1380 trace_mptcp_subflow_get_send(subflow); 1381 ssk = mptcp_subflow_tcp_sock(subflow); 1382 if (!mptcp_subflow_active(subflow)) 1383 continue; 1384 1385 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1386 nr_active += !backup; 1387 pace = subflow->avg_pacing_rate; 1388 if (unlikely(!pace)) { 1389 /* init pacing rate from socket */ 1390 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1391 pace = subflow->avg_pacing_rate; 1392 if (!pace) 1393 continue; 1394 } 1395 1396 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1397 if (linger_time < send_info[backup].linger_time) { 1398 send_info[backup].ssk = ssk; 1399 send_info[backup].linger_time = linger_time; 1400 } 1401 } 1402 __mptcp_set_timeout(sk, tout); 1403 1404 /* pick the best backup if no other subflow is active */ 1405 if (!nr_active) 1406 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1407 1408 /* According to the blest algorithm, to avoid HoL blocking for the 1409 * faster flow, we need to: 1410 * - estimate the faster flow linger time 1411 * - use the above to estimate the amount of byte transferred 1412 * by the faster flow 1413 * - check that the amount of queued data is greter than the above, 1414 * otherwise do not use the picked, slower, subflow 1415 * We select the subflow with the shorter estimated time to flush 1416 * the queued mem, which basically ensure the above. We just need 1417 * to check that subflow has a non empty cwin. 1418 */ 1419 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1420 if (!ssk || !sk_stream_memory_free(ssk)) 1421 return NULL; 1422 1423 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt); 1424 wmem = READ_ONCE(ssk->sk_wmem_queued); 1425 if (!burst) 1426 return ssk; 1427 1428 subflow = mptcp_subflow_ctx(ssk); 1429 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1430 READ_ONCE(ssk->sk_pacing_rate) * burst, 1431 burst + wmem); 1432 msk->snd_burst = burst; 1433 return ssk; 1434 } 1435 1436 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1437 { 1438 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1439 release_sock(ssk); 1440 } 1441 1442 static void mptcp_update_post_push(struct mptcp_sock *msk, 1443 struct mptcp_data_frag *dfrag, 1444 u32 sent) 1445 { 1446 u64 snd_nxt_new = dfrag->data_seq; 1447 1448 dfrag->already_sent += sent; 1449 1450 msk->snd_burst -= sent; 1451 1452 snd_nxt_new += dfrag->already_sent; 1453 1454 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1455 * is recovering after a failover. In that event, this re-sends 1456 * old segments. 1457 * 1458 * Thus compute snd_nxt_new candidate based on 1459 * the dfrag->data_seq that was sent and the data 1460 * that has been handed to the subflow for transmission 1461 * and skip update in case it was old dfrag. 1462 */ 1463 if (likely(after64(snd_nxt_new, msk->snd_nxt))) { 1464 msk->bytes_sent += snd_nxt_new - msk->snd_nxt; 1465 WRITE_ONCE(msk->snd_nxt, snd_nxt_new); 1466 } 1467 } 1468 1469 void mptcp_check_and_set_pending(struct sock *sk) 1470 { 1471 if (mptcp_send_head(sk)) { 1472 mptcp_data_lock(sk); 1473 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING); 1474 mptcp_data_unlock(sk); 1475 } 1476 } 1477 1478 static int __subflow_push_pending(struct sock *sk, struct sock *ssk, 1479 struct mptcp_sendmsg_info *info) 1480 { 1481 struct mptcp_sock *msk = mptcp_sk(sk); 1482 struct mptcp_data_frag *dfrag; 1483 int len, copied = 0, err = 0; 1484 1485 while ((dfrag = mptcp_send_head(sk))) { 1486 info->sent = dfrag->already_sent; 1487 info->limit = dfrag->data_len; 1488 len = dfrag->data_len - dfrag->already_sent; 1489 while (len > 0) { 1490 int ret = 0; 1491 1492 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info); 1493 if (ret <= 0) { 1494 err = copied ? : ret; 1495 goto out; 1496 } 1497 1498 info->sent += ret; 1499 copied += ret; 1500 len -= ret; 1501 1502 mptcp_update_post_push(msk, dfrag, ret); 1503 } 1504 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk)); 1505 1506 if (msk->snd_burst <= 0 || 1507 !sk_stream_memory_free(ssk) || 1508 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) { 1509 err = copied; 1510 goto out; 1511 } 1512 mptcp_set_timeout(sk); 1513 } 1514 err = copied; 1515 1516 out: 1517 if (err > 0) 1518 msk->last_data_sent = tcp_jiffies32; 1519 return err; 1520 } 1521 1522 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1523 { 1524 struct sock *prev_ssk = NULL, *ssk = NULL; 1525 struct mptcp_sock *msk = mptcp_sk(sk); 1526 struct mptcp_sendmsg_info info = { 1527 .flags = flags, 1528 }; 1529 bool do_check_data_fin = false; 1530 int push_count = 1; 1531 1532 while (mptcp_send_head(sk) && (push_count > 0)) { 1533 struct mptcp_subflow_context *subflow; 1534 int ret = 0; 1535 1536 if (mptcp_sched_get_send(msk)) 1537 break; 1538 1539 push_count = 0; 1540 1541 mptcp_for_each_subflow(msk, subflow) { 1542 if (READ_ONCE(subflow->scheduled)) { 1543 mptcp_subflow_set_scheduled(subflow, false); 1544 1545 prev_ssk = ssk; 1546 ssk = mptcp_subflow_tcp_sock(subflow); 1547 if (ssk != prev_ssk) { 1548 /* First check. If the ssk has changed since 1549 * the last round, release prev_ssk 1550 */ 1551 if (prev_ssk) 1552 mptcp_push_release(prev_ssk, &info); 1553 1554 /* Need to lock the new subflow only if different 1555 * from the previous one, otherwise we are still 1556 * helding the relevant lock 1557 */ 1558 lock_sock(ssk); 1559 } 1560 1561 push_count++; 1562 1563 ret = __subflow_push_pending(sk, ssk, &info); 1564 if (ret <= 0) { 1565 if (ret != -EAGAIN || 1566 (1 << ssk->sk_state) & 1567 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE)) 1568 push_count--; 1569 continue; 1570 } 1571 do_check_data_fin = true; 1572 } 1573 } 1574 } 1575 1576 /* at this point we held the socket lock for the last subflow we used */ 1577 if (ssk) 1578 mptcp_push_release(ssk, &info); 1579 1580 /* ensure the rtx timer is running */ 1581 if (!mptcp_rtx_timer_pending(sk)) 1582 mptcp_reset_rtx_timer(sk); 1583 if (do_check_data_fin) 1584 mptcp_check_send_data_fin(sk); 1585 } 1586 1587 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first) 1588 { 1589 struct mptcp_sock *msk = mptcp_sk(sk); 1590 struct mptcp_sendmsg_info info = { 1591 .data_lock_held = true, 1592 }; 1593 bool keep_pushing = true; 1594 struct sock *xmit_ssk; 1595 int copied = 0; 1596 1597 info.flags = 0; 1598 while (mptcp_send_head(sk) && keep_pushing) { 1599 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1600 int ret = 0; 1601 1602 /* check for a different subflow usage only after 1603 * spooling the first chunk of data 1604 */ 1605 if (first) { 1606 mptcp_subflow_set_scheduled(subflow, false); 1607 ret = __subflow_push_pending(sk, ssk, &info); 1608 first = false; 1609 if (ret <= 0) 1610 break; 1611 copied += ret; 1612 continue; 1613 } 1614 1615 if (mptcp_sched_get_send(msk)) 1616 goto out; 1617 1618 if (READ_ONCE(subflow->scheduled)) { 1619 mptcp_subflow_set_scheduled(subflow, false); 1620 ret = __subflow_push_pending(sk, ssk, &info); 1621 if (ret <= 0) 1622 keep_pushing = false; 1623 copied += ret; 1624 } 1625 1626 mptcp_for_each_subflow(msk, subflow) { 1627 if (READ_ONCE(subflow->scheduled)) { 1628 xmit_ssk = mptcp_subflow_tcp_sock(subflow); 1629 if (xmit_ssk != ssk) { 1630 mptcp_subflow_delegate(subflow, 1631 MPTCP_DELEGATE_SEND); 1632 keep_pushing = false; 1633 } 1634 } 1635 } 1636 } 1637 1638 out: 1639 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1640 * not going to flush it via release_sock() 1641 */ 1642 if (copied) { 1643 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1644 info.size_goal); 1645 if (!mptcp_rtx_timer_pending(sk)) 1646 mptcp_reset_rtx_timer(sk); 1647 1648 if (msk->snd_data_fin_enable && 1649 msk->snd_nxt + 1 == msk->write_seq) 1650 mptcp_schedule_work(sk); 1651 } 1652 } 1653 1654 static int mptcp_disconnect(struct sock *sk, int flags); 1655 1656 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1657 size_t len, int *copied_syn) 1658 { 1659 unsigned int saved_flags = msg->msg_flags; 1660 struct mptcp_sock *msk = mptcp_sk(sk); 1661 struct sock *ssk; 1662 int ret; 1663 1664 /* on flags based fastopen the mptcp is supposed to create the 1665 * first subflow right now. Otherwise we are in the defer_connect 1666 * path, and the first subflow must be already present. 1667 * Since the defer_connect flag is cleared after the first succsful 1668 * fastopen attempt, no need to check for additional subflow status. 1669 */ 1670 if (msg->msg_flags & MSG_FASTOPEN) { 1671 ssk = __mptcp_nmpc_sk(msk); 1672 if (IS_ERR(ssk)) 1673 return PTR_ERR(ssk); 1674 } 1675 if (!msk->first) 1676 return -EINVAL; 1677 1678 ssk = msk->first; 1679 1680 lock_sock(ssk); 1681 msg->msg_flags |= MSG_DONTWAIT; 1682 msk->fastopening = 1; 1683 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL); 1684 msk->fastopening = 0; 1685 msg->msg_flags = saved_flags; 1686 release_sock(ssk); 1687 1688 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */ 1689 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) { 1690 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1691 msg->msg_namelen, msg->msg_flags, 1); 1692 1693 /* Keep the same behaviour of plain TCP: zero the copied bytes in 1694 * case of any error, except timeout or signal 1695 */ 1696 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR) 1697 *copied_syn = 0; 1698 } else if (ret && ret != -EINPROGRESS) { 1699 /* The disconnect() op called by tcp_sendmsg_fastopen()/ 1700 * __inet_stream_connect() can fail, due to looking check, 1701 * see mptcp_disconnect(). 1702 * Attempt it again outside the problematic scope. 1703 */ 1704 if (!mptcp_disconnect(sk, 0)) { 1705 sk->sk_disconnects++; 1706 sk->sk_socket->state = SS_UNCONNECTED; 1707 } 1708 } 1709 inet_clear_bit(DEFER_CONNECT, sk); 1710 1711 return ret; 1712 } 1713 1714 static int do_copy_data_nocache(struct sock *sk, int copy, 1715 struct iov_iter *from, char *to) 1716 { 1717 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) { 1718 if (!copy_from_iter_full_nocache(to, copy, from)) 1719 return -EFAULT; 1720 } else if (!copy_from_iter_full(to, copy, from)) { 1721 return -EFAULT; 1722 } 1723 return 0; 1724 } 1725 1726 /* open-code sk_stream_memory_free() plus sent limit computation to 1727 * avoid indirect calls in fast-path. 1728 * Called under the msk socket lock, so we can avoid a bunch of ONCE 1729 * annotations. 1730 */ 1731 static u32 mptcp_send_limit(const struct sock *sk) 1732 { 1733 const struct mptcp_sock *msk = mptcp_sk(sk); 1734 u32 limit, not_sent; 1735 1736 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf)) 1737 return 0; 1738 1739 limit = mptcp_notsent_lowat(sk); 1740 if (limit == UINT_MAX) 1741 return UINT_MAX; 1742 1743 not_sent = msk->write_seq - msk->snd_nxt; 1744 if (not_sent >= limit) 1745 return 0; 1746 1747 return limit - not_sent; 1748 } 1749 1750 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1751 { 1752 struct mptcp_sock *msk = mptcp_sk(sk); 1753 struct page_frag *pfrag; 1754 size_t copied = 0; 1755 int ret = 0; 1756 long timeo; 1757 1758 /* silently ignore everything else */ 1759 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN; 1760 1761 lock_sock(sk); 1762 1763 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) || 1764 msg->msg_flags & MSG_FASTOPEN)) { 1765 int copied_syn = 0; 1766 1767 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn); 1768 copied += copied_syn; 1769 if (ret == -EINPROGRESS && copied_syn > 0) 1770 goto out; 1771 else if (ret) 1772 goto do_error; 1773 } 1774 1775 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1776 1777 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1778 ret = sk_stream_wait_connect(sk, &timeo); 1779 if (ret) 1780 goto do_error; 1781 } 1782 1783 ret = -EPIPE; 1784 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))) 1785 goto do_error; 1786 1787 pfrag = sk_page_frag(sk); 1788 1789 while (msg_data_left(msg)) { 1790 int total_ts, frag_truesize = 0; 1791 struct mptcp_data_frag *dfrag; 1792 bool dfrag_collapsed; 1793 size_t psize, offset; 1794 u32 copy_limit; 1795 1796 /* ensure fitting the notsent_lowat() constraint */ 1797 copy_limit = mptcp_send_limit(sk); 1798 if (!copy_limit) 1799 goto wait_for_memory; 1800 1801 /* reuse tail pfrag, if possible, or carve a new one from the 1802 * page allocator 1803 */ 1804 dfrag = mptcp_pending_tail(sk); 1805 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1806 if (!dfrag_collapsed) { 1807 if (!mptcp_page_frag_refill(sk, pfrag)) 1808 goto wait_for_memory; 1809 1810 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1811 frag_truesize = dfrag->overhead; 1812 } 1813 1814 /* we do not bound vs wspace, to allow a single packet. 1815 * memory accounting will prevent execessive memory usage 1816 * anyway 1817 */ 1818 offset = dfrag->offset + dfrag->data_len; 1819 psize = pfrag->size - offset; 1820 psize = min_t(size_t, psize, msg_data_left(msg)); 1821 psize = min_t(size_t, psize, copy_limit); 1822 total_ts = psize + frag_truesize; 1823 1824 if (!sk_wmem_schedule(sk, total_ts)) 1825 goto wait_for_memory; 1826 1827 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter, 1828 page_address(dfrag->page) + offset); 1829 if (ret) 1830 goto do_error; 1831 1832 /* data successfully copied into the write queue */ 1833 sk_forward_alloc_add(sk, -total_ts); 1834 copied += psize; 1835 dfrag->data_len += psize; 1836 frag_truesize += psize; 1837 pfrag->offset += frag_truesize; 1838 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1839 1840 /* charge data on mptcp pending queue to the msk socket 1841 * Note: we charge such data both to sk and ssk 1842 */ 1843 sk_wmem_queued_add(sk, frag_truesize); 1844 if (!dfrag_collapsed) { 1845 get_page(dfrag->page); 1846 list_add_tail(&dfrag->list, &msk->rtx_queue); 1847 if (!msk->first_pending) 1848 WRITE_ONCE(msk->first_pending, dfrag); 1849 } 1850 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk, 1851 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1852 !dfrag_collapsed); 1853 1854 continue; 1855 1856 wait_for_memory: 1857 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1858 __mptcp_push_pending(sk, msg->msg_flags); 1859 ret = sk_stream_wait_memory(sk, &timeo); 1860 if (ret) 1861 goto do_error; 1862 } 1863 1864 if (copied) 1865 __mptcp_push_pending(sk, msg->msg_flags); 1866 1867 out: 1868 release_sock(sk); 1869 return copied; 1870 1871 do_error: 1872 if (copied) 1873 goto out; 1874 1875 copied = sk_stream_error(sk, msg->msg_flags, ret); 1876 goto out; 1877 } 1878 1879 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied); 1880 1881 static int __mptcp_recvmsg_mskq(struct sock *sk, 1882 struct msghdr *msg, 1883 size_t len, int flags, 1884 struct scm_timestamping_internal *tss, 1885 int *cmsg_flags) 1886 { 1887 struct mptcp_sock *msk = mptcp_sk(sk); 1888 struct sk_buff *skb, *tmp; 1889 int copied = 0; 1890 1891 skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) { 1892 u32 offset = MPTCP_SKB_CB(skb)->offset; 1893 u32 data_len = skb->len - offset; 1894 u32 count = min_t(size_t, len - copied, data_len); 1895 int err; 1896 1897 if (!(flags & MSG_TRUNC)) { 1898 err = skb_copy_datagram_msg(skb, offset, msg, count); 1899 if (unlikely(err < 0)) { 1900 if (!copied) 1901 return err; 1902 break; 1903 } 1904 } 1905 1906 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 1907 tcp_update_recv_tstamps(skb, tss); 1908 *cmsg_flags |= MPTCP_CMSG_TS; 1909 } 1910 1911 copied += count; 1912 1913 if (count < data_len) { 1914 if (!(flags & MSG_PEEK)) { 1915 MPTCP_SKB_CB(skb)->offset += count; 1916 MPTCP_SKB_CB(skb)->map_seq += count; 1917 msk->bytes_consumed += count; 1918 } 1919 break; 1920 } 1921 1922 if (!(flags & MSG_PEEK)) { 1923 /* avoid the indirect call, we know the destructor is sock_wfree */ 1924 skb->destructor = NULL; 1925 atomic_sub(skb->truesize, &sk->sk_rmem_alloc); 1926 sk_mem_uncharge(sk, skb->truesize); 1927 __skb_unlink(skb, &sk->sk_receive_queue); 1928 __kfree_skb(skb); 1929 msk->bytes_consumed += count; 1930 } 1931 1932 if (copied >= len) 1933 break; 1934 } 1935 1936 mptcp_rcv_space_adjust(msk, copied); 1937 return copied; 1938 } 1939 1940 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 1941 * 1942 * Only difference: Use highest rtt estimate of the subflows in use. 1943 */ 1944 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 1945 { 1946 struct mptcp_subflow_context *subflow; 1947 struct sock *sk = (struct sock *)msk; 1948 u8 scaling_ratio = U8_MAX; 1949 u32 time, advmss = 1; 1950 u64 rtt_us, mstamp; 1951 1952 msk_owned_by_me(msk); 1953 1954 if (copied <= 0) 1955 return; 1956 1957 if (!msk->rcvspace_init) 1958 mptcp_rcv_space_init(msk, msk->first); 1959 1960 msk->rcvq_space.copied += copied; 1961 1962 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 1963 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 1964 1965 rtt_us = msk->rcvq_space.rtt_us; 1966 if (rtt_us && time < (rtt_us >> 3)) 1967 return; 1968 1969 rtt_us = 0; 1970 mptcp_for_each_subflow(msk, subflow) { 1971 const struct tcp_sock *tp; 1972 u64 sf_rtt_us; 1973 u32 sf_advmss; 1974 1975 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 1976 1977 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 1978 sf_advmss = READ_ONCE(tp->advmss); 1979 1980 rtt_us = max(sf_rtt_us, rtt_us); 1981 advmss = max(sf_advmss, advmss); 1982 scaling_ratio = min(tp->scaling_ratio, scaling_ratio); 1983 } 1984 1985 msk->rcvq_space.rtt_us = rtt_us; 1986 msk->scaling_ratio = scaling_ratio; 1987 if (time < (rtt_us >> 3) || rtt_us == 0) 1988 return; 1989 1990 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 1991 goto new_measure; 1992 1993 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 1994 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 1995 u64 rcvwin, grow; 1996 int rcvbuf; 1997 1998 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss; 1999 2000 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space); 2001 2002 do_div(grow, msk->rcvq_space.space); 2003 rcvwin += (grow << 1); 2004 2005 rcvbuf = min_t(u64, mptcp_space_from_win(sk, rcvwin), 2006 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 2007 2008 if (rcvbuf > sk->sk_rcvbuf) { 2009 u32 window_clamp; 2010 2011 window_clamp = mptcp_win_from_space(sk, rcvbuf); 2012 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 2013 2014 /* Make subflows follow along. If we do not do this, we 2015 * get drops at subflow level if skbs can't be moved to 2016 * the mptcp rx queue fast enough (announced rcv_win can 2017 * exceed ssk->sk_rcvbuf). 2018 */ 2019 mptcp_for_each_subflow(msk, subflow) { 2020 struct sock *ssk; 2021 bool slow; 2022 2023 ssk = mptcp_subflow_tcp_sock(subflow); 2024 slow = lock_sock_fast(ssk); 2025 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf); 2026 WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp); 2027 if (tcp_can_send_ack(ssk)) 2028 tcp_cleanup_rbuf(ssk, 1); 2029 unlock_sock_fast(ssk, slow); 2030 } 2031 } 2032 } 2033 2034 msk->rcvq_space.space = msk->rcvq_space.copied; 2035 new_measure: 2036 msk->rcvq_space.copied = 0; 2037 msk->rcvq_space.time = mstamp; 2038 } 2039 2040 static bool __mptcp_move_skbs(struct sock *sk) 2041 { 2042 struct mptcp_subflow_context *subflow; 2043 struct mptcp_sock *msk = mptcp_sk(sk); 2044 unsigned int moved = 0; 2045 bool ret, done; 2046 2047 /* verify we can move any data from the subflow, eventually updating */ 2048 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 2049 mptcp_for_each_subflow(msk, subflow) 2050 __mptcp_rcvbuf_update(sk, subflow->tcp_sock); 2051 2052 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 2053 return false; 2054 2055 do { 2056 struct sock *ssk = mptcp_subflow_recv_lookup(msk); 2057 bool slowpath; 2058 2059 if (unlikely(!ssk)) 2060 break; 2061 2062 slowpath = lock_sock_fast(ssk); 2063 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved); 2064 2065 if (unlikely(ssk->sk_err)) 2066 __mptcp_error_report(sk); 2067 unlock_sock_fast(ssk, slowpath); 2068 } while (!done); 2069 2070 ret = moved > 0 || __mptcp_ofo_queue(msk); 2071 if (ret) 2072 mptcp_check_data_fin((struct sock *)msk); 2073 return ret; 2074 } 2075 2076 static unsigned int mptcp_inq_hint(const struct sock *sk) 2077 { 2078 const struct mptcp_sock *msk = mptcp_sk(sk); 2079 const struct sk_buff *skb; 2080 2081 skb = skb_peek(&sk->sk_receive_queue); 2082 if (skb) { 2083 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq; 2084 2085 if (hint_val >= INT_MAX) 2086 return INT_MAX; 2087 2088 return (unsigned int)hint_val; 2089 } 2090 2091 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2092 return 1; 2093 2094 return 0; 2095 } 2096 2097 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2098 int flags, int *addr_len) 2099 { 2100 struct mptcp_sock *msk = mptcp_sk(sk); 2101 struct scm_timestamping_internal tss; 2102 int copied = 0, cmsg_flags = 0; 2103 int target; 2104 long timeo; 2105 2106 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2107 if (unlikely(flags & MSG_ERRQUEUE)) 2108 return inet_recv_error(sk, msg, len, addr_len); 2109 2110 lock_sock(sk); 2111 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2112 copied = -ENOTCONN; 2113 goto out_err; 2114 } 2115 2116 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2117 2118 len = min_t(size_t, len, INT_MAX); 2119 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2120 2121 if (unlikely(msk->recvmsg_inq)) 2122 cmsg_flags = MPTCP_CMSG_INQ; 2123 2124 while (copied < len) { 2125 int err, bytes_read; 2126 2127 bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags, &tss, &cmsg_flags); 2128 if (unlikely(bytes_read < 0)) { 2129 if (!copied) 2130 copied = bytes_read; 2131 goto out_err; 2132 } 2133 2134 copied += bytes_read; 2135 2136 if (skb_queue_empty(&sk->sk_receive_queue) && __mptcp_move_skbs(sk)) 2137 continue; 2138 2139 /* only the MPTCP socket status is relevant here. The exit 2140 * conditions mirror closely tcp_recvmsg() 2141 */ 2142 if (copied >= target) 2143 break; 2144 2145 if (copied) { 2146 if (sk->sk_err || 2147 sk->sk_state == TCP_CLOSE || 2148 (sk->sk_shutdown & RCV_SHUTDOWN) || 2149 !timeo || 2150 signal_pending(current)) 2151 break; 2152 } else { 2153 if (sk->sk_err) { 2154 copied = sock_error(sk); 2155 break; 2156 } 2157 2158 if (sk->sk_shutdown & RCV_SHUTDOWN) { 2159 /* race breaker: the shutdown could be after the 2160 * previous receive queue check 2161 */ 2162 if (__mptcp_move_skbs(sk)) 2163 continue; 2164 break; 2165 } 2166 2167 if (sk->sk_state == TCP_CLOSE) { 2168 copied = -ENOTCONN; 2169 break; 2170 } 2171 2172 if (!timeo) { 2173 copied = -EAGAIN; 2174 break; 2175 } 2176 2177 if (signal_pending(current)) { 2178 copied = sock_intr_errno(timeo); 2179 break; 2180 } 2181 } 2182 2183 pr_debug("block timeout %ld\n", timeo); 2184 mptcp_cleanup_rbuf(msk, copied); 2185 err = sk_wait_data(sk, &timeo, NULL); 2186 if (err < 0) { 2187 err = copied ? : err; 2188 goto out_err; 2189 } 2190 } 2191 2192 mptcp_cleanup_rbuf(msk, copied); 2193 2194 out_err: 2195 if (cmsg_flags && copied >= 0) { 2196 if (cmsg_flags & MPTCP_CMSG_TS) 2197 tcp_recv_timestamp(msg, sk, &tss); 2198 2199 if (cmsg_flags & MPTCP_CMSG_INQ) { 2200 unsigned int inq = mptcp_inq_hint(sk); 2201 2202 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2203 } 2204 } 2205 2206 pr_debug("msk=%p rx queue empty=%d copied=%d\n", 2207 msk, skb_queue_empty(&sk->sk_receive_queue), copied); 2208 2209 release_sock(sk); 2210 return copied; 2211 } 2212 2213 static void mptcp_retransmit_timer(struct timer_list *t) 2214 { 2215 struct inet_connection_sock *icsk = from_timer(icsk, t, 2216 icsk_retransmit_timer); 2217 struct sock *sk = &icsk->icsk_inet.sk; 2218 struct mptcp_sock *msk = mptcp_sk(sk); 2219 2220 bh_lock_sock(sk); 2221 if (!sock_owned_by_user(sk)) { 2222 /* we need a process context to retransmit */ 2223 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2224 mptcp_schedule_work(sk); 2225 } else { 2226 /* delegate our work to tcp_release_cb() */ 2227 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2228 } 2229 bh_unlock_sock(sk); 2230 sock_put(sk); 2231 } 2232 2233 static void mptcp_tout_timer(struct timer_list *t) 2234 { 2235 struct sock *sk = from_timer(sk, t, sk_timer); 2236 2237 mptcp_schedule_work(sk); 2238 sock_put(sk); 2239 } 2240 2241 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2242 * level. 2243 * 2244 * A backup subflow is returned only if that is the only kind available. 2245 */ 2246 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2247 { 2248 struct sock *backup = NULL, *pick = NULL; 2249 struct mptcp_subflow_context *subflow; 2250 int min_stale_count = INT_MAX; 2251 2252 mptcp_for_each_subflow(msk, subflow) { 2253 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2254 2255 if (!__mptcp_subflow_active(subflow)) 2256 continue; 2257 2258 /* still data outstanding at TCP level? skip this */ 2259 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2260 mptcp_pm_subflow_chk_stale(msk, ssk); 2261 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2262 continue; 2263 } 2264 2265 if (subflow->backup || subflow->request_bkup) { 2266 if (!backup) 2267 backup = ssk; 2268 continue; 2269 } 2270 2271 if (!pick) 2272 pick = ssk; 2273 } 2274 2275 if (pick) 2276 return pick; 2277 2278 /* use backup only if there are no progresses anywhere */ 2279 return min_stale_count > 1 ? backup : NULL; 2280 } 2281 2282 bool __mptcp_retransmit_pending_data(struct sock *sk) 2283 { 2284 struct mptcp_data_frag *cur, *rtx_head; 2285 struct mptcp_sock *msk = mptcp_sk(sk); 2286 2287 if (__mptcp_check_fallback(msk)) 2288 return false; 2289 2290 /* the closing socket has some data untransmitted and/or unacked: 2291 * some data in the mptcp rtx queue has not really xmitted yet. 2292 * keep it simple and re-inject the whole mptcp level rtx queue 2293 */ 2294 mptcp_data_lock(sk); 2295 __mptcp_clean_una_wakeup(sk); 2296 rtx_head = mptcp_rtx_head(sk); 2297 if (!rtx_head) { 2298 mptcp_data_unlock(sk); 2299 return false; 2300 } 2301 2302 msk->recovery_snd_nxt = msk->snd_nxt; 2303 msk->recovery = true; 2304 mptcp_data_unlock(sk); 2305 2306 msk->first_pending = rtx_head; 2307 msk->snd_burst = 0; 2308 2309 /* be sure to clear the "sent status" on all re-injected fragments */ 2310 list_for_each_entry(cur, &msk->rtx_queue, list) { 2311 if (!cur->already_sent) 2312 break; 2313 cur->already_sent = 0; 2314 } 2315 2316 return true; 2317 } 2318 2319 /* flags for __mptcp_close_ssk() */ 2320 #define MPTCP_CF_PUSH BIT(1) 2321 #define MPTCP_CF_FASTCLOSE BIT(2) 2322 2323 /* be sure to send a reset only if the caller asked for it, also 2324 * clean completely the subflow status when the subflow reaches 2325 * TCP_CLOSE state 2326 */ 2327 static void __mptcp_subflow_disconnect(struct sock *ssk, 2328 struct mptcp_subflow_context *subflow, 2329 unsigned int flags) 2330 { 2331 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 2332 (flags & MPTCP_CF_FASTCLOSE)) { 2333 /* The MPTCP code never wait on the subflow sockets, TCP-level 2334 * disconnect should never fail 2335 */ 2336 WARN_ON_ONCE(tcp_disconnect(ssk, 0)); 2337 mptcp_subflow_ctx_reset(subflow); 2338 } else { 2339 tcp_shutdown(ssk, SEND_SHUTDOWN); 2340 } 2341 } 2342 2343 /* subflow sockets can be either outgoing (connect) or incoming 2344 * (accept). 2345 * 2346 * Outgoing subflows use in-kernel sockets. 2347 * Incoming subflows do not have their own 'struct socket' allocated, 2348 * so we need to use tcp_close() after detaching them from the mptcp 2349 * parent socket. 2350 */ 2351 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2352 struct mptcp_subflow_context *subflow, 2353 unsigned int flags) 2354 { 2355 struct mptcp_sock *msk = mptcp_sk(sk); 2356 bool dispose_it, need_push = false; 2357 2358 /* If the first subflow moved to a close state before accept, e.g. due 2359 * to an incoming reset or listener shutdown, the subflow socket is 2360 * already deleted by inet_child_forget() and the mptcp socket can't 2361 * survive too. 2362 */ 2363 if (msk->in_accept_queue && msk->first == ssk && 2364 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) { 2365 /* ensure later check in mptcp_worker() will dispose the msk */ 2366 sock_set_flag(sk, SOCK_DEAD); 2367 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1)); 2368 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2369 mptcp_subflow_drop_ctx(ssk); 2370 goto out_release; 2371 } 2372 2373 dispose_it = msk->free_first || ssk != msk->first; 2374 if (dispose_it) 2375 list_del(&subflow->node); 2376 2377 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2378 2379 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) { 2380 /* be sure to force the tcp_close path 2381 * to generate the egress reset 2382 */ 2383 ssk->sk_lingertime = 0; 2384 sock_set_flag(ssk, SOCK_LINGER); 2385 subflow->send_fastclose = 1; 2386 } 2387 2388 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2389 if (!dispose_it) { 2390 __mptcp_subflow_disconnect(ssk, subflow, flags); 2391 release_sock(ssk); 2392 2393 goto out; 2394 } 2395 2396 subflow->disposable = 1; 2397 2398 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2399 * the ssk has been already destroyed, we just need to release the 2400 * reference owned by msk; 2401 */ 2402 if (!inet_csk(ssk)->icsk_ulp_ops) { 2403 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD)); 2404 kfree_rcu(subflow, rcu); 2405 } else { 2406 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2407 __tcp_close(ssk, 0); 2408 2409 /* close acquired an extra ref */ 2410 __sock_put(ssk); 2411 } 2412 2413 out_release: 2414 __mptcp_subflow_error_report(sk, ssk); 2415 release_sock(ssk); 2416 2417 sock_put(ssk); 2418 2419 if (ssk == msk->first) 2420 WRITE_ONCE(msk->first, NULL); 2421 2422 out: 2423 __mptcp_sync_sndbuf(sk); 2424 if (need_push) 2425 __mptcp_push_pending(sk, 0); 2426 2427 /* Catch every 'all subflows closed' scenario, including peers silently 2428 * closing them, e.g. due to timeout. 2429 * For established sockets, allow an additional timeout before closing, 2430 * as the protocol can still create more subflows. 2431 */ 2432 if (list_is_singular(&msk->conn_list) && msk->first && 2433 inet_sk_state_load(msk->first) == TCP_CLOSE) { 2434 if (sk->sk_state != TCP_ESTABLISHED || 2435 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) { 2436 mptcp_set_state(sk, TCP_CLOSE); 2437 mptcp_close_wake_up(sk); 2438 } else { 2439 mptcp_start_tout_timer(sk); 2440 } 2441 } 2442 } 2443 2444 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2445 struct mptcp_subflow_context *subflow) 2446 { 2447 /* The first subflow can already be closed and still in the list */ 2448 if (subflow->close_event_done) 2449 return; 2450 2451 subflow->close_event_done = true; 2452 2453 if (sk->sk_state == TCP_ESTABLISHED) 2454 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2455 2456 /* subflow aborted before reaching the fully_established status 2457 * attempt the creation of the next subflow 2458 */ 2459 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow); 2460 2461 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2462 } 2463 2464 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2465 { 2466 return 0; 2467 } 2468 2469 static void __mptcp_close_subflow(struct sock *sk) 2470 { 2471 struct mptcp_subflow_context *subflow, *tmp; 2472 struct mptcp_sock *msk = mptcp_sk(sk); 2473 2474 might_sleep(); 2475 2476 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2477 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2478 int ssk_state = inet_sk_state_load(ssk); 2479 2480 if (ssk_state != TCP_CLOSE && 2481 (ssk_state != TCP_CLOSE_WAIT || 2482 inet_sk_state_load(sk) != TCP_ESTABLISHED)) 2483 continue; 2484 2485 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2486 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2487 continue; 2488 2489 mptcp_close_ssk(sk, ssk, subflow); 2490 } 2491 2492 } 2493 2494 static bool mptcp_close_tout_expired(const struct sock *sk) 2495 { 2496 if (!inet_csk(sk)->icsk_mtup.probe_timestamp || 2497 sk->sk_state == TCP_CLOSE) 2498 return false; 2499 2500 return time_after32(tcp_jiffies32, 2501 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk)); 2502 } 2503 2504 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2505 { 2506 struct mptcp_subflow_context *subflow, *tmp; 2507 struct sock *sk = (struct sock *)msk; 2508 2509 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2510 return; 2511 2512 mptcp_token_destroy(msk); 2513 2514 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2515 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2516 bool slow; 2517 2518 slow = lock_sock_fast(tcp_sk); 2519 if (tcp_sk->sk_state != TCP_CLOSE) { 2520 mptcp_send_active_reset_reason(tcp_sk); 2521 tcp_set_state(tcp_sk, TCP_CLOSE); 2522 } 2523 unlock_sock_fast(tcp_sk, slow); 2524 } 2525 2526 /* Mirror the tcp_reset() error propagation */ 2527 switch (sk->sk_state) { 2528 case TCP_SYN_SENT: 2529 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 2530 break; 2531 case TCP_CLOSE_WAIT: 2532 WRITE_ONCE(sk->sk_err, EPIPE); 2533 break; 2534 case TCP_CLOSE: 2535 return; 2536 default: 2537 WRITE_ONCE(sk->sk_err, ECONNRESET); 2538 } 2539 2540 mptcp_set_state(sk, TCP_CLOSE); 2541 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2542 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2543 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2544 2545 /* the calling mptcp_worker will properly destroy the socket */ 2546 if (sock_flag(sk, SOCK_DEAD)) 2547 return; 2548 2549 sk->sk_state_change(sk); 2550 sk_error_report(sk); 2551 } 2552 2553 static void __mptcp_retrans(struct sock *sk) 2554 { 2555 struct mptcp_sock *msk = mptcp_sk(sk); 2556 struct mptcp_subflow_context *subflow; 2557 struct mptcp_sendmsg_info info = {}; 2558 struct mptcp_data_frag *dfrag; 2559 struct sock *ssk; 2560 int ret, err; 2561 u16 len = 0; 2562 2563 mptcp_clean_una_wakeup(sk); 2564 2565 /* first check ssk: need to kick "stale" logic */ 2566 err = mptcp_sched_get_retrans(msk); 2567 dfrag = mptcp_rtx_head(sk); 2568 if (!dfrag) { 2569 if (mptcp_data_fin_enabled(msk)) { 2570 struct inet_connection_sock *icsk = inet_csk(sk); 2571 2572 icsk->icsk_retransmits++; 2573 mptcp_set_datafin_timeout(sk); 2574 mptcp_send_ack(msk); 2575 2576 goto reset_timer; 2577 } 2578 2579 if (!mptcp_send_head(sk)) 2580 return; 2581 2582 goto reset_timer; 2583 } 2584 2585 if (err) 2586 goto reset_timer; 2587 2588 mptcp_for_each_subflow(msk, subflow) { 2589 if (READ_ONCE(subflow->scheduled)) { 2590 u16 copied = 0; 2591 2592 mptcp_subflow_set_scheduled(subflow, false); 2593 2594 ssk = mptcp_subflow_tcp_sock(subflow); 2595 2596 lock_sock(ssk); 2597 2598 /* limit retransmission to the bytes already sent on some subflows */ 2599 info.sent = 0; 2600 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : 2601 dfrag->already_sent; 2602 while (info.sent < info.limit) { 2603 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2604 if (ret <= 0) 2605 break; 2606 2607 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2608 copied += ret; 2609 info.sent += ret; 2610 } 2611 if (copied) { 2612 len = max(copied, len); 2613 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2614 info.size_goal); 2615 WRITE_ONCE(msk->allow_infinite_fallback, false); 2616 } 2617 2618 release_sock(ssk); 2619 } 2620 } 2621 2622 msk->bytes_retrans += len; 2623 dfrag->already_sent = max(dfrag->already_sent, len); 2624 2625 reset_timer: 2626 mptcp_check_and_set_pending(sk); 2627 2628 if (!mptcp_rtx_timer_pending(sk)) 2629 mptcp_reset_rtx_timer(sk); 2630 } 2631 2632 /* schedule the timeout timer for the relevant event: either close timeout 2633 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one 2634 */ 2635 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout) 2636 { 2637 struct sock *sk = (struct sock *)msk; 2638 unsigned long timeout, close_timeout; 2639 2640 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp) 2641 return; 2642 2643 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp - 2644 tcp_jiffies32 + jiffies + mptcp_close_timeout(sk); 2645 2646 /* the close timeout takes precedence on the fail one, and here at least one of 2647 * them is active 2648 */ 2649 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout; 2650 2651 sk_reset_timer(sk, &sk->sk_timer, timeout); 2652 } 2653 2654 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2655 { 2656 struct sock *ssk = msk->first; 2657 bool slow; 2658 2659 if (!ssk) 2660 return; 2661 2662 pr_debug("MP_FAIL doesn't respond, reset the subflow\n"); 2663 2664 slow = lock_sock_fast(ssk); 2665 mptcp_subflow_reset(ssk); 2666 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0); 2667 unlock_sock_fast(ssk, slow); 2668 } 2669 2670 static void mptcp_do_fastclose(struct sock *sk) 2671 { 2672 struct mptcp_subflow_context *subflow, *tmp; 2673 struct mptcp_sock *msk = mptcp_sk(sk); 2674 2675 mptcp_set_state(sk, TCP_CLOSE); 2676 mptcp_for_each_subflow_safe(msk, subflow, tmp) 2677 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), 2678 subflow, MPTCP_CF_FASTCLOSE); 2679 } 2680 2681 static void mptcp_worker(struct work_struct *work) 2682 { 2683 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2684 struct sock *sk = (struct sock *)msk; 2685 unsigned long fail_tout; 2686 int state; 2687 2688 lock_sock(sk); 2689 state = sk->sk_state; 2690 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN))) 2691 goto unlock; 2692 2693 mptcp_check_fastclose(msk); 2694 2695 mptcp_pm_nl_work(msk); 2696 2697 mptcp_check_send_data_fin(sk); 2698 mptcp_check_data_fin_ack(sk); 2699 mptcp_check_data_fin(sk); 2700 2701 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2702 __mptcp_close_subflow(sk); 2703 2704 if (mptcp_close_tout_expired(sk)) { 2705 mptcp_do_fastclose(sk); 2706 mptcp_close_wake_up(sk); 2707 } 2708 2709 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) { 2710 __mptcp_destroy_sock(sk); 2711 goto unlock; 2712 } 2713 2714 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2715 __mptcp_retrans(sk); 2716 2717 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0; 2718 if (fail_tout && time_after(jiffies, fail_tout)) 2719 mptcp_mp_fail_no_response(msk); 2720 2721 unlock: 2722 release_sock(sk); 2723 sock_put(sk); 2724 } 2725 2726 static void __mptcp_init_sock(struct sock *sk) 2727 { 2728 struct mptcp_sock *msk = mptcp_sk(sk); 2729 2730 INIT_LIST_HEAD(&msk->conn_list); 2731 INIT_LIST_HEAD(&msk->join_list); 2732 INIT_LIST_HEAD(&msk->rtx_queue); 2733 INIT_WORK(&msk->work, mptcp_worker); 2734 msk->out_of_order_queue = RB_ROOT; 2735 msk->first_pending = NULL; 2736 msk->timer_ival = TCP_RTO_MIN; 2737 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO; 2738 2739 WRITE_ONCE(msk->first, NULL); 2740 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2741 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2742 WRITE_ONCE(msk->allow_infinite_fallback, true); 2743 msk->recovery = false; 2744 msk->subflow_id = 1; 2745 msk->last_data_sent = tcp_jiffies32; 2746 msk->last_data_recv = tcp_jiffies32; 2747 msk->last_ack_recv = tcp_jiffies32; 2748 2749 mptcp_pm_data_init(msk); 2750 2751 /* re-use the csk retrans timer for MPTCP-level retrans */ 2752 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2753 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0); 2754 } 2755 2756 static void mptcp_ca_reset(struct sock *sk) 2757 { 2758 struct inet_connection_sock *icsk = inet_csk(sk); 2759 2760 tcp_assign_congestion_control(sk); 2761 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name, 2762 sizeof(mptcp_sk(sk)->ca_name)); 2763 2764 /* no need to keep a reference to the ops, the name will suffice */ 2765 tcp_cleanup_congestion_control(sk); 2766 icsk->icsk_ca_ops = NULL; 2767 } 2768 2769 static int mptcp_init_sock(struct sock *sk) 2770 { 2771 struct net *net = sock_net(sk); 2772 int ret; 2773 2774 __mptcp_init_sock(sk); 2775 2776 if (!mptcp_is_enabled(net)) 2777 return -ENOPROTOOPT; 2778 2779 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2780 return -ENOMEM; 2781 2782 rcu_read_lock(); 2783 ret = mptcp_init_sched(mptcp_sk(sk), 2784 mptcp_sched_find(mptcp_get_scheduler(net))); 2785 rcu_read_unlock(); 2786 if (ret) 2787 return ret; 2788 2789 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); 2790 2791 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2792 * propagate the correct value 2793 */ 2794 mptcp_ca_reset(sk); 2795 2796 sk_sockets_allocated_inc(sk); 2797 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]); 2798 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]); 2799 2800 return 0; 2801 } 2802 2803 static void __mptcp_clear_xmit(struct sock *sk) 2804 { 2805 struct mptcp_sock *msk = mptcp_sk(sk); 2806 struct mptcp_data_frag *dtmp, *dfrag; 2807 2808 WRITE_ONCE(msk->first_pending, NULL); 2809 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2810 dfrag_clear(sk, dfrag); 2811 } 2812 2813 void mptcp_cancel_work(struct sock *sk) 2814 { 2815 struct mptcp_sock *msk = mptcp_sk(sk); 2816 2817 if (cancel_work_sync(&msk->work)) 2818 __sock_put(sk); 2819 } 2820 2821 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2822 { 2823 lock_sock(ssk); 2824 2825 switch (ssk->sk_state) { 2826 case TCP_LISTEN: 2827 if (!(how & RCV_SHUTDOWN)) 2828 break; 2829 fallthrough; 2830 case TCP_SYN_SENT: 2831 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK)); 2832 break; 2833 default: 2834 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2835 pr_debug("Fallback\n"); 2836 ssk->sk_shutdown |= how; 2837 tcp_shutdown(ssk, how); 2838 2839 /* simulate the data_fin ack reception to let the state 2840 * machine move forward 2841 */ 2842 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt); 2843 mptcp_schedule_work(sk); 2844 } else { 2845 pr_debug("Sending DATA_FIN on subflow %p\n", ssk); 2846 tcp_send_ack(ssk); 2847 if (!mptcp_rtx_timer_pending(sk)) 2848 mptcp_reset_rtx_timer(sk); 2849 } 2850 break; 2851 } 2852 2853 release_sock(ssk); 2854 } 2855 2856 void mptcp_set_state(struct sock *sk, int state) 2857 { 2858 int oldstate = sk->sk_state; 2859 2860 switch (state) { 2861 case TCP_ESTABLISHED: 2862 if (oldstate != TCP_ESTABLISHED) 2863 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2864 break; 2865 case TCP_CLOSE_WAIT: 2866 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state: 2867 * MPTCP "accepted" sockets will be created later on. So no 2868 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT. 2869 */ 2870 break; 2871 default: 2872 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 2873 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2874 } 2875 2876 inet_sk_state_store(sk, state); 2877 } 2878 2879 static const unsigned char new_state[16] = { 2880 /* current state: new state: action: */ 2881 [0 /* (Invalid) */] = TCP_CLOSE, 2882 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2883 [TCP_SYN_SENT] = TCP_CLOSE, 2884 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 2885 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 2886 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 2887 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 2888 [TCP_CLOSE] = TCP_CLOSE, 2889 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 2890 [TCP_LAST_ACK] = TCP_LAST_ACK, 2891 [TCP_LISTEN] = TCP_CLOSE, 2892 [TCP_CLOSING] = TCP_CLOSING, 2893 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 2894 }; 2895 2896 static int mptcp_close_state(struct sock *sk) 2897 { 2898 int next = (int)new_state[sk->sk_state]; 2899 int ns = next & TCP_STATE_MASK; 2900 2901 mptcp_set_state(sk, ns); 2902 2903 return next & TCP_ACTION_FIN; 2904 } 2905 2906 static void mptcp_check_send_data_fin(struct sock *sk) 2907 { 2908 struct mptcp_subflow_context *subflow; 2909 struct mptcp_sock *msk = mptcp_sk(sk); 2910 2911 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n", 2912 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 2913 msk->snd_nxt, msk->write_seq); 2914 2915 /* we still need to enqueue subflows or not really shutting down, 2916 * skip this 2917 */ 2918 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 2919 mptcp_send_head(sk)) 2920 return; 2921 2922 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 2923 2924 mptcp_for_each_subflow(msk, subflow) { 2925 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2926 2927 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 2928 } 2929 } 2930 2931 static void __mptcp_wr_shutdown(struct sock *sk) 2932 { 2933 struct mptcp_sock *msk = mptcp_sk(sk); 2934 2935 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n", 2936 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 2937 !!mptcp_send_head(sk)); 2938 2939 /* will be ignored by fallback sockets */ 2940 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 2941 WRITE_ONCE(msk->snd_data_fin_enable, 1); 2942 2943 mptcp_check_send_data_fin(sk); 2944 } 2945 2946 static void __mptcp_destroy_sock(struct sock *sk) 2947 { 2948 struct mptcp_sock *msk = mptcp_sk(sk); 2949 2950 pr_debug("msk=%p\n", msk); 2951 2952 might_sleep(); 2953 2954 mptcp_stop_rtx_timer(sk); 2955 sk_stop_timer(sk, &sk->sk_timer); 2956 msk->pm.status = 0; 2957 mptcp_release_sched(msk); 2958 2959 sk->sk_prot->destroy(sk); 2960 2961 sk_stream_kill_queues(sk); 2962 xfrm_sk_free_policy(sk); 2963 2964 sock_put(sk); 2965 } 2966 2967 void __mptcp_unaccepted_force_close(struct sock *sk) 2968 { 2969 sock_set_flag(sk, SOCK_DEAD); 2970 mptcp_do_fastclose(sk); 2971 __mptcp_destroy_sock(sk); 2972 } 2973 2974 static __poll_t mptcp_check_readable(struct sock *sk) 2975 { 2976 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0; 2977 } 2978 2979 static void mptcp_check_listen_stop(struct sock *sk) 2980 { 2981 struct sock *ssk; 2982 2983 if (inet_sk_state_load(sk) != TCP_LISTEN) 2984 return; 2985 2986 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 2987 ssk = mptcp_sk(sk)->first; 2988 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN)) 2989 return; 2990 2991 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2992 tcp_set_state(ssk, TCP_CLOSE); 2993 mptcp_subflow_queue_clean(sk, ssk); 2994 inet_csk_listen_stop(ssk); 2995 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED); 2996 release_sock(ssk); 2997 } 2998 2999 bool __mptcp_close(struct sock *sk, long timeout) 3000 { 3001 struct mptcp_subflow_context *subflow; 3002 struct mptcp_sock *msk = mptcp_sk(sk); 3003 bool do_cancel_work = false; 3004 int subflows_alive = 0; 3005 3006 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3007 3008 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 3009 mptcp_check_listen_stop(sk); 3010 mptcp_set_state(sk, TCP_CLOSE); 3011 goto cleanup; 3012 } 3013 3014 if (mptcp_data_avail(msk) || timeout < 0) { 3015 /* If the msk has read data, or the caller explicitly ask it, 3016 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose 3017 */ 3018 mptcp_do_fastclose(sk); 3019 timeout = 0; 3020 } else if (mptcp_close_state(sk)) { 3021 __mptcp_wr_shutdown(sk); 3022 } 3023 3024 sk_stream_wait_close(sk, timeout); 3025 3026 cleanup: 3027 /* orphan all the subflows */ 3028 mptcp_for_each_subflow(msk, subflow) { 3029 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3030 bool slow = lock_sock_fast_nested(ssk); 3031 3032 subflows_alive += ssk->sk_state != TCP_CLOSE; 3033 3034 /* since the close timeout takes precedence on the fail one, 3035 * cancel the latter 3036 */ 3037 if (ssk == msk->first) 3038 subflow->fail_tout = 0; 3039 3040 /* detach from the parent socket, but allow data_ready to 3041 * push incoming data into the mptcp stack, to properly ack it 3042 */ 3043 ssk->sk_socket = NULL; 3044 ssk->sk_wq = NULL; 3045 unlock_sock_fast(ssk, slow); 3046 } 3047 sock_orphan(sk); 3048 3049 /* all the subflows are closed, only timeout can change the msk 3050 * state, let's not keep resources busy for no reasons 3051 */ 3052 if (subflows_alive == 0) 3053 mptcp_set_state(sk, TCP_CLOSE); 3054 3055 sock_hold(sk); 3056 pr_debug("msk=%p state=%d\n", sk, sk->sk_state); 3057 mptcp_pm_connection_closed(msk); 3058 3059 if (sk->sk_state == TCP_CLOSE) { 3060 __mptcp_destroy_sock(sk); 3061 do_cancel_work = true; 3062 } else { 3063 mptcp_start_tout_timer(sk); 3064 } 3065 3066 return do_cancel_work; 3067 } 3068 3069 static void mptcp_close(struct sock *sk, long timeout) 3070 { 3071 bool do_cancel_work; 3072 3073 lock_sock(sk); 3074 3075 do_cancel_work = __mptcp_close(sk, timeout); 3076 release_sock(sk); 3077 if (do_cancel_work) 3078 mptcp_cancel_work(sk); 3079 3080 sock_put(sk); 3081 } 3082 3083 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 3084 { 3085 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3086 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 3087 struct ipv6_pinfo *msk6 = inet6_sk(msk); 3088 3089 msk->sk_v6_daddr = ssk->sk_v6_daddr; 3090 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 3091 3092 if (msk6 && ssk6) { 3093 msk6->saddr = ssk6->saddr; 3094 msk6->flow_label = ssk6->flow_label; 3095 } 3096 #endif 3097 3098 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 3099 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 3100 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 3101 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 3102 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 3103 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 3104 } 3105 3106 static int mptcp_disconnect(struct sock *sk, int flags) 3107 { 3108 struct mptcp_sock *msk = mptcp_sk(sk); 3109 3110 /* We are on the fastopen error path. We can't call straight into the 3111 * subflows cleanup code due to lock nesting (we are already under 3112 * msk->firstsocket lock). 3113 */ 3114 if (msk->fastopening) 3115 return -EBUSY; 3116 3117 mptcp_check_listen_stop(sk); 3118 mptcp_set_state(sk, TCP_CLOSE); 3119 3120 mptcp_stop_rtx_timer(sk); 3121 mptcp_stop_tout_timer(sk); 3122 3123 mptcp_pm_connection_closed(msk); 3124 3125 /* msk->subflow is still intact, the following will not free the first 3126 * subflow 3127 */ 3128 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE); 3129 WRITE_ONCE(msk->flags, 0); 3130 msk->cb_flags = 0; 3131 msk->recovery = false; 3132 WRITE_ONCE(msk->can_ack, false); 3133 WRITE_ONCE(msk->fully_established, false); 3134 WRITE_ONCE(msk->rcv_data_fin, false); 3135 WRITE_ONCE(msk->snd_data_fin_enable, false); 3136 WRITE_ONCE(msk->rcv_fastclose, false); 3137 WRITE_ONCE(msk->use_64bit_ack, false); 3138 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 3139 mptcp_pm_data_reset(msk); 3140 mptcp_ca_reset(sk); 3141 msk->bytes_consumed = 0; 3142 msk->bytes_acked = 0; 3143 msk->bytes_received = 0; 3144 msk->bytes_sent = 0; 3145 msk->bytes_retrans = 0; 3146 msk->rcvspace_init = 0; 3147 3148 WRITE_ONCE(sk->sk_shutdown, 0); 3149 sk_error_report(sk); 3150 return 0; 3151 } 3152 3153 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3154 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 3155 { 3156 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo); 3157 3158 return (struct ipv6_pinfo *)(((u8 *)sk) + offset); 3159 } 3160 3161 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk) 3162 { 3163 const struct ipv6_pinfo *np = inet6_sk(sk); 3164 struct ipv6_txoptions *opt; 3165 struct ipv6_pinfo *newnp; 3166 3167 newnp = inet6_sk(newsk); 3168 3169 rcu_read_lock(); 3170 opt = rcu_dereference(np->opt); 3171 if (opt) { 3172 opt = ipv6_dup_options(newsk, opt); 3173 if (!opt) 3174 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__); 3175 } 3176 RCU_INIT_POINTER(newnp->opt, opt); 3177 rcu_read_unlock(); 3178 } 3179 #endif 3180 3181 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk) 3182 { 3183 struct ip_options_rcu *inet_opt, *newopt = NULL; 3184 const struct inet_sock *inet = inet_sk(sk); 3185 struct inet_sock *newinet; 3186 3187 newinet = inet_sk(newsk); 3188 3189 rcu_read_lock(); 3190 inet_opt = rcu_dereference(inet->inet_opt); 3191 if (inet_opt) { 3192 newopt = sock_kmalloc(newsk, sizeof(*inet_opt) + 3193 inet_opt->opt.optlen, GFP_ATOMIC); 3194 if (newopt) 3195 memcpy(newopt, inet_opt, sizeof(*inet_opt) + 3196 inet_opt->opt.optlen); 3197 else 3198 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__); 3199 } 3200 RCU_INIT_POINTER(newinet->inet_opt, newopt); 3201 rcu_read_unlock(); 3202 } 3203 3204 struct sock *mptcp_sk_clone_init(const struct sock *sk, 3205 const struct mptcp_options_received *mp_opt, 3206 struct sock *ssk, 3207 struct request_sock *req) 3208 { 3209 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 3210 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 3211 struct mptcp_subflow_context *subflow; 3212 struct mptcp_sock *msk; 3213 3214 if (!nsk) 3215 return NULL; 3216 3217 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3218 if (nsk->sk_family == AF_INET6) 3219 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 3220 #endif 3221 3222 __mptcp_init_sock(nsk); 3223 3224 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3225 if (nsk->sk_family == AF_INET6) 3226 mptcp_copy_ip6_options(nsk, sk); 3227 else 3228 #endif 3229 mptcp_copy_ip_options(nsk, sk); 3230 3231 msk = mptcp_sk(nsk); 3232 WRITE_ONCE(msk->local_key, subflow_req->local_key); 3233 WRITE_ONCE(msk->token, subflow_req->token); 3234 msk->in_accept_queue = 1; 3235 WRITE_ONCE(msk->fully_established, false); 3236 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 3237 WRITE_ONCE(msk->csum_enabled, true); 3238 3239 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1); 3240 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3241 WRITE_ONCE(msk->snd_una, msk->write_seq); 3242 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 3243 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 3244 mptcp_init_sched(msk, mptcp_sk(sk)->sched); 3245 3246 /* passive msk is created after the first/MPC subflow */ 3247 msk->subflow_id = 2; 3248 3249 sock_reset_flag(nsk, SOCK_RCU_FREE); 3250 security_inet_csk_clone(nsk, req); 3251 3252 /* this can't race with mptcp_close(), as the msk is 3253 * not yet exposted to user-space 3254 */ 3255 mptcp_set_state(nsk, TCP_ESTABLISHED); 3256 3257 /* The msk maintain a ref to each subflow in the connections list */ 3258 WRITE_ONCE(msk->first, ssk); 3259 subflow = mptcp_subflow_ctx(ssk); 3260 list_add(&subflow->node, &msk->conn_list); 3261 sock_hold(ssk); 3262 3263 /* new mpc subflow takes ownership of the newly 3264 * created mptcp socket 3265 */ 3266 mptcp_token_accept(subflow_req, msk); 3267 3268 /* set msk addresses early to ensure mptcp_pm_get_local_id() 3269 * uses the correct data 3270 */ 3271 mptcp_copy_inaddrs(nsk, ssk); 3272 __mptcp_propagate_sndbuf(nsk, ssk); 3273 3274 mptcp_rcv_space_init(msk, ssk); 3275 3276 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 3277 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 3278 bh_unlock_sock(nsk); 3279 3280 /* note: the newly allocated socket refcount is 2 now */ 3281 return nsk; 3282 } 3283 3284 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3285 { 3286 const struct tcp_sock *tp = tcp_sk(ssk); 3287 3288 msk->rcvspace_init = 1; 3289 msk->rcvq_space.copied = 0; 3290 msk->rcvq_space.rtt_us = 0; 3291 3292 msk->rcvq_space.time = tp->tcp_mstamp; 3293 3294 /* initial rcv_space offering made to peer */ 3295 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3296 TCP_INIT_CWND * tp->advmss); 3297 if (msk->rcvq_space.space == 0) 3298 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3299 } 3300 3301 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags) 3302 { 3303 struct mptcp_subflow_context *subflow, *tmp; 3304 struct sock *sk = (struct sock *)msk; 3305 3306 __mptcp_clear_xmit(sk); 3307 3308 /* join list will be eventually flushed (with rst) at sock lock release time */ 3309 mptcp_for_each_subflow_safe(msk, subflow, tmp) 3310 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags); 3311 3312 __skb_queue_purge(&sk->sk_receive_queue); 3313 skb_rbtree_purge(&msk->out_of_order_queue); 3314 3315 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3316 * inet_sock_destruct() will dispose it 3317 */ 3318 mptcp_token_destroy(msk); 3319 mptcp_pm_free_anno_list(msk); 3320 mptcp_free_local_addr_list(msk); 3321 } 3322 3323 static void mptcp_destroy(struct sock *sk) 3324 { 3325 struct mptcp_sock *msk = mptcp_sk(sk); 3326 3327 /* allow the following to close even the initial subflow */ 3328 msk->free_first = 1; 3329 mptcp_destroy_common(msk, 0); 3330 sk_sockets_allocated_dec(sk); 3331 } 3332 3333 void __mptcp_data_acked(struct sock *sk) 3334 { 3335 if (!sock_owned_by_user(sk)) 3336 __mptcp_clean_una(sk); 3337 else 3338 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3339 } 3340 3341 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3342 { 3343 if (!mptcp_send_head(sk)) 3344 return; 3345 3346 if (!sock_owned_by_user(sk)) 3347 __mptcp_subflow_push_pending(sk, ssk, false); 3348 else 3349 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3350 } 3351 3352 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3353 BIT(MPTCP_RETRANSMIT) | \ 3354 BIT(MPTCP_FLUSH_JOIN_LIST) | \ 3355 BIT(MPTCP_DEQUEUE)) 3356 3357 /* processes deferred events and flush wmem */ 3358 static void mptcp_release_cb(struct sock *sk) 3359 __must_hold(&sk->sk_lock.slock) 3360 { 3361 struct mptcp_sock *msk = mptcp_sk(sk); 3362 3363 for (;;) { 3364 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED); 3365 struct list_head join_list; 3366 3367 if (!flags) 3368 break; 3369 3370 INIT_LIST_HEAD(&join_list); 3371 list_splice_init(&msk->join_list, &join_list); 3372 3373 /* the following actions acquire the subflow socket lock 3374 * 3375 * 1) can't be invoked in atomic scope 3376 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3377 * datapath acquires the msk socket spinlock while helding 3378 * the subflow socket lock 3379 */ 3380 msk->cb_flags &= ~flags; 3381 spin_unlock_bh(&sk->sk_lock.slock); 3382 3383 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3384 __mptcp_flush_join_list(sk, &join_list); 3385 if (flags & BIT(MPTCP_PUSH_PENDING)) 3386 __mptcp_push_pending(sk, 0); 3387 if (flags & BIT(MPTCP_RETRANSMIT)) 3388 __mptcp_retrans(sk); 3389 if ((flags & BIT(MPTCP_DEQUEUE)) && __mptcp_move_skbs(sk)) { 3390 /* notify ack seq update */ 3391 mptcp_cleanup_rbuf(msk, 0); 3392 sk->sk_data_ready(sk); 3393 } 3394 3395 cond_resched(); 3396 spin_lock_bh(&sk->sk_lock.slock); 3397 } 3398 3399 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3400 __mptcp_clean_una_wakeup(sk); 3401 if (unlikely(msk->cb_flags)) { 3402 /* be sure to sync the msk state before taking actions 3403 * depending on sk_state (MPTCP_ERROR_REPORT) 3404 * On sk release avoid actions depending on the first subflow 3405 */ 3406 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first) 3407 __mptcp_sync_state(sk, msk->pending_state); 3408 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3409 __mptcp_error_report(sk); 3410 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags)) 3411 __mptcp_sync_sndbuf(sk); 3412 } 3413 } 3414 3415 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3416 * TCP can't schedule delack timer before the subflow is fully established. 3417 * MPTCP uses the delack timer to do 3rd ack retransmissions 3418 */ 3419 static void schedule_3rdack_retransmission(struct sock *ssk) 3420 { 3421 struct inet_connection_sock *icsk = inet_csk(ssk); 3422 struct tcp_sock *tp = tcp_sk(ssk); 3423 unsigned long timeout; 3424 3425 if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established)) 3426 return; 3427 3428 /* reschedule with a timeout above RTT, as we must look only for drop */ 3429 if (tp->srtt_us) 3430 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3431 else 3432 timeout = TCP_TIMEOUT_INIT; 3433 timeout += jiffies; 3434 3435 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3436 smp_store_release(&icsk->icsk_ack.pending, 3437 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER); 3438 icsk->icsk_ack.timeout = timeout; 3439 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3440 } 3441 3442 void mptcp_subflow_process_delegated(struct sock *ssk, long status) 3443 { 3444 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3445 struct sock *sk = subflow->conn; 3446 3447 if (status & BIT(MPTCP_DELEGATE_SEND)) { 3448 mptcp_data_lock(sk); 3449 if (!sock_owned_by_user(sk)) 3450 __mptcp_subflow_push_pending(sk, ssk, true); 3451 else 3452 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3453 mptcp_data_unlock(sk); 3454 } 3455 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) { 3456 mptcp_data_lock(sk); 3457 if (!sock_owned_by_user(sk)) 3458 __mptcp_sync_sndbuf(sk); 3459 else 3460 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags); 3461 mptcp_data_unlock(sk); 3462 } 3463 if (status & BIT(MPTCP_DELEGATE_ACK)) 3464 schedule_3rdack_retransmission(ssk); 3465 } 3466 3467 static int mptcp_hash(struct sock *sk) 3468 { 3469 /* should never be called, 3470 * we hash the TCP subflows not the MPTCP socket 3471 */ 3472 WARN_ON_ONCE(1); 3473 return 0; 3474 } 3475 3476 static void mptcp_unhash(struct sock *sk) 3477 { 3478 /* called from sk_common_release(), but nothing to do here */ 3479 } 3480 3481 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3482 { 3483 struct mptcp_sock *msk = mptcp_sk(sk); 3484 3485 pr_debug("msk=%p, ssk=%p\n", msk, msk->first); 3486 if (WARN_ON_ONCE(!msk->first)) 3487 return -EINVAL; 3488 3489 return inet_csk_get_port(msk->first, snum); 3490 } 3491 3492 void mptcp_finish_connect(struct sock *ssk) 3493 { 3494 struct mptcp_subflow_context *subflow; 3495 struct mptcp_sock *msk; 3496 struct sock *sk; 3497 3498 subflow = mptcp_subflow_ctx(ssk); 3499 sk = subflow->conn; 3500 msk = mptcp_sk(sk); 3501 3502 pr_debug("msk=%p, token=%u\n", sk, subflow->token); 3503 3504 subflow->map_seq = subflow->iasn; 3505 subflow->map_subflow_seq = 1; 3506 3507 /* the socket is not connected yet, no msk/subflow ops can access/race 3508 * accessing the field below 3509 */ 3510 WRITE_ONCE(msk->local_key, subflow->local_key); 3511 3512 mptcp_pm_new_connection(msk, ssk, 0); 3513 } 3514 3515 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3516 { 3517 write_lock_bh(&sk->sk_callback_lock); 3518 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3519 sk_set_socket(sk, parent); 3520 sk->sk_uid = SOCK_INODE(parent)->i_uid; 3521 write_unlock_bh(&sk->sk_callback_lock); 3522 } 3523 3524 bool mptcp_finish_join(struct sock *ssk) 3525 { 3526 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3527 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3528 struct sock *parent = (void *)msk; 3529 bool ret = true; 3530 3531 pr_debug("msk=%p, subflow=%p\n", msk, subflow); 3532 3533 /* mptcp socket already closing? */ 3534 if (!mptcp_is_fully_established(parent)) { 3535 subflow->reset_reason = MPTCP_RST_EMPTCP; 3536 return false; 3537 } 3538 3539 /* active subflow, already present inside the conn_list */ 3540 if (!list_empty(&subflow->node)) { 3541 mptcp_subflow_joined(msk, ssk); 3542 mptcp_propagate_sndbuf(parent, ssk); 3543 return true; 3544 } 3545 3546 if (!mptcp_pm_allow_new_subflow(msk)) 3547 goto err_prohibited; 3548 3549 /* If we can't acquire msk socket lock here, let the release callback 3550 * handle it 3551 */ 3552 mptcp_data_lock(parent); 3553 if (!sock_owned_by_user(parent)) { 3554 ret = __mptcp_finish_join(msk, ssk); 3555 if (ret) { 3556 sock_hold(ssk); 3557 list_add_tail(&subflow->node, &msk->conn_list); 3558 } 3559 } else { 3560 sock_hold(ssk); 3561 list_add_tail(&subflow->node, &msk->join_list); 3562 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3563 } 3564 mptcp_data_unlock(parent); 3565 3566 if (!ret) { 3567 err_prohibited: 3568 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3569 return false; 3570 } 3571 3572 return true; 3573 } 3574 3575 static void mptcp_shutdown(struct sock *sk, int how) 3576 { 3577 pr_debug("sk=%p, how=%d\n", sk, how); 3578 3579 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3580 __mptcp_wr_shutdown(sk); 3581 } 3582 3583 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3584 { 3585 const struct sock *sk = (void *)msk; 3586 u64 delta; 3587 3588 if (sk->sk_state == TCP_LISTEN) 3589 return -EINVAL; 3590 3591 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3592 return 0; 3593 3594 delta = msk->write_seq - v; 3595 if (__mptcp_check_fallback(msk) && msk->first) { 3596 struct tcp_sock *tp = tcp_sk(msk->first); 3597 3598 /* the first subflow is disconnected after close - see 3599 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3600 * so ignore that status, too. 3601 */ 3602 if (!((1 << msk->first->sk_state) & 3603 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3604 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3605 } 3606 if (delta > INT_MAX) 3607 delta = INT_MAX; 3608 3609 return (int)delta; 3610 } 3611 3612 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg) 3613 { 3614 struct mptcp_sock *msk = mptcp_sk(sk); 3615 bool slow; 3616 3617 switch (cmd) { 3618 case SIOCINQ: 3619 if (sk->sk_state == TCP_LISTEN) 3620 return -EINVAL; 3621 3622 lock_sock(sk); 3623 if (__mptcp_move_skbs(sk)) 3624 mptcp_cleanup_rbuf(msk, 0); 3625 *karg = mptcp_inq_hint(sk); 3626 release_sock(sk); 3627 break; 3628 case SIOCOUTQ: 3629 slow = lock_sock_fast(sk); 3630 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3631 unlock_sock_fast(sk, slow); 3632 break; 3633 case SIOCOUTQNSD: 3634 slow = lock_sock_fast(sk); 3635 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt); 3636 unlock_sock_fast(sk, slow); 3637 break; 3638 default: 3639 return -ENOIOCTLCMD; 3640 } 3641 3642 return 0; 3643 } 3644 3645 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) 3646 { 3647 struct mptcp_subflow_context *subflow; 3648 struct mptcp_sock *msk = mptcp_sk(sk); 3649 int err = -EINVAL; 3650 struct sock *ssk; 3651 3652 ssk = __mptcp_nmpc_sk(msk); 3653 if (IS_ERR(ssk)) 3654 return PTR_ERR(ssk); 3655 3656 mptcp_set_state(sk, TCP_SYN_SENT); 3657 subflow = mptcp_subflow_ctx(ssk); 3658 #ifdef CONFIG_TCP_MD5SIG 3659 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3660 * TCP option space. 3661 */ 3662 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info)) 3663 mptcp_subflow_early_fallback(msk, subflow); 3664 #endif 3665 if (subflow->request_mptcp) { 3666 if (mptcp_active_should_disable(sk)) { 3667 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEACTIVEDISABLED); 3668 mptcp_subflow_early_fallback(msk, subflow); 3669 } else if (mptcp_token_new_connect(ssk) < 0) { 3670 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT); 3671 mptcp_subflow_early_fallback(msk, subflow); 3672 } 3673 } 3674 3675 WRITE_ONCE(msk->write_seq, subflow->idsn); 3676 WRITE_ONCE(msk->snd_nxt, subflow->idsn); 3677 WRITE_ONCE(msk->snd_una, subflow->idsn); 3678 if (likely(!__mptcp_check_fallback(msk))) 3679 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE); 3680 3681 /* if reaching here via the fastopen/sendmsg path, the caller already 3682 * acquired the subflow socket lock, too. 3683 */ 3684 if (!msk->fastopening) 3685 lock_sock(ssk); 3686 3687 /* the following mirrors closely a very small chunk of code from 3688 * __inet_stream_connect() 3689 */ 3690 if (ssk->sk_state != TCP_CLOSE) 3691 goto out; 3692 3693 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) { 3694 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len); 3695 if (err) 3696 goto out; 3697 } 3698 3699 err = ssk->sk_prot->connect(ssk, uaddr, addr_len); 3700 if (err < 0) 3701 goto out; 3702 3703 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk)); 3704 3705 out: 3706 if (!msk->fastopening) 3707 release_sock(ssk); 3708 3709 /* on successful connect, the msk state will be moved to established by 3710 * subflow_finish_connect() 3711 */ 3712 if (unlikely(err)) { 3713 /* avoid leaving a dangling token in an unconnected socket */ 3714 mptcp_token_destroy(msk); 3715 mptcp_set_state(sk, TCP_CLOSE); 3716 return err; 3717 } 3718 3719 mptcp_copy_inaddrs(sk, ssk); 3720 return 0; 3721 } 3722 3723 static struct proto mptcp_prot = { 3724 .name = "MPTCP", 3725 .owner = THIS_MODULE, 3726 .init = mptcp_init_sock, 3727 .connect = mptcp_connect, 3728 .disconnect = mptcp_disconnect, 3729 .close = mptcp_close, 3730 .setsockopt = mptcp_setsockopt, 3731 .getsockopt = mptcp_getsockopt, 3732 .shutdown = mptcp_shutdown, 3733 .destroy = mptcp_destroy, 3734 .sendmsg = mptcp_sendmsg, 3735 .ioctl = mptcp_ioctl, 3736 .recvmsg = mptcp_recvmsg, 3737 .release_cb = mptcp_release_cb, 3738 .hash = mptcp_hash, 3739 .unhash = mptcp_unhash, 3740 .get_port = mptcp_get_port, 3741 .stream_memory_free = mptcp_stream_memory_free, 3742 .sockets_allocated = &mptcp_sockets_allocated, 3743 3744 .memory_allocated = &tcp_memory_allocated, 3745 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc, 3746 3747 .memory_pressure = &tcp_memory_pressure, 3748 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3749 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3750 .sysctl_mem = sysctl_tcp_mem, 3751 .obj_size = sizeof(struct mptcp_sock), 3752 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3753 .no_autobind = true, 3754 }; 3755 3756 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3757 { 3758 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3759 struct sock *ssk, *sk = sock->sk; 3760 int err = -EINVAL; 3761 3762 lock_sock(sk); 3763 ssk = __mptcp_nmpc_sk(msk); 3764 if (IS_ERR(ssk)) { 3765 err = PTR_ERR(ssk); 3766 goto unlock; 3767 } 3768 3769 if (sk->sk_family == AF_INET) 3770 err = inet_bind_sk(ssk, uaddr, addr_len); 3771 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3772 else if (sk->sk_family == AF_INET6) 3773 err = inet6_bind_sk(ssk, uaddr, addr_len); 3774 #endif 3775 if (!err) 3776 mptcp_copy_inaddrs(sk, ssk); 3777 3778 unlock: 3779 release_sock(sk); 3780 return err; 3781 } 3782 3783 static int mptcp_listen(struct socket *sock, int backlog) 3784 { 3785 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3786 struct sock *sk = sock->sk; 3787 struct sock *ssk; 3788 int err; 3789 3790 pr_debug("msk=%p\n", msk); 3791 3792 lock_sock(sk); 3793 3794 err = -EINVAL; 3795 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM) 3796 goto unlock; 3797 3798 ssk = __mptcp_nmpc_sk(msk); 3799 if (IS_ERR(ssk)) { 3800 err = PTR_ERR(ssk); 3801 goto unlock; 3802 } 3803 3804 mptcp_set_state(sk, TCP_LISTEN); 3805 sock_set_flag(sk, SOCK_RCU_FREE); 3806 3807 lock_sock(ssk); 3808 err = __inet_listen_sk(ssk, backlog); 3809 release_sock(ssk); 3810 mptcp_set_state(sk, inet_sk_state_load(ssk)); 3811 3812 if (!err) { 3813 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 3814 mptcp_copy_inaddrs(sk, ssk); 3815 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED); 3816 } 3817 3818 unlock: 3819 release_sock(sk); 3820 return err; 3821 } 3822 3823 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3824 struct proto_accept_arg *arg) 3825 { 3826 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3827 struct sock *ssk, *newsk; 3828 3829 pr_debug("msk=%p\n", msk); 3830 3831 /* Buggy applications can call accept on socket states other then LISTEN 3832 * but no need to allocate the first subflow just to error out. 3833 */ 3834 ssk = READ_ONCE(msk->first); 3835 if (!ssk) 3836 return -EINVAL; 3837 3838 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk)); 3839 newsk = inet_csk_accept(ssk, arg); 3840 if (!newsk) 3841 return arg->err; 3842 3843 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk)); 3844 if (sk_is_mptcp(newsk)) { 3845 struct mptcp_subflow_context *subflow; 3846 struct sock *new_mptcp_sock; 3847 3848 subflow = mptcp_subflow_ctx(newsk); 3849 new_mptcp_sock = subflow->conn; 3850 3851 /* is_mptcp should be false if subflow->conn is missing, see 3852 * subflow_syn_recv_sock() 3853 */ 3854 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3855 tcp_sk(newsk)->is_mptcp = 0; 3856 goto tcpfallback; 3857 } 3858 3859 newsk = new_mptcp_sock; 3860 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3861 3862 newsk->sk_kern_sock = arg->kern; 3863 lock_sock(newsk); 3864 __inet_accept(sock, newsock, newsk); 3865 3866 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags); 3867 msk = mptcp_sk(newsk); 3868 msk->in_accept_queue = 0; 3869 3870 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 3871 * This is needed so NOSPACE flag can be set from tcp stack. 3872 */ 3873 mptcp_for_each_subflow(msk, subflow) { 3874 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3875 3876 if (!ssk->sk_socket) 3877 mptcp_sock_graft(ssk, newsock); 3878 } 3879 3880 /* Do late cleanup for the first subflow as necessary. Also 3881 * deal with bad peers not doing a complete shutdown. 3882 */ 3883 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) { 3884 __mptcp_close_ssk(newsk, msk->first, 3885 mptcp_subflow_ctx(msk->first), 0); 3886 if (unlikely(list_is_singular(&msk->conn_list))) 3887 mptcp_set_state(newsk, TCP_CLOSE); 3888 } 3889 } else { 3890 tcpfallback: 3891 newsk->sk_kern_sock = arg->kern; 3892 lock_sock(newsk); 3893 __inet_accept(sock, newsock, newsk); 3894 /* we are being invoked after accepting a non-mp-capable 3895 * flow: sk is a tcp_sk, not an mptcp one. 3896 * 3897 * Hand the socket over to tcp so all further socket ops 3898 * bypass mptcp. 3899 */ 3900 WRITE_ONCE(newsock->sk->sk_socket->ops, 3901 mptcp_fallback_tcp_ops(newsock->sk)); 3902 } 3903 release_sock(newsk); 3904 3905 return 0; 3906 } 3907 3908 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 3909 { 3910 struct sock *sk = (struct sock *)msk; 3911 3912 if (__mptcp_stream_is_writeable(sk, 1)) 3913 return EPOLLOUT | EPOLLWRNORM; 3914 3915 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 3916 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */ 3917 if (__mptcp_stream_is_writeable(sk, 1)) 3918 return EPOLLOUT | EPOLLWRNORM; 3919 3920 return 0; 3921 } 3922 3923 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 3924 struct poll_table_struct *wait) 3925 { 3926 struct sock *sk = sock->sk; 3927 struct mptcp_sock *msk; 3928 __poll_t mask = 0; 3929 u8 shutdown; 3930 int state; 3931 3932 msk = mptcp_sk(sk); 3933 sock_poll_wait(file, sock, wait); 3934 3935 state = inet_sk_state_load(sk); 3936 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags); 3937 if (state == TCP_LISTEN) { 3938 struct sock *ssk = READ_ONCE(msk->first); 3939 3940 if (WARN_ON_ONCE(!ssk)) 3941 return 0; 3942 3943 return inet_csk_listen_poll(ssk); 3944 } 3945 3946 shutdown = READ_ONCE(sk->sk_shutdown); 3947 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 3948 mask |= EPOLLHUP; 3949 if (shutdown & RCV_SHUTDOWN) 3950 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 3951 3952 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 3953 mask |= mptcp_check_readable(sk); 3954 if (shutdown & SEND_SHUTDOWN) 3955 mask |= EPOLLOUT | EPOLLWRNORM; 3956 else 3957 mask |= mptcp_check_writeable(msk); 3958 } else if (state == TCP_SYN_SENT && 3959 inet_test_bit(DEFER_CONNECT, sk)) { 3960 /* cf tcp_poll() note about TFO */ 3961 mask |= EPOLLOUT | EPOLLWRNORM; 3962 } 3963 3964 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */ 3965 smp_rmb(); 3966 if (READ_ONCE(sk->sk_err)) 3967 mask |= EPOLLERR; 3968 3969 return mask; 3970 } 3971 3972 static const struct proto_ops mptcp_stream_ops = { 3973 .family = PF_INET, 3974 .owner = THIS_MODULE, 3975 .release = inet_release, 3976 .bind = mptcp_bind, 3977 .connect = inet_stream_connect, 3978 .socketpair = sock_no_socketpair, 3979 .accept = mptcp_stream_accept, 3980 .getname = inet_getname, 3981 .poll = mptcp_poll, 3982 .ioctl = inet_ioctl, 3983 .gettstamp = sock_gettstamp, 3984 .listen = mptcp_listen, 3985 .shutdown = inet_shutdown, 3986 .setsockopt = sock_common_setsockopt, 3987 .getsockopt = sock_common_getsockopt, 3988 .sendmsg = inet_sendmsg, 3989 .recvmsg = inet_recvmsg, 3990 .mmap = sock_no_mmap, 3991 .set_rcvlowat = mptcp_set_rcvlowat, 3992 }; 3993 3994 static struct inet_protosw mptcp_protosw = { 3995 .type = SOCK_STREAM, 3996 .protocol = IPPROTO_MPTCP, 3997 .prot = &mptcp_prot, 3998 .ops = &mptcp_stream_ops, 3999 .flags = INET_PROTOSW_ICSK, 4000 }; 4001 4002 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 4003 { 4004 struct mptcp_delegated_action *delegated; 4005 struct mptcp_subflow_context *subflow; 4006 int work_done = 0; 4007 4008 delegated = container_of(napi, struct mptcp_delegated_action, napi); 4009 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 4010 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4011 4012 bh_lock_sock_nested(ssk); 4013 if (!sock_owned_by_user(ssk)) { 4014 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0)); 4015 } else { 4016 /* tcp_release_cb_override already processed 4017 * the action or will do at next release_sock(). 4018 * In both case must dequeue the subflow here - on the same 4019 * CPU that scheduled it. 4020 */ 4021 smp_wmb(); 4022 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status); 4023 } 4024 bh_unlock_sock(ssk); 4025 sock_put(ssk); 4026 4027 if (++work_done == budget) 4028 return budget; 4029 } 4030 4031 /* always provide a 0 'work_done' argument, so that napi_complete_done 4032 * will not try accessing the NULL napi->dev ptr 4033 */ 4034 napi_complete_done(napi, 0); 4035 return work_done; 4036 } 4037 4038 void __init mptcp_proto_init(void) 4039 { 4040 struct mptcp_delegated_action *delegated; 4041 int cpu; 4042 4043 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 4044 4045 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 4046 panic("Failed to allocate MPTCP pcpu counter\n"); 4047 4048 mptcp_napi_dev = alloc_netdev_dummy(0); 4049 if (!mptcp_napi_dev) 4050 panic("Failed to allocate MPTCP dummy netdev\n"); 4051 for_each_possible_cpu(cpu) { 4052 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 4053 INIT_LIST_HEAD(&delegated->head); 4054 netif_napi_add_tx(mptcp_napi_dev, &delegated->napi, 4055 mptcp_napi_poll); 4056 napi_enable(&delegated->napi); 4057 } 4058 4059 mptcp_subflow_init(); 4060 mptcp_pm_init(); 4061 mptcp_sched_init(); 4062 mptcp_token_init(); 4063 4064 if (proto_register(&mptcp_prot, 1) != 0) 4065 panic("Failed to register MPTCP proto.\n"); 4066 4067 inet_register_protosw(&mptcp_protosw); 4068 4069 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 4070 } 4071 4072 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 4073 static const struct proto_ops mptcp_v6_stream_ops = { 4074 .family = PF_INET6, 4075 .owner = THIS_MODULE, 4076 .release = inet6_release, 4077 .bind = mptcp_bind, 4078 .connect = inet_stream_connect, 4079 .socketpair = sock_no_socketpair, 4080 .accept = mptcp_stream_accept, 4081 .getname = inet6_getname, 4082 .poll = mptcp_poll, 4083 .ioctl = inet6_ioctl, 4084 .gettstamp = sock_gettstamp, 4085 .listen = mptcp_listen, 4086 .shutdown = inet_shutdown, 4087 .setsockopt = sock_common_setsockopt, 4088 .getsockopt = sock_common_getsockopt, 4089 .sendmsg = inet6_sendmsg, 4090 .recvmsg = inet6_recvmsg, 4091 .mmap = sock_no_mmap, 4092 #ifdef CONFIG_COMPAT 4093 .compat_ioctl = inet6_compat_ioctl, 4094 #endif 4095 .set_rcvlowat = mptcp_set_rcvlowat, 4096 }; 4097 4098 static struct proto mptcp_v6_prot; 4099 4100 static struct inet_protosw mptcp_v6_protosw = { 4101 .type = SOCK_STREAM, 4102 .protocol = IPPROTO_MPTCP, 4103 .prot = &mptcp_v6_prot, 4104 .ops = &mptcp_v6_stream_ops, 4105 .flags = INET_PROTOSW_ICSK, 4106 }; 4107 4108 int __init mptcp_proto_v6_init(void) 4109 { 4110 int err; 4111 4112 mptcp_v6_prot = mptcp_prot; 4113 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name)); 4114 mptcp_v6_prot.slab = NULL; 4115 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 4116 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np); 4117 4118 err = proto_register(&mptcp_v6_prot, 1); 4119 if (err) 4120 return err; 4121 4122 err = inet6_register_protosw(&mptcp_v6_protosw); 4123 if (err) 4124 proto_unregister(&mptcp_v6_prot); 4125 4126 return err; 4127 } 4128 #endif 4129