1 // SPDX-License-Identifier: GPL-2.0 2 /* Multipath TCP 3 * 4 * Copyright (c) 2017 - 2019, Intel Corporation. 5 */ 6 7 #define pr_fmt(fmt) "MPTCP: " fmt 8 9 #include <linux/kernel.h> 10 #include <linux/module.h> 11 #include <linux/netdevice.h> 12 #include <linux/sched/signal.h> 13 #include <linux/atomic.h> 14 #include <net/aligned_data.h> 15 #include <net/rps.h> 16 #include <net/sock.h> 17 #include <net/inet_common.h> 18 #include <net/inet_hashtables.h> 19 #include <net/protocol.h> 20 #include <net/tcp_states.h> 21 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 22 #include <net/transp_v6.h> 23 #endif 24 #include <net/mptcp.h> 25 #include <net/hotdata.h> 26 #include <net/xfrm.h> 27 #include <asm/ioctls.h> 28 #include "protocol.h" 29 #include "mib.h" 30 31 #define CREATE_TRACE_POINTS 32 #include <trace/events/mptcp.h> 33 34 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 35 struct mptcp6_sock { 36 struct mptcp_sock msk; 37 struct ipv6_pinfo np; 38 }; 39 #endif 40 41 enum { 42 MPTCP_CMSG_TS = BIT(0), 43 MPTCP_CMSG_INQ = BIT(1), 44 }; 45 46 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp; 47 48 static void __mptcp_destroy_sock(struct sock *sk); 49 static void mptcp_check_send_data_fin(struct sock *sk); 50 51 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions) = { 52 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 53 }; 54 static struct net_device *mptcp_napi_dev; 55 56 /* Returns end sequence number of the receiver's advertised window */ 57 static u64 mptcp_wnd_end(const struct mptcp_sock *msk) 58 { 59 return READ_ONCE(msk->wnd_end); 60 } 61 62 static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk) 63 { 64 unsigned short family = READ_ONCE(sk->sk_family); 65 66 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 67 if (family == AF_INET6) 68 return &inet6_stream_ops; 69 #endif 70 WARN_ON_ONCE(family != AF_INET); 71 return &inet_stream_ops; 72 } 73 74 bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib) 75 { 76 struct net *net = sock_net((struct sock *)msk); 77 78 if (__mptcp_check_fallback(msk)) 79 return true; 80 81 /* The caller possibly is not holding the msk socket lock, but 82 * in the fallback case only the current subflow is touching 83 * the OoO queue. 84 */ 85 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue)) 86 return false; 87 88 spin_lock_bh(&msk->fallback_lock); 89 if (!msk->allow_infinite_fallback) { 90 spin_unlock_bh(&msk->fallback_lock); 91 return false; 92 } 93 94 msk->allow_subflows = false; 95 set_bit(MPTCP_FALLBACK_DONE, &msk->flags); 96 __MPTCP_INC_STATS(net, fb_mib); 97 spin_unlock_bh(&msk->fallback_lock); 98 return true; 99 } 100 101 static int __mptcp_socket_create(struct mptcp_sock *msk) 102 { 103 struct mptcp_subflow_context *subflow; 104 struct sock *sk = (struct sock *)msk; 105 struct socket *ssock; 106 int err; 107 108 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock); 109 if (err) 110 return err; 111 112 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio; 113 WRITE_ONCE(msk->first, ssock->sk); 114 subflow = mptcp_subflow_ctx(ssock->sk); 115 list_add(&subflow->node, &msk->conn_list); 116 sock_hold(ssock->sk); 117 subflow->request_mptcp = 1; 118 subflow->subflow_id = msk->subflow_id++; 119 120 /* This is the first subflow, always with id 0 */ 121 WRITE_ONCE(subflow->local_id, 0); 122 mptcp_sock_graft(msk->first, sk->sk_socket); 123 iput(SOCK_INODE(ssock)); 124 125 return 0; 126 } 127 128 /* If the MPC handshake is not started, returns the first subflow, 129 * eventually allocating it. 130 */ 131 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk) 132 { 133 struct sock *sk = (struct sock *)msk; 134 int ret; 135 136 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) 137 return ERR_PTR(-EINVAL); 138 139 if (!msk->first) { 140 ret = __mptcp_socket_create(msk); 141 if (ret) 142 return ERR_PTR(ret); 143 } 144 145 return msk->first; 146 } 147 148 static void mptcp_drop(struct sock *sk, struct sk_buff *skb) 149 { 150 sk_drops_skbadd(sk, skb); 151 __kfree_skb(skb); 152 } 153 154 static bool __mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 155 struct sk_buff *from, bool *fragstolen, 156 int *delta) 157 { 158 int limit = READ_ONCE(sk->sk_rcvbuf); 159 160 if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) || 161 MPTCP_SKB_CB(from)->offset || 162 ((to->len + from->len) > (limit >> 3)) || 163 !skb_try_coalesce(to, from, fragstolen, delta)) 164 return false; 165 166 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n", 167 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq, 168 to->len, MPTCP_SKB_CB(from)->end_seq); 169 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq; 170 return true; 171 } 172 173 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to, 174 struct sk_buff *from) 175 { 176 bool fragstolen; 177 int delta; 178 179 if (!__mptcp_try_coalesce(sk, to, from, &fragstolen, &delta)) 180 return false; 181 182 /* note the fwd memory can reach a negative value after accounting 183 * for the delta, but the later skb free will restore a non 184 * negative one 185 */ 186 atomic_add(delta, &sk->sk_rmem_alloc); 187 sk_mem_charge(sk, delta); 188 kfree_skb_partial(from, fragstolen); 189 190 return true; 191 } 192 193 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to, 194 struct sk_buff *from) 195 { 196 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq) 197 return false; 198 199 return mptcp_try_coalesce((struct sock *)msk, to, from); 200 } 201 202 /* "inspired" by tcp_rcvbuf_grow(), main difference: 203 * - mptcp does not maintain a msk-level window clamp 204 * - returns true when the receive buffer is actually updated 205 */ 206 static bool mptcp_rcvbuf_grow(struct sock *sk, u32 newval) 207 { 208 struct mptcp_sock *msk = mptcp_sk(sk); 209 const struct net *net = sock_net(sk); 210 u32 rcvwin, rcvbuf, cap, oldval; 211 u64 grow; 212 213 oldval = msk->rcvq_space.space; 214 msk->rcvq_space.space = newval; 215 if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) || 216 (sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 217 return false; 218 219 /* DRS is always one RTT late. */ 220 rcvwin = newval << 1; 221 222 /* slow start: allow the sender to double its rate. */ 223 grow = (u64)rcvwin * (newval - oldval); 224 do_div(grow, oldval); 225 rcvwin += grow << 1; 226 227 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue)) 228 rcvwin += MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq - msk->ack_seq; 229 230 cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 231 232 rcvbuf = min_t(u32, mptcp_space_from_win(sk, rcvwin), cap); 233 if (rcvbuf > sk->sk_rcvbuf) { 234 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 235 return true; 236 } 237 return false; 238 } 239 240 /* "inspired" by tcp_data_queue_ofo(), main differences: 241 * - use mptcp seqs 242 * - don't cope with sacks 243 */ 244 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb) 245 { 246 struct sock *sk = (struct sock *)msk; 247 struct rb_node **p, *parent; 248 u64 seq, end_seq, max_seq; 249 struct sk_buff *skb1; 250 251 seq = MPTCP_SKB_CB(skb)->map_seq; 252 end_seq = MPTCP_SKB_CB(skb)->end_seq; 253 max_seq = atomic64_read(&msk->rcv_wnd_sent); 254 255 pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq, 256 RB_EMPTY_ROOT(&msk->out_of_order_queue)); 257 if (after64(end_seq, max_seq)) { 258 /* out of window */ 259 mptcp_drop(sk, skb); 260 pr_debug("oow by %lld, rcv_wnd_sent %llu\n", 261 (unsigned long long)end_seq - (unsigned long)max_seq, 262 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent)); 263 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW); 264 return; 265 } 266 267 p = &msk->out_of_order_queue.rb_node; 268 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE); 269 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) { 270 rb_link_node(&skb->rbnode, NULL, p); 271 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 272 msk->ooo_last_skb = skb; 273 goto end; 274 } 275 276 /* with 2 subflows, adding at end of ooo queue is quite likely 277 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 278 */ 279 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) { 280 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 281 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 282 return; 283 } 284 285 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 286 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) { 287 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL); 288 parent = &msk->ooo_last_skb->rbnode; 289 p = &parent->rb_right; 290 goto insert; 291 } 292 293 /* Find place to insert this segment. Handle overlaps on the way. */ 294 parent = NULL; 295 while (*p) { 296 parent = *p; 297 skb1 = rb_to_skb(parent); 298 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 299 p = &parent->rb_left; 300 continue; 301 } 302 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) { 303 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) { 304 /* All the bits are present. Drop. */ 305 mptcp_drop(sk, skb); 306 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 307 return; 308 } 309 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) { 310 /* partial overlap: 311 * | skb | 312 * | skb1 | 313 * continue traversing 314 */ 315 } else { 316 /* skb's seq == skb1's seq and skb covers skb1. 317 * Replace skb1 with skb. 318 */ 319 rb_replace_node(&skb1->rbnode, &skb->rbnode, 320 &msk->out_of_order_queue); 321 mptcp_drop(sk, skb1); 322 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 323 goto merge_right; 324 } 325 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) { 326 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE); 327 return; 328 } 329 p = &parent->rb_right; 330 } 331 332 insert: 333 /* Insert segment into RB tree. */ 334 rb_link_node(&skb->rbnode, parent, p); 335 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue); 336 337 merge_right: 338 /* Remove other segments covered by skb. */ 339 while ((skb1 = skb_rb_next(skb)) != NULL) { 340 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) 341 break; 342 rb_erase(&skb1->rbnode, &msk->out_of_order_queue); 343 mptcp_drop(sk, skb1); 344 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 345 } 346 /* If there is no skb after us, we are the last_skb ! */ 347 if (!skb1) 348 msk->ooo_last_skb = skb; 349 350 end: 351 skb_condense(skb); 352 skb_set_owner_r(skb, sk); 353 /* do not grow rcvbuf for not-yet-accepted or orphaned sockets. */ 354 if (sk->sk_socket) 355 mptcp_rcvbuf_grow(sk, msk->rcvq_space.space); 356 } 357 358 static void mptcp_init_skb(struct sock *ssk, struct sk_buff *skb, int offset, 359 int copy_len) 360 { 361 const struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 362 bool has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp; 363 364 /* the skb map_seq accounts for the skb offset: 365 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq 366 * value 367 */ 368 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow); 369 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len; 370 MPTCP_SKB_CB(skb)->offset = offset; 371 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp; 372 MPTCP_SKB_CB(skb)->cant_coalesce = 0; 373 374 __skb_unlink(skb, &ssk->sk_receive_queue); 375 376 skb_ext_reset(skb); 377 skb_dst_drop(skb); 378 } 379 380 static bool __mptcp_move_skb(struct sock *sk, struct sk_buff *skb) 381 { 382 u64 copy_len = MPTCP_SKB_CB(skb)->end_seq - MPTCP_SKB_CB(skb)->map_seq; 383 struct mptcp_sock *msk = mptcp_sk(sk); 384 struct sk_buff *tail; 385 386 /* try to fetch required memory from subflow */ 387 if (!sk_rmem_schedule(sk, skb, skb->truesize)) { 388 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED); 389 goto drop; 390 } 391 392 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) { 393 /* in sequence */ 394 msk->bytes_received += copy_len; 395 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len); 396 tail = skb_peek_tail(&sk->sk_receive_queue); 397 if (tail && mptcp_try_coalesce(sk, tail, skb)) 398 return true; 399 400 skb_set_owner_r(skb, sk); 401 __skb_queue_tail(&sk->sk_receive_queue, skb); 402 return true; 403 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) { 404 mptcp_data_queue_ofo(msk, skb); 405 return false; 406 } 407 408 /* old data, keep it simple and drop the whole pkt, sender 409 * will retransmit as needed, if needed. 410 */ 411 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 412 drop: 413 mptcp_drop(sk, skb); 414 return false; 415 } 416 417 static void mptcp_stop_rtx_timer(struct sock *sk) 418 { 419 struct inet_connection_sock *icsk = inet_csk(sk); 420 421 sk_stop_timer(sk, &icsk->icsk_retransmit_timer); 422 mptcp_sk(sk)->timer_ival = 0; 423 } 424 425 static void mptcp_close_wake_up(struct sock *sk) 426 { 427 if (sock_flag(sk, SOCK_DEAD)) 428 return; 429 430 sk->sk_state_change(sk); 431 if (sk->sk_shutdown == SHUTDOWN_MASK || 432 sk->sk_state == TCP_CLOSE) 433 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 434 else 435 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 436 } 437 438 static void mptcp_shutdown_subflows(struct mptcp_sock *msk) 439 { 440 struct mptcp_subflow_context *subflow; 441 442 mptcp_for_each_subflow(msk, subflow) { 443 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 444 bool slow; 445 446 slow = lock_sock_fast(ssk); 447 tcp_shutdown(ssk, SEND_SHUTDOWN); 448 unlock_sock_fast(ssk, slow); 449 } 450 } 451 452 /* called under the msk socket lock */ 453 static bool mptcp_pending_data_fin_ack(struct sock *sk) 454 { 455 struct mptcp_sock *msk = mptcp_sk(sk); 456 457 return ((1 << sk->sk_state) & 458 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) && 459 msk->write_seq == READ_ONCE(msk->snd_una); 460 } 461 462 static void mptcp_check_data_fin_ack(struct sock *sk) 463 { 464 struct mptcp_sock *msk = mptcp_sk(sk); 465 466 /* Look for an acknowledged DATA_FIN */ 467 if (mptcp_pending_data_fin_ack(sk)) { 468 WRITE_ONCE(msk->snd_data_fin_enable, 0); 469 470 switch (sk->sk_state) { 471 case TCP_FIN_WAIT1: 472 mptcp_set_state(sk, TCP_FIN_WAIT2); 473 break; 474 case TCP_CLOSING: 475 case TCP_LAST_ACK: 476 mptcp_shutdown_subflows(msk); 477 mptcp_set_state(sk, TCP_CLOSE); 478 break; 479 } 480 481 mptcp_close_wake_up(sk); 482 } 483 } 484 485 /* can be called with no lock acquired */ 486 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq) 487 { 488 struct mptcp_sock *msk = mptcp_sk(sk); 489 490 if (READ_ONCE(msk->rcv_data_fin) && 491 ((1 << inet_sk_state_load(sk)) & 492 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) { 493 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq); 494 495 if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) { 496 if (seq) 497 *seq = rcv_data_fin_seq; 498 499 return true; 500 } 501 } 502 503 return false; 504 } 505 506 static void mptcp_set_datafin_timeout(struct sock *sk) 507 { 508 struct inet_connection_sock *icsk = inet_csk(sk); 509 u32 retransmits; 510 511 retransmits = min_t(u32, icsk->icsk_retransmits, 512 ilog2(TCP_RTO_MAX / TCP_RTO_MIN)); 513 514 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits; 515 } 516 517 static void __mptcp_set_timeout(struct sock *sk, long tout) 518 { 519 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN; 520 } 521 522 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow) 523 { 524 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 525 526 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ? 527 icsk_timeout(inet_csk(ssk)) - jiffies : 0; 528 } 529 530 static void mptcp_set_timeout(struct sock *sk) 531 { 532 struct mptcp_subflow_context *subflow; 533 long tout = 0; 534 535 mptcp_for_each_subflow(mptcp_sk(sk), subflow) 536 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 537 __mptcp_set_timeout(sk, tout); 538 } 539 540 static inline bool tcp_can_send_ack(const struct sock *ssk) 541 { 542 return !((1 << inet_sk_state_load(ssk)) & 543 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN)); 544 } 545 546 void __mptcp_subflow_send_ack(struct sock *ssk) 547 { 548 if (tcp_can_send_ack(ssk)) 549 tcp_send_ack(ssk); 550 } 551 552 static void mptcp_subflow_send_ack(struct sock *ssk) 553 { 554 bool slow; 555 556 slow = lock_sock_fast(ssk); 557 __mptcp_subflow_send_ack(ssk); 558 unlock_sock_fast(ssk, slow); 559 } 560 561 static void mptcp_send_ack(struct mptcp_sock *msk) 562 { 563 struct mptcp_subflow_context *subflow; 564 565 mptcp_for_each_subflow(msk, subflow) 566 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow)); 567 } 568 569 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied) 570 { 571 bool slow; 572 573 slow = lock_sock_fast(ssk); 574 if (tcp_can_send_ack(ssk)) 575 tcp_cleanup_rbuf(ssk, copied); 576 unlock_sock_fast(ssk, slow); 577 } 578 579 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty) 580 { 581 const struct inet_connection_sock *icsk = inet_csk(ssk); 582 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending); 583 const struct tcp_sock *tp = tcp_sk(ssk); 584 585 return (ack_pending & ICSK_ACK_SCHED) && 586 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) > 587 READ_ONCE(icsk->icsk_ack.rcv_mss)) || 588 (rx_empty && ack_pending & 589 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED))); 590 } 591 592 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied) 593 { 594 int old_space = READ_ONCE(msk->old_wspace); 595 struct mptcp_subflow_context *subflow; 596 struct sock *sk = (struct sock *)msk; 597 int space = __mptcp_space(sk); 598 bool cleanup, rx_empty; 599 600 cleanup = (space > 0) && (space >= (old_space << 1)) && copied; 601 rx_empty = !sk_rmem_alloc_get(sk) && copied; 602 603 mptcp_for_each_subflow(msk, subflow) { 604 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 605 606 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty)) 607 mptcp_subflow_cleanup_rbuf(ssk, copied); 608 } 609 } 610 611 static void mptcp_check_data_fin(struct sock *sk) 612 { 613 struct mptcp_sock *msk = mptcp_sk(sk); 614 u64 rcv_data_fin_seq; 615 616 /* Need to ack a DATA_FIN received from a peer while this side 617 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2. 618 * msk->rcv_data_fin was set when parsing the incoming options 619 * at the subflow level and the msk lock was not held, so this 620 * is the first opportunity to act on the DATA_FIN and change 621 * the msk state. 622 * 623 * If we are caught up to the sequence number of the incoming 624 * DATA_FIN, send the DATA_ACK now and do state transition. If 625 * not caught up, do nothing and let the recv code send DATA_ACK 626 * when catching up. 627 */ 628 629 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) { 630 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1); 631 WRITE_ONCE(msk->rcv_data_fin, 0); 632 633 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 634 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 635 636 switch (sk->sk_state) { 637 case TCP_ESTABLISHED: 638 mptcp_set_state(sk, TCP_CLOSE_WAIT); 639 break; 640 case TCP_FIN_WAIT1: 641 mptcp_set_state(sk, TCP_CLOSING); 642 break; 643 case TCP_FIN_WAIT2: 644 mptcp_shutdown_subflows(msk); 645 mptcp_set_state(sk, TCP_CLOSE); 646 break; 647 default: 648 /* Other states not expected */ 649 WARN_ON_ONCE(1); 650 break; 651 } 652 653 if (!__mptcp_check_fallback(msk)) 654 mptcp_send_ack(msk); 655 mptcp_close_wake_up(sk); 656 } 657 } 658 659 static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk) 660 { 661 if (!mptcp_try_fallback(ssk, MPTCP_MIB_DSSCORRUPTIONFALLBACK)) { 662 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET); 663 mptcp_subflow_reset(ssk); 664 } 665 } 666 667 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk, 668 struct sock *ssk) 669 { 670 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 671 struct sock *sk = (struct sock *)msk; 672 bool more_data_avail; 673 struct tcp_sock *tp; 674 bool ret = false; 675 676 pr_debug("msk=%p ssk=%p\n", msk, ssk); 677 tp = tcp_sk(ssk); 678 do { 679 u32 map_remaining, offset; 680 u32 seq = tp->copied_seq; 681 struct sk_buff *skb; 682 bool fin; 683 684 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 685 break; 686 687 /* try to move as much data as available */ 688 map_remaining = subflow->map_data_len - 689 mptcp_subflow_get_map_offset(subflow); 690 691 skb = skb_peek(&ssk->sk_receive_queue); 692 if (unlikely(!skb)) 693 break; 694 695 if (__mptcp_check_fallback(msk)) { 696 /* Under fallback skbs have no MPTCP extension and TCP could 697 * collapse them between the dummy map creation and the 698 * current dequeue. Be sure to adjust the map size. 699 */ 700 map_remaining = skb->len; 701 subflow->map_data_len = skb->len; 702 } 703 704 offset = seq - TCP_SKB_CB(skb)->seq; 705 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 706 if (fin) 707 seq++; 708 709 if (offset < skb->len) { 710 size_t len = skb->len - offset; 711 712 mptcp_init_skb(ssk, skb, offset, len); 713 skb_orphan(skb); 714 ret = __mptcp_move_skb(sk, skb) || ret; 715 seq += len; 716 717 if (unlikely(map_remaining < len)) { 718 DEBUG_NET_WARN_ON_ONCE(1); 719 mptcp_dss_corruption(msk, ssk); 720 } 721 } else { 722 if (unlikely(!fin)) { 723 DEBUG_NET_WARN_ON_ONCE(1); 724 mptcp_dss_corruption(msk, ssk); 725 } 726 727 sk_eat_skb(ssk, skb); 728 } 729 730 WRITE_ONCE(tp->copied_seq, seq); 731 more_data_avail = mptcp_subflow_data_available(ssk); 732 733 } while (more_data_avail); 734 735 if (ret) 736 msk->last_data_recv = tcp_jiffies32; 737 return ret; 738 } 739 740 static bool __mptcp_ofo_queue(struct mptcp_sock *msk) 741 { 742 struct sock *sk = (struct sock *)msk; 743 struct sk_buff *skb, *tail; 744 bool moved = false; 745 struct rb_node *p; 746 u64 end_seq; 747 748 p = rb_first(&msk->out_of_order_queue); 749 pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue)); 750 while (p) { 751 skb = rb_to_skb(p); 752 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) 753 break; 754 755 p = rb_next(p); 756 rb_erase(&skb->rbnode, &msk->out_of_order_queue); 757 758 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq, 759 msk->ack_seq))) { 760 mptcp_drop(sk, skb); 761 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA); 762 continue; 763 } 764 765 end_seq = MPTCP_SKB_CB(skb)->end_seq; 766 tail = skb_peek_tail(&sk->sk_receive_queue); 767 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) { 768 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq; 769 770 /* skip overlapping data, if any */ 771 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n", 772 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq, 773 delta); 774 MPTCP_SKB_CB(skb)->offset += delta; 775 MPTCP_SKB_CB(skb)->map_seq += delta; 776 __skb_queue_tail(&sk->sk_receive_queue, skb); 777 } 778 msk->bytes_received += end_seq - msk->ack_seq; 779 WRITE_ONCE(msk->ack_seq, end_seq); 780 moved = true; 781 } 782 return moved; 783 } 784 785 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk) 786 { 787 int err = sock_error(ssk); 788 int ssk_state; 789 790 if (!err) 791 return false; 792 793 /* only propagate errors on fallen-back sockets or 794 * on MPC connect 795 */ 796 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk))) 797 return false; 798 799 /* We need to propagate only transition to CLOSE state. 800 * Orphaned socket will see such state change via 801 * subflow_sched_work_if_closed() and that path will properly 802 * destroy the msk as needed. 803 */ 804 ssk_state = inet_sk_state_load(ssk); 805 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD)) 806 mptcp_set_state(sk, ssk_state); 807 WRITE_ONCE(sk->sk_err, -err); 808 809 /* This barrier is coupled with smp_rmb() in mptcp_poll() */ 810 smp_wmb(); 811 sk_error_report(sk); 812 return true; 813 } 814 815 void __mptcp_error_report(struct sock *sk) 816 { 817 struct mptcp_subflow_context *subflow; 818 struct mptcp_sock *msk = mptcp_sk(sk); 819 820 mptcp_for_each_subflow(msk, subflow) 821 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow))) 822 break; 823 } 824 825 /* In most cases we will be able to lock the mptcp socket. If its already 826 * owned, we need to defer to the work queue to avoid ABBA deadlock. 827 */ 828 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk) 829 { 830 struct sock *sk = (struct sock *)msk; 831 bool moved; 832 833 moved = __mptcp_move_skbs_from_subflow(msk, ssk); 834 __mptcp_ofo_queue(msk); 835 if (unlikely(ssk->sk_err)) 836 __mptcp_subflow_error_report(sk, ssk); 837 838 /* If the moves have caught up with the DATA_FIN sequence number 839 * it's time to ack the DATA_FIN and change socket state, but 840 * this is not a good place to change state. Let the workqueue 841 * do it. 842 */ 843 if (mptcp_pending_data_fin(sk, NULL)) 844 mptcp_schedule_work(sk); 845 return moved; 846 } 847 848 static void __mptcp_data_ready(struct sock *sk, struct sock *ssk) 849 { 850 struct mptcp_sock *msk = mptcp_sk(sk); 851 852 /* Wake-up the reader only for in-sequence data */ 853 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk)) 854 sk->sk_data_ready(sk); 855 } 856 857 void mptcp_data_ready(struct sock *sk, struct sock *ssk) 858 { 859 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 860 861 /* The peer can send data while we are shutting down this 862 * subflow at msk destruction time, but we must avoid enqueuing 863 * more data to the msk receive queue 864 */ 865 if (unlikely(subflow->disposable)) 866 return; 867 868 mptcp_data_lock(sk); 869 if (!sock_owned_by_user(sk)) 870 __mptcp_data_ready(sk, ssk); 871 else 872 __set_bit(MPTCP_DEQUEUE, &mptcp_sk(sk)->cb_flags); 873 mptcp_data_unlock(sk); 874 } 875 876 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk) 877 { 878 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq); 879 msk->allow_infinite_fallback = false; 880 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC); 881 } 882 883 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk) 884 { 885 struct sock *sk = (struct sock *)msk; 886 887 if (sk->sk_state != TCP_ESTABLISHED) 888 return false; 889 890 spin_lock_bh(&msk->fallback_lock); 891 if (!msk->allow_subflows) { 892 spin_unlock_bh(&msk->fallback_lock); 893 return false; 894 } 895 mptcp_subflow_joined(msk, ssk); 896 spin_unlock_bh(&msk->fallback_lock); 897 898 /* attach to msk socket only after we are sure we will deal with it 899 * at close time 900 */ 901 if (sk->sk_socket && !ssk->sk_socket) 902 mptcp_sock_graft(ssk, sk->sk_socket); 903 904 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++; 905 mptcp_sockopt_sync_locked(msk, ssk); 906 mptcp_stop_tout_timer(sk); 907 __mptcp_propagate_sndbuf(sk, ssk); 908 return true; 909 } 910 911 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list) 912 { 913 struct mptcp_subflow_context *tmp, *subflow; 914 struct mptcp_sock *msk = mptcp_sk(sk); 915 916 list_for_each_entry_safe(subflow, tmp, join_list, node) { 917 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 918 bool slow = lock_sock_fast(ssk); 919 920 list_move_tail(&subflow->node, &msk->conn_list); 921 if (!__mptcp_finish_join(msk, ssk)) 922 mptcp_subflow_reset(ssk); 923 unlock_sock_fast(ssk, slow); 924 } 925 } 926 927 static bool mptcp_rtx_timer_pending(struct sock *sk) 928 { 929 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer); 930 } 931 932 static void mptcp_reset_rtx_timer(struct sock *sk) 933 { 934 struct inet_connection_sock *icsk = inet_csk(sk); 935 unsigned long tout; 936 937 /* prevent rescheduling on close */ 938 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE)) 939 return; 940 941 tout = mptcp_sk(sk)->timer_ival; 942 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout); 943 } 944 945 bool mptcp_schedule_work(struct sock *sk) 946 { 947 if (inet_sk_state_load(sk) == TCP_CLOSE) 948 return false; 949 950 /* Get a reference on this socket, mptcp_worker() will release it. 951 * As mptcp_worker() might complete before us, we can not avoid 952 * a sock_hold()/sock_put() if schedule_work() returns false. 953 */ 954 sock_hold(sk); 955 956 if (schedule_work(&mptcp_sk(sk)->work)) 957 return true; 958 959 sock_put(sk); 960 return false; 961 } 962 963 static bool mptcp_skb_can_collapse_to(u64 write_seq, 964 const struct sk_buff *skb, 965 const struct mptcp_ext *mpext) 966 { 967 if (!tcp_skb_can_collapse_to(skb)) 968 return false; 969 970 /* can collapse only if MPTCP level sequence is in order and this 971 * mapping has not been xmitted yet 972 */ 973 return mpext && mpext->data_seq + mpext->data_len == write_seq && 974 !mpext->frozen; 975 } 976 977 /* we can append data to the given data frag if: 978 * - there is space available in the backing page_frag 979 * - the data frag tail matches the current page_frag free offset 980 * - the data frag end sequence number matches the current write seq 981 */ 982 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk, 983 const struct page_frag *pfrag, 984 const struct mptcp_data_frag *df) 985 { 986 return df && pfrag->page == df->page && 987 pfrag->size - pfrag->offset > 0 && 988 pfrag->offset == (df->offset + df->data_len) && 989 df->data_seq + df->data_len == msk->write_seq; 990 } 991 992 static void dfrag_uncharge(struct sock *sk, int len) 993 { 994 sk_mem_uncharge(sk, len); 995 sk_wmem_queued_add(sk, -len); 996 } 997 998 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag) 999 { 1000 int len = dfrag->data_len + dfrag->overhead; 1001 1002 list_del(&dfrag->list); 1003 dfrag_uncharge(sk, len); 1004 put_page(dfrag->page); 1005 } 1006 1007 /* called under both the msk socket lock and the data lock */ 1008 static void __mptcp_clean_una(struct sock *sk) 1009 { 1010 struct mptcp_sock *msk = mptcp_sk(sk); 1011 struct mptcp_data_frag *dtmp, *dfrag; 1012 u64 snd_una; 1013 1014 snd_una = msk->snd_una; 1015 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) { 1016 if (after64(dfrag->data_seq + dfrag->data_len, snd_una)) 1017 break; 1018 1019 if (unlikely(dfrag == msk->first_pending)) { 1020 /* in recovery mode can see ack after the current snd head */ 1021 if (WARN_ON_ONCE(!msk->recovery)) 1022 break; 1023 1024 msk->first_pending = mptcp_send_next(sk); 1025 } 1026 1027 dfrag_clear(sk, dfrag); 1028 } 1029 1030 dfrag = mptcp_rtx_head(sk); 1031 if (dfrag && after64(snd_una, dfrag->data_seq)) { 1032 u64 delta = snd_una - dfrag->data_seq; 1033 1034 /* prevent wrap around in recovery mode */ 1035 if (unlikely(delta > dfrag->already_sent)) { 1036 if (WARN_ON_ONCE(!msk->recovery)) 1037 goto out; 1038 if (WARN_ON_ONCE(delta > dfrag->data_len)) 1039 goto out; 1040 dfrag->already_sent += delta - dfrag->already_sent; 1041 } 1042 1043 dfrag->data_seq += delta; 1044 dfrag->offset += delta; 1045 dfrag->data_len -= delta; 1046 dfrag->already_sent -= delta; 1047 1048 dfrag_uncharge(sk, delta); 1049 } 1050 1051 /* all retransmitted data acked, recovery completed */ 1052 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt)) 1053 msk->recovery = false; 1054 1055 out: 1056 if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) { 1057 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk)) 1058 mptcp_stop_rtx_timer(sk); 1059 } else { 1060 mptcp_reset_rtx_timer(sk); 1061 } 1062 1063 if (mptcp_pending_data_fin_ack(sk)) 1064 mptcp_schedule_work(sk); 1065 } 1066 1067 static void __mptcp_clean_una_wakeup(struct sock *sk) 1068 { 1069 lockdep_assert_held_once(&sk->sk_lock.slock); 1070 1071 __mptcp_clean_una(sk); 1072 mptcp_write_space(sk); 1073 } 1074 1075 static void mptcp_clean_una_wakeup(struct sock *sk) 1076 { 1077 mptcp_data_lock(sk); 1078 __mptcp_clean_una_wakeup(sk); 1079 mptcp_data_unlock(sk); 1080 } 1081 1082 static void mptcp_enter_memory_pressure(struct sock *sk) 1083 { 1084 struct mptcp_subflow_context *subflow; 1085 struct mptcp_sock *msk = mptcp_sk(sk); 1086 bool first = true; 1087 1088 mptcp_for_each_subflow(msk, subflow) { 1089 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1090 1091 if (first) 1092 tcp_enter_memory_pressure(ssk); 1093 sk_stream_moderate_sndbuf(ssk); 1094 1095 first = false; 1096 } 1097 __mptcp_sync_sndbuf(sk); 1098 } 1099 1100 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1101 * data 1102 */ 1103 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1104 { 1105 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1106 pfrag, sk->sk_allocation))) 1107 return true; 1108 1109 mptcp_enter_memory_pressure(sk); 1110 return false; 1111 } 1112 1113 static struct mptcp_data_frag * 1114 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1115 int orig_offset) 1116 { 1117 int offset = ALIGN(orig_offset, sizeof(long)); 1118 struct mptcp_data_frag *dfrag; 1119 1120 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1121 dfrag->data_len = 0; 1122 dfrag->data_seq = msk->write_seq; 1123 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1124 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1125 dfrag->already_sent = 0; 1126 dfrag->page = pfrag->page; 1127 1128 return dfrag; 1129 } 1130 1131 struct mptcp_sendmsg_info { 1132 int mss_now; 1133 int size_goal; 1134 u16 limit; 1135 u16 sent; 1136 unsigned int flags; 1137 bool data_lock_held; 1138 }; 1139 1140 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk, 1141 u64 data_seq, int avail_size) 1142 { 1143 u64 window_end = mptcp_wnd_end(msk); 1144 u64 mptcp_snd_wnd; 1145 1146 if (__mptcp_check_fallback(msk)) 1147 return avail_size; 1148 1149 mptcp_snd_wnd = window_end - data_seq; 1150 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size); 1151 1152 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) { 1153 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd); 1154 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED); 1155 } 1156 1157 return avail_size; 1158 } 1159 1160 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1161 { 1162 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1163 1164 if (!mpext) 1165 return false; 1166 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1167 return true; 1168 } 1169 1170 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1171 { 1172 struct sk_buff *skb; 1173 1174 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1175 if (likely(skb)) { 1176 if (likely(__mptcp_add_ext(skb, gfp))) { 1177 skb_reserve(skb, MAX_TCP_HEADER); 1178 skb->ip_summed = CHECKSUM_PARTIAL; 1179 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1180 return skb; 1181 } 1182 __kfree_skb(skb); 1183 } else { 1184 mptcp_enter_memory_pressure(sk); 1185 } 1186 return NULL; 1187 } 1188 1189 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1190 { 1191 struct sk_buff *skb; 1192 1193 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1194 if (!skb) 1195 return NULL; 1196 1197 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1198 tcp_skb_entail(ssk, skb); 1199 return skb; 1200 } 1201 tcp_skb_tsorted_anchor_cleanup(skb); 1202 kfree_skb(skb); 1203 return NULL; 1204 } 1205 1206 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1207 { 1208 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1209 1210 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1211 } 1212 1213 /* note: this always recompute the csum on the whole skb, even 1214 * if we just appended a single frag. More status info needed 1215 */ 1216 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1217 { 1218 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1219 __wsum csum = ~csum_unfold(mpext->csum); 1220 int offset = skb->len - added; 1221 1222 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1223 } 1224 1225 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1226 struct sock *ssk, 1227 struct mptcp_ext *mpext) 1228 { 1229 if (!mpext) 1230 return; 1231 1232 mpext->infinite_map = 1; 1233 mpext->data_len = 0; 1234 1235 if (!mptcp_try_fallback(ssk, MPTCP_MIB_INFINITEMAPTX)) { 1236 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED); 1237 mptcp_subflow_reset(ssk); 1238 return; 1239 } 1240 1241 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1242 } 1243 1244 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1)) 1245 1246 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1247 struct mptcp_data_frag *dfrag, 1248 struct mptcp_sendmsg_info *info) 1249 { 1250 u64 data_seq = dfrag->data_seq + info->sent; 1251 int offset = dfrag->offset + info->sent; 1252 struct mptcp_sock *msk = mptcp_sk(sk); 1253 bool zero_window_probe = false; 1254 struct mptcp_ext *mpext = NULL; 1255 bool can_coalesce = false; 1256 bool reuse_skb = true; 1257 struct sk_buff *skb; 1258 size_t copy; 1259 int i; 1260 1261 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n", 1262 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1263 1264 if (WARN_ON_ONCE(info->sent > info->limit || 1265 info->limit > dfrag->data_len)) 1266 return 0; 1267 1268 if (unlikely(!__tcp_can_send(ssk))) 1269 return -EAGAIN; 1270 1271 /* compute send limit */ 1272 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE)) 1273 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE; 1274 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1275 copy = info->size_goal; 1276 1277 skb = tcp_write_queue_tail(ssk); 1278 if (skb && copy > skb->len) { 1279 /* Limit the write to the size available in the 1280 * current skb, if any, so that we create at most a new skb. 1281 * Explicitly tells TCP internals to avoid collapsing on later 1282 * queue management operation, to avoid breaking the ext <-> 1283 * SSN association set here 1284 */ 1285 mpext = mptcp_get_ext(skb); 1286 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1287 TCP_SKB_CB(skb)->eor = 1; 1288 tcp_mark_push(tcp_sk(ssk), skb); 1289 goto alloc_skb; 1290 } 1291 1292 i = skb_shinfo(skb)->nr_frags; 1293 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1294 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) { 1295 tcp_mark_push(tcp_sk(ssk), skb); 1296 goto alloc_skb; 1297 } 1298 1299 copy -= skb->len; 1300 } else { 1301 alloc_skb: 1302 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1303 if (!skb) 1304 return -ENOMEM; 1305 1306 i = skb_shinfo(skb)->nr_frags; 1307 reuse_skb = false; 1308 mpext = mptcp_get_ext(skb); 1309 } 1310 1311 /* Zero window and all data acked? Probe. */ 1312 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy); 1313 if (copy == 0) { 1314 u64 snd_una = READ_ONCE(msk->snd_una); 1315 1316 /* No need for zero probe if there are any data pending 1317 * either at the msk or ssk level; skb is the current write 1318 * queue tail and can be empty at this point. 1319 */ 1320 if (snd_una != msk->snd_nxt || skb->len || 1321 skb != tcp_send_head(ssk)) { 1322 tcp_remove_empty_skb(ssk); 1323 return 0; 1324 } 1325 1326 zero_window_probe = true; 1327 data_seq = snd_una - 1; 1328 copy = 1; 1329 } 1330 1331 copy = min_t(size_t, copy, info->limit - info->sent); 1332 if (!sk_wmem_schedule(ssk, copy)) { 1333 tcp_remove_empty_skb(ssk); 1334 return -ENOMEM; 1335 } 1336 1337 if (can_coalesce) { 1338 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1339 } else { 1340 get_page(dfrag->page); 1341 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1342 } 1343 1344 skb->len += copy; 1345 skb->data_len += copy; 1346 skb->truesize += copy; 1347 sk_wmem_queued_add(ssk, copy); 1348 sk_mem_charge(ssk, copy); 1349 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1350 TCP_SKB_CB(skb)->end_seq += copy; 1351 tcp_skb_pcount_set(skb, 0); 1352 1353 /* on skb reuse we just need to update the DSS len */ 1354 if (reuse_skb) { 1355 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1356 mpext->data_len += copy; 1357 goto out; 1358 } 1359 1360 memset(mpext, 0, sizeof(*mpext)); 1361 mpext->data_seq = data_seq; 1362 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1363 mpext->data_len = copy; 1364 mpext->use_map = 1; 1365 mpext->dsn64 = 1; 1366 1367 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n", 1368 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1369 mpext->dsn64); 1370 1371 if (zero_window_probe) { 1372 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_WINPROBE); 1373 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1374 mpext->frozen = 1; 1375 if (READ_ONCE(msk->csum_enabled)) 1376 mptcp_update_data_checksum(skb, copy); 1377 tcp_push_pending_frames(ssk); 1378 return 0; 1379 } 1380 out: 1381 if (READ_ONCE(msk->csum_enabled)) 1382 mptcp_update_data_checksum(skb, copy); 1383 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1384 mptcp_update_infinite_map(msk, ssk, mpext); 1385 trace_mptcp_sendmsg_frag(mpext); 1386 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1387 return copy; 1388 } 1389 1390 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1391 sizeof(struct tcphdr) - \ 1392 MAX_TCP_OPTION_SPACE - \ 1393 sizeof(struct ipv6hdr) - \ 1394 sizeof(struct frag_hdr)) 1395 1396 struct subflow_send_info { 1397 struct sock *ssk; 1398 u64 linger_time; 1399 }; 1400 1401 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1402 { 1403 if (!subflow->stale) 1404 return; 1405 1406 subflow->stale = 0; 1407 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1408 } 1409 1410 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1411 { 1412 if (unlikely(subflow->stale)) { 1413 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1414 1415 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1416 return false; 1417 1418 mptcp_subflow_set_active(subflow); 1419 } 1420 return __mptcp_subflow_active(subflow); 1421 } 1422 1423 #define SSK_MODE_ACTIVE 0 1424 #define SSK_MODE_BACKUP 1 1425 #define SSK_MODE_MAX 2 1426 1427 /* implement the mptcp packet scheduler; 1428 * returns the subflow that will transmit the next DSS 1429 * additionally updates the rtx timeout 1430 */ 1431 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1432 { 1433 struct subflow_send_info send_info[SSK_MODE_MAX]; 1434 struct mptcp_subflow_context *subflow; 1435 struct sock *sk = (struct sock *)msk; 1436 u32 pace, burst, wmem; 1437 int i, nr_active = 0; 1438 struct sock *ssk; 1439 u64 linger_time; 1440 long tout = 0; 1441 1442 /* pick the subflow with the lower wmem/wspace ratio */ 1443 for (i = 0; i < SSK_MODE_MAX; ++i) { 1444 send_info[i].ssk = NULL; 1445 send_info[i].linger_time = -1; 1446 } 1447 1448 mptcp_for_each_subflow(msk, subflow) { 1449 bool backup = subflow->backup || subflow->request_bkup; 1450 1451 trace_mptcp_subflow_get_send(subflow); 1452 ssk = mptcp_subflow_tcp_sock(subflow); 1453 if (!mptcp_subflow_active(subflow)) 1454 continue; 1455 1456 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1457 nr_active += !backup; 1458 pace = subflow->avg_pacing_rate; 1459 if (unlikely(!pace)) { 1460 /* init pacing rate from socket */ 1461 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1462 pace = subflow->avg_pacing_rate; 1463 if (!pace) 1464 continue; 1465 } 1466 1467 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1468 if (linger_time < send_info[backup].linger_time) { 1469 send_info[backup].ssk = ssk; 1470 send_info[backup].linger_time = linger_time; 1471 } 1472 } 1473 __mptcp_set_timeout(sk, tout); 1474 1475 /* pick the best backup if no other subflow is active */ 1476 if (!nr_active) 1477 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1478 1479 /* According to the blest algorithm, to avoid HoL blocking for the 1480 * faster flow, we need to: 1481 * - estimate the faster flow linger time 1482 * - use the above to estimate the amount of byte transferred 1483 * by the faster flow 1484 * - check that the amount of queued data is greater than the above, 1485 * otherwise do not use the picked, slower, subflow 1486 * We select the subflow with the shorter estimated time to flush 1487 * the queued mem, which basically ensure the above. We just need 1488 * to check that subflow has a non empty cwin. 1489 */ 1490 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1491 if (!ssk || !sk_stream_memory_free(ssk)) 1492 return NULL; 1493 1494 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt); 1495 wmem = READ_ONCE(ssk->sk_wmem_queued); 1496 if (!burst) 1497 return ssk; 1498 1499 subflow = mptcp_subflow_ctx(ssk); 1500 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1501 READ_ONCE(ssk->sk_pacing_rate) * burst, 1502 burst + wmem); 1503 msk->snd_burst = burst; 1504 return ssk; 1505 } 1506 1507 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1508 { 1509 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1510 release_sock(ssk); 1511 } 1512 1513 static void mptcp_update_post_push(struct mptcp_sock *msk, 1514 struct mptcp_data_frag *dfrag, 1515 u32 sent) 1516 { 1517 u64 snd_nxt_new = dfrag->data_seq; 1518 1519 dfrag->already_sent += sent; 1520 1521 msk->snd_burst -= sent; 1522 1523 snd_nxt_new += dfrag->already_sent; 1524 1525 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1526 * is recovering after a failover. In that event, this re-sends 1527 * old segments. 1528 * 1529 * Thus compute snd_nxt_new candidate based on 1530 * the dfrag->data_seq that was sent and the data 1531 * that has been handed to the subflow for transmission 1532 * and skip update in case it was old dfrag. 1533 */ 1534 if (likely(after64(snd_nxt_new, msk->snd_nxt))) { 1535 msk->bytes_sent += snd_nxt_new - msk->snd_nxt; 1536 WRITE_ONCE(msk->snd_nxt, snd_nxt_new); 1537 } 1538 } 1539 1540 void mptcp_check_and_set_pending(struct sock *sk) 1541 { 1542 if (mptcp_send_head(sk)) { 1543 mptcp_data_lock(sk); 1544 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING); 1545 mptcp_data_unlock(sk); 1546 } 1547 } 1548 1549 static int __subflow_push_pending(struct sock *sk, struct sock *ssk, 1550 struct mptcp_sendmsg_info *info) 1551 { 1552 struct mptcp_sock *msk = mptcp_sk(sk); 1553 struct mptcp_data_frag *dfrag; 1554 int len, copied = 0, err = 0; 1555 1556 while ((dfrag = mptcp_send_head(sk))) { 1557 info->sent = dfrag->already_sent; 1558 info->limit = dfrag->data_len; 1559 len = dfrag->data_len - dfrag->already_sent; 1560 while (len > 0) { 1561 int ret = 0; 1562 1563 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info); 1564 if (ret <= 0) { 1565 err = copied ? : ret; 1566 goto out; 1567 } 1568 1569 info->sent += ret; 1570 copied += ret; 1571 len -= ret; 1572 1573 mptcp_update_post_push(msk, dfrag, ret); 1574 } 1575 msk->first_pending = mptcp_send_next(sk); 1576 1577 if (msk->snd_burst <= 0 || 1578 !sk_stream_memory_free(ssk) || 1579 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) { 1580 err = copied; 1581 goto out; 1582 } 1583 mptcp_set_timeout(sk); 1584 } 1585 err = copied; 1586 1587 out: 1588 if (err > 0) 1589 msk->last_data_sent = tcp_jiffies32; 1590 return err; 1591 } 1592 1593 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1594 { 1595 struct sock *prev_ssk = NULL, *ssk = NULL; 1596 struct mptcp_sock *msk = mptcp_sk(sk); 1597 struct mptcp_sendmsg_info info = { 1598 .flags = flags, 1599 }; 1600 bool do_check_data_fin = false; 1601 int push_count = 1; 1602 1603 while (mptcp_send_head(sk) && (push_count > 0)) { 1604 struct mptcp_subflow_context *subflow; 1605 int ret = 0; 1606 1607 if (mptcp_sched_get_send(msk)) 1608 break; 1609 1610 push_count = 0; 1611 1612 mptcp_for_each_subflow(msk, subflow) { 1613 if (READ_ONCE(subflow->scheduled)) { 1614 mptcp_subflow_set_scheduled(subflow, false); 1615 1616 prev_ssk = ssk; 1617 ssk = mptcp_subflow_tcp_sock(subflow); 1618 if (ssk != prev_ssk) { 1619 /* First check. If the ssk has changed since 1620 * the last round, release prev_ssk 1621 */ 1622 if (prev_ssk) 1623 mptcp_push_release(prev_ssk, &info); 1624 1625 /* Need to lock the new subflow only if different 1626 * from the previous one, otherwise we are still 1627 * helding the relevant lock 1628 */ 1629 lock_sock(ssk); 1630 } 1631 1632 push_count++; 1633 1634 ret = __subflow_push_pending(sk, ssk, &info); 1635 if (ret <= 0) { 1636 if (ret != -EAGAIN || 1637 (1 << ssk->sk_state) & 1638 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE)) 1639 push_count--; 1640 continue; 1641 } 1642 do_check_data_fin = true; 1643 } 1644 } 1645 } 1646 1647 /* at this point we held the socket lock for the last subflow we used */ 1648 if (ssk) 1649 mptcp_push_release(ssk, &info); 1650 1651 /* ensure the rtx timer is running */ 1652 if (!mptcp_rtx_timer_pending(sk)) 1653 mptcp_reset_rtx_timer(sk); 1654 if (do_check_data_fin) 1655 mptcp_check_send_data_fin(sk); 1656 } 1657 1658 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first) 1659 { 1660 struct mptcp_sock *msk = mptcp_sk(sk); 1661 struct mptcp_sendmsg_info info = { 1662 .data_lock_held = true, 1663 }; 1664 bool keep_pushing = true; 1665 struct sock *xmit_ssk; 1666 int copied = 0; 1667 1668 info.flags = 0; 1669 while (mptcp_send_head(sk) && keep_pushing) { 1670 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1671 int ret = 0; 1672 1673 /* check for a different subflow usage only after 1674 * spooling the first chunk of data 1675 */ 1676 if (first) { 1677 mptcp_subflow_set_scheduled(subflow, false); 1678 ret = __subflow_push_pending(sk, ssk, &info); 1679 first = false; 1680 if (ret <= 0) 1681 break; 1682 copied += ret; 1683 continue; 1684 } 1685 1686 if (mptcp_sched_get_send(msk)) 1687 goto out; 1688 1689 if (READ_ONCE(subflow->scheduled)) { 1690 mptcp_subflow_set_scheduled(subflow, false); 1691 ret = __subflow_push_pending(sk, ssk, &info); 1692 if (ret <= 0) 1693 keep_pushing = false; 1694 copied += ret; 1695 } 1696 1697 mptcp_for_each_subflow(msk, subflow) { 1698 if (READ_ONCE(subflow->scheduled)) { 1699 xmit_ssk = mptcp_subflow_tcp_sock(subflow); 1700 if (xmit_ssk != ssk) { 1701 mptcp_subflow_delegate(subflow, 1702 MPTCP_DELEGATE_SEND); 1703 keep_pushing = false; 1704 } 1705 } 1706 } 1707 } 1708 1709 out: 1710 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1711 * not going to flush it via release_sock() 1712 */ 1713 if (copied) { 1714 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1715 info.size_goal); 1716 if (!mptcp_rtx_timer_pending(sk)) 1717 mptcp_reset_rtx_timer(sk); 1718 1719 if (msk->snd_data_fin_enable && 1720 msk->snd_nxt + 1 == msk->write_seq) 1721 mptcp_schedule_work(sk); 1722 } 1723 } 1724 1725 static int mptcp_disconnect(struct sock *sk, int flags); 1726 1727 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1728 size_t len, int *copied_syn) 1729 { 1730 unsigned int saved_flags = msg->msg_flags; 1731 struct mptcp_sock *msk = mptcp_sk(sk); 1732 struct sock *ssk; 1733 int ret; 1734 1735 /* on flags based fastopen the mptcp is supposed to create the 1736 * first subflow right now. Otherwise we are in the defer_connect 1737 * path, and the first subflow must be already present. 1738 * Since the defer_connect flag is cleared after the first succsful 1739 * fastopen attempt, no need to check for additional subflow status. 1740 */ 1741 if (msg->msg_flags & MSG_FASTOPEN) { 1742 ssk = __mptcp_nmpc_sk(msk); 1743 if (IS_ERR(ssk)) 1744 return PTR_ERR(ssk); 1745 } 1746 if (!msk->first) 1747 return -EINVAL; 1748 1749 ssk = msk->first; 1750 1751 lock_sock(ssk); 1752 msg->msg_flags |= MSG_DONTWAIT; 1753 msk->fastopening = 1; 1754 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL); 1755 msk->fastopening = 0; 1756 msg->msg_flags = saved_flags; 1757 release_sock(ssk); 1758 1759 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */ 1760 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) { 1761 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1762 msg->msg_namelen, msg->msg_flags, 1); 1763 1764 /* Keep the same behaviour of plain TCP: zero the copied bytes in 1765 * case of any error, except timeout or signal 1766 */ 1767 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR) 1768 *copied_syn = 0; 1769 } else if (ret && ret != -EINPROGRESS) { 1770 /* The disconnect() op called by tcp_sendmsg_fastopen()/ 1771 * __inet_stream_connect() can fail, due to looking check, 1772 * see mptcp_disconnect(). 1773 * Attempt it again outside the problematic scope. 1774 */ 1775 if (!mptcp_disconnect(sk, 0)) { 1776 sk->sk_disconnects++; 1777 sk->sk_socket->state = SS_UNCONNECTED; 1778 } 1779 } 1780 inet_clear_bit(DEFER_CONNECT, sk); 1781 1782 return ret; 1783 } 1784 1785 static int do_copy_data_nocache(struct sock *sk, int copy, 1786 struct iov_iter *from, char *to) 1787 { 1788 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) { 1789 if (!copy_from_iter_full_nocache(to, copy, from)) 1790 return -EFAULT; 1791 } else if (!copy_from_iter_full(to, copy, from)) { 1792 return -EFAULT; 1793 } 1794 return 0; 1795 } 1796 1797 /* open-code sk_stream_memory_free() plus sent limit computation to 1798 * avoid indirect calls in fast-path. 1799 * Called under the msk socket lock, so we can avoid a bunch of ONCE 1800 * annotations. 1801 */ 1802 static u32 mptcp_send_limit(const struct sock *sk) 1803 { 1804 const struct mptcp_sock *msk = mptcp_sk(sk); 1805 u32 limit, not_sent; 1806 1807 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf)) 1808 return 0; 1809 1810 limit = mptcp_notsent_lowat(sk); 1811 if (limit == UINT_MAX) 1812 return UINT_MAX; 1813 1814 not_sent = msk->write_seq - msk->snd_nxt; 1815 if (not_sent >= limit) 1816 return 0; 1817 1818 return limit - not_sent; 1819 } 1820 1821 static void mptcp_rps_record_subflows(const struct mptcp_sock *msk) 1822 { 1823 struct mptcp_subflow_context *subflow; 1824 1825 if (!rfs_is_needed()) 1826 return; 1827 1828 mptcp_for_each_subflow(msk, subflow) { 1829 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1830 1831 sock_rps_record_flow(ssk); 1832 } 1833 } 1834 1835 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1836 { 1837 struct mptcp_sock *msk = mptcp_sk(sk); 1838 struct page_frag *pfrag; 1839 size_t copied = 0; 1840 int ret = 0; 1841 long timeo; 1842 1843 /* silently ignore everything else */ 1844 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN; 1845 1846 lock_sock(sk); 1847 1848 mptcp_rps_record_subflows(msk); 1849 1850 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) || 1851 msg->msg_flags & MSG_FASTOPEN)) { 1852 int copied_syn = 0; 1853 1854 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn); 1855 copied += copied_syn; 1856 if (ret == -EINPROGRESS && copied_syn > 0) 1857 goto out; 1858 else if (ret) 1859 goto do_error; 1860 } 1861 1862 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1863 1864 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1865 ret = sk_stream_wait_connect(sk, &timeo); 1866 if (ret) 1867 goto do_error; 1868 } 1869 1870 ret = -EPIPE; 1871 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))) 1872 goto do_error; 1873 1874 pfrag = sk_page_frag(sk); 1875 1876 while (msg_data_left(msg)) { 1877 int total_ts, frag_truesize = 0; 1878 struct mptcp_data_frag *dfrag; 1879 bool dfrag_collapsed; 1880 size_t psize, offset; 1881 u32 copy_limit; 1882 1883 /* ensure fitting the notsent_lowat() constraint */ 1884 copy_limit = mptcp_send_limit(sk); 1885 if (!copy_limit) 1886 goto wait_for_memory; 1887 1888 /* reuse tail pfrag, if possible, or carve a new one from the 1889 * page allocator 1890 */ 1891 dfrag = mptcp_pending_tail(sk); 1892 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1893 if (!dfrag_collapsed) { 1894 if (!mptcp_page_frag_refill(sk, pfrag)) 1895 goto wait_for_memory; 1896 1897 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1898 frag_truesize = dfrag->overhead; 1899 } 1900 1901 /* we do not bound vs wspace, to allow a single packet. 1902 * memory accounting will prevent execessive memory usage 1903 * anyway 1904 */ 1905 offset = dfrag->offset + dfrag->data_len; 1906 psize = pfrag->size - offset; 1907 psize = min_t(size_t, psize, msg_data_left(msg)); 1908 psize = min_t(size_t, psize, copy_limit); 1909 total_ts = psize + frag_truesize; 1910 1911 if (!sk_wmem_schedule(sk, total_ts)) 1912 goto wait_for_memory; 1913 1914 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter, 1915 page_address(dfrag->page) + offset); 1916 if (ret) 1917 goto do_error; 1918 1919 /* data successfully copied into the write queue */ 1920 sk_forward_alloc_add(sk, -total_ts); 1921 copied += psize; 1922 dfrag->data_len += psize; 1923 frag_truesize += psize; 1924 pfrag->offset += frag_truesize; 1925 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1926 1927 /* charge data on mptcp pending queue to the msk socket 1928 * Note: we charge such data both to sk and ssk 1929 */ 1930 sk_wmem_queued_add(sk, frag_truesize); 1931 if (!dfrag_collapsed) { 1932 get_page(dfrag->page); 1933 list_add_tail(&dfrag->list, &msk->rtx_queue); 1934 if (!msk->first_pending) 1935 msk->first_pending = dfrag; 1936 } 1937 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk, 1938 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1939 !dfrag_collapsed); 1940 1941 continue; 1942 1943 wait_for_memory: 1944 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1945 __mptcp_push_pending(sk, msg->msg_flags); 1946 ret = sk_stream_wait_memory(sk, &timeo); 1947 if (ret) 1948 goto do_error; 1949 } 1950 1951 if (copied) 1952 __mptcp_push_pending(sk, msg->msg_flags); 1953 1954 out: 1955 release_sock(sk); 1956 return copied; 1957 1958 do_error: 1959 if (copied) 1960 goto out; 1961 1962 copied = sk_stream_error(sk, msg->msg_flags, ret); 1963 goto out; 1964 } 1965 1966 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied); 1967 1968 static int __mptcp_recvmsg_mskq(struct sock *sk, struct msghdr *msg, 1969 size_t len, int flags, int copied_total, 1970 struct scm_timestamping_internal *tss, 1971 int *cmsg_flags) 1972 { 1973 struct mptcp_sock *msk = mptcp_sk(sk); 1974 struct sk_buff *skb, *tmp; 1975 int total_data_len = 0; 1976 int copied = 0; 1977 1978 skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) { 1979 u32 delta, offset = MPTCP_SKB_CB(skb)->offset; 1980 u32 data_len = skb->len - offset; 1981 u32 count; 1982 int err; 1983 1984 if (flags & MSG_PEEK) { 1985 /* skip already peeked skbs */ 1986 if (total_data_len + data_len <= copied_total) { 1987 total_data_len += data_len; 1988 continue; 1989 } 1990 1991 /* skip the already peeked data in the current skb */ 1992 delta = copied_total - total_data_len; 1993 offset += delta; 1994 data_len -= delta; 1995 } 1996 1997 count = min_t(size_t, len - copied, data_len); 1998 if (!(flags & MSG_TRUNC)) { 1999 err = skb_copy_datagram_msg(skb, offset, msg, count); 2000 if (unlikely(err < 0)) { 2001 if (!copied) 2002 return err; 2003 break; 2004 } 2005 } 2006 2007 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 2008 tcp_update_recv_tstamps(skb, tss); 2009 *cmsg_flags |= MPTCP_CMSG_TS; 2010 } 2011 2012 copied += count; 2013 2014 if (!(flags & MSG_PEEK)) { 2015 msk->bytes_consumed += count; 2016 if (count < data_len) { 2017 MPTCP_SKB_CB(skb)->offset += count; 2018 MPTCP_SKB_CB(skb)->map_seq += count; 2019 break; 2020 } 2021 2022 /* avoid the indirect call, we know the destructor is sock_rfree */ 2023 skb->destructor = NULL; 2024 skb->sk = NULL; 2025 atomic_sub(skb->truesize, &sk->sk_rmem_alloc); 2026 sk_mem_uncharge(sk, skb->truesize); 2027 __skb_unlink(skb, &sk->sk_receive_queue); 2028 skb_attempt_defer_free(skb); 2029 } 2030 2031 if (copied >= len) 2032 break; 2033 } 2034 2035 mptcp_rcv_space_adjust(msk, copied); 2036 return copied; 2037 } 2038 2039 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 2040 * 2041 * Only difference: Use highest rtt estimate of the subflows in use. 2042 */ 2043 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 2044 { 2045 struct mptcp_subflow_context *subflow; 2046 struct sock *sk = (struct sock *)msk; 2047 u8 scaling_ratio = U8_MAX; 2048 u32 time, advmss = 1; 2049 u64 rtt_us, mstamp; 2050 2051 msk_owned_by_me(msk); 2052 2053 if (copied <= 0) 2054 return; 2055 2056 if (!msk->rcvspace_init) 2057 mptcp_rcv_space_init(msk, msk->first); 2058 2059 msk->rcvq_space.copied += copied; 2060 2061 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 2062 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 2063 2064 rtt_us = msk->rcvq_space.rtt_us; 2065 if (rtt_us && time < (rtt_us >> 3)) 2066 return; 2067 2068 rtt_us = 0; 2069 mptcp_for_each_subflow(msk, subflow) { 2070 const struct tcp_sock *tp; 2071 u64 sf_rtt_us; 2072 u32 sf_advmss; 2073 2074 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 2075 2076 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 2077 sf_advmss = READ_ONCE(tp->advmss); 2078 2079 rtt_us = max(sf_rtt_us, rtt_us); 2080 advmss = max(sf_advmss, advmss); 2081 scaling_ratio = min(tp->scaling_ratio, scaling_ratio); 2082 } 2083 2084 msk->rcvq_space.rtt_us = rtt_us; 2085 msk->scaling_ratio = scaling_ratio; 2086 if (time < (rtt_us >> 3) || rtt_us == 0) 2087 return; 2088 2089 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 2090 goto new_measure; 2091 2092 if (mptcp_rcvbuf_grow(sk, msk->rcvq_space.copied)) { 2093 /* Make subflows follow along. If we do not do this, we 2094 * get drops at subflow level if skbs can't be moved to 2095 * the mptcp rx queue fast enough (announced rcv_win can 2096 * exceed ssk->sk_rcvbuf). 2097 */ 2098 mptcp_for_each_subflow(msk, subflow) { 2099 struct sock *ssk; 2100 bool slow; 2101 2102 ssk = mptcp_subflow_tcp_sock(subflow); 2103 slow = lock_sock_fast(ssk); 2104 /* subflows can be added before tcp_init_transfer() */ 2105 if (tcp_sk(ssk)->rcvq_space.space) 2106 tcp_rcvbuf_grow(ssk, msk->rcvq_space.copied); 2107 unlock_sock_fast(ssk, slow); 2108 } 2109 } 2110 2111 new_measure: 2112 msk->rcvq_space.copied = 0; 2113 msk->rcvq_space.time = mstamp; 2114 } 2115 2116 static struct mptcp_subflow_context * 2117 __mptcp_first_ready_from(struct mptcp_sock *msk, 2118 struct mptcp_subflow_context *subflow) 2119 { 2120 struct mptcp_subflow_context *start_subflow = subflow; 2121 2122 while (!READ_ONCE(subflow->data_avail)) { 2123 subflow = mptcp_next_subflow(msk, subflow); 2124 if (subflow == start_subflow) 2125 return NULL; 2126 } 2127 return subflow; 2128 } 2129 2130 static bool __mptcp_move_skbs(struct sock *sk) 2131 { 2132 struct mptcp_subflow_context *subflow; 2133 struct mptcp_sock *msk = mptcp_sk(sk); 2134 bool ret = false; 2135 2136 if (list_empty(&msk->conn_list)) 2137 return false; 2138 2139 subflow = list_first_entry(&msk->conn_list, 2140 struct mptcp_subflow_context, node); 2141 for (;;) { 2142 struct sock *ssk; 2143 bool slowpath; 2144 2145 /* 2146 * As an optimization avoid traversing the subflows list 2147 * and ev. acquiring the subflow socket lock before baling out 2148 */ 2149 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 2150 break; 2151 2152 subflow = __mptcp_first_ready_from(msk, subflow); 2153 if (!subflow) 2154 break; 2155 2156 ssk = mptcp_subflow_tcp_sock(subflow); 2157 slowpath = lock_sock_fast(ssk); 2158 ret = __mptcp_move_skbs_from_subflow(msk, ssk) || ret; 2159 if (unlikely(ssk->sk_err)) 2160 __mptcp_error_report(sk); 2161 unlock_sock_fast(ssk, slowpath); 2162 2163 subflow = mptcp_next_subflow(msk, subflow); 2164 } 2165 2166 __mptcp_ofo_queue(msk); 2167 if (ret) 2168 mptcp_check_data_fin((struct sock *)msk); 2169 return ret; 2170 } 2171 2172 static unsigned int mptcp_inq_hint(const struct sock *sk) 2173 { 2174 const struct mptcp_sock *msk = mptcp_sk(sk); 2175 const struct sk_buff *skb; 2176 2177 skb = skb_peek(&sk->sk_receive_queue); 2178 if (skb) { 2179 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq; 2180 2181 if (hint_val >= INT_MAX) 2182 return INT_MAX; 2183 2184 return (unsigned int)hint_val; 2185 } 2186 2187 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2188 return 1; 2189 2190 return 0; 2191 } 2192 2193 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2194 int flags, int *addr_len) 2195 { 2196 struct mptcp_sock *msk = mptcp_sk(sk); 2197 struct scm_timestamping_internal tss; 2198 int copied = 0, cmsg_flags = 0; 2199 int target; 2200 long timeo; 2201 2202 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2203 if (unlikely(flags & MSG_ERRQUEUE)) 2204 return inet_recv_error(sk, msg, len, addr_len); 2205 2206 lock_sock(sk); 2207 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2208 copied = -ENOTCONN; 2209 goto out_err; 2210 } 2211 2212 mptcp_rps_record_subflows(msk); 2213 2214 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2215 2216 len = min_t(size_t, len, INT_MAX); 2217 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2218 2219 if (unlikely(msk->recvmsg_inq)) 2220 cmsg_flags = MPTCP_CMSG_INQ; 2221 2222 while (copied < len) { 2223 int err, bytes_read; 2224 2225 bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags, 2226 copied, &tss, &cmsg_flags); 2227 if (unlikely(bytes_read < 0)) { 2228 if (!copied) 2229 copied = bytes_read; 2230 goto out_err; 2231 } 2232 2233 copied += bytes_read; 2234 2235 if (skb_queue_empty(&sk->sk_receive_queue) && __mptcp_move_skbs(sk)) 2236 continue; 2237 2238 /* only the MPTCP socket status is relevant here. The exit 2239 * conditions mirror closely tcp_recvmsg() 2240 */ 2241 if (copied >= target) 2242 break; 2243 2244 if (copied) { 2245 if (sk->sk_err || 2246 sk->sk_state == TCP_CLOSE || 2247 (sk->sk_shutdown & RCV_SHUTDOWN) || 2248 !timeo || 2249 signal_pending(current)) 2250 break; 2251 } else { 2252 if (sk->sk_err) { 2253 copied = sock_error(sk); 2254 break; 2255 } 2256 2257 if (sk->sk_shutdown & RCV_SHUTDOWN) 2258 break; 2259 2260 if (sk->sk_state == TCP_CLOSE) { 2261 copied = -ENOTCONN; 2262 break; 2263 } 2264 2265 if (!timeo) { 2266 copied = -EAGAIN; 2267 break; 2268 } 2269 2270 if (signal_pending(current)) { 2271 copied = sock_intr_errno(timeo); 2272 break; 2273 } 2274 } 2275 2276 pr_debug("block timeout %ld\n", timeo); 2277 mptcp_cleanup_rbuf(msk, copied); 2278 err = sk_wait_data(sk, &timeo, NULL); 2279 if (err < 0) { 2280 err = copied ? : err; 2281 goto out_err; 2282 } 2283 } 2284 2285 mptcp_cleanup_rbuf(msk, copied); 2286 2287 out_err: 2288 if (cmsg_flags && copied >= 0) { 2289 if (cmsg_flags & MPTCP_CMSG_TS) 2290 tcp_recv_timestamp(msg, sk, &tss); 2291 2292 if (cmsg_flags & MPTCP_CMSG_INQ) { 2293 unsigned int inq = mptcp_inq_hint(sk); 2294 2295 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2296 } 2297 } 2298 2299 pr_debug("msk=%p rx queue empty=%d copied=%d\n", 2300 msk, skb_queue_empty(&sk->sk_receive_queue), copied); 2301 2302 release_sock(sk); 2303 return copied; 2304 } 2305 2306 static void mptcp_retransmit_timer(struct timer_list *t) 2307 { 2308 struct inet_connection_sock *icsk = timer_container_of(icsk, t, 2309 icsk_retransmit_timer); 2310 struct sock *sk = &icsk->icsk_inet.sk; 2311 struct mptcp_sock *msk = mptcp_sk(sk); 2312 2313 bh_lock_sock(sk); 2314 if (!sock_owned_by_user(sk)) { 2315 /* we need a process context to retransmit */ 2316 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2317 mptcp_schedule_work(sk); 2318 } else { 2319 /* delegate our work to tcp_release_cb() */ 2320 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2321 } 2322 bh_unlock_sock(sk); 2323 sock_put(sk); 2324 } 2325 2326 static void mptcp_tout_timer(struct timer_list *t) 2327 { 2328 struct sock *sk = timer_container_of(sk, t, sk_timer); 2329 2330 mptcp_schedule_work(sk); 2331 sock_put(sk); 2332 } 2333 2334 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2335 * level. 2336 * 2337 * A backup subflow is returned only if that is the only kind available. 2338 */ 2339 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2340 { 2341 struct sock *backup = NULL, *pick = NULL; 2342 struct mptcp_subflow_context *subflow; 2343 int min_stale_count = INT_MAX; 2344 2345 mptcp_for_each_subflow(msk, subflow) { 2346 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2347 2348 if (!__mptcp_subflow_active(subflow)) 2349 continue; 2350 2351 /* still data outstanding at TCP level? skip this */ 2352 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2353 mptcp_pm_subflow_chk_stale(msk, ssk); 2354 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2355 continue; 2356 } 2357 2358 if (subflow->backup || subflow->request_bkup) { 2359 if (!backup) 2360 backup = ssk; 2361 continue; 2362 } 2363 2364 if (!pick) 2365 pick = ssk; 2366 } 2367 2368 if (pick) 2369 return pick; 2370 2371 /* use backup only if there are no progresses anywhere */ 2372 return min_stale_count > 1 ? backup : NULL; 2373 } 2374 2375 bool __mptcp_retransmit_pending_data(struct sock *sk) 2376 { 2377 struct mptcp_data_frag *cur, *rtx_head; 2378 struct mptcp_sock *msk = mptcp_sk(sk); 2379 2380 if (__mptcp_check_fallback(msk)) 2381 return false; 2382 2383 /* the closing socket has some data untransmitted and/or unacked: 2384 * some data in the mptcp rtx queue has not really xmitted yet. 2385 * keep it simple and re-inject the whole mptcp level rtx queue 2386 */ 2387 mptcp_data_lock(sk); 2388 __mptcp_clean_una_wakeup(sk); 2389 rtx_head = mptcp_rtx_head(sk); 2390 if (!rtx_head) { 2391 mptcp_data_unlock(sk); 2392 return false; 2393 } 2394 2395 msk->recovery_snd_nxt = msk->snd_nxt; 2396 msk->recovery = true; 2397 mptcp_data_unlock(sk); 2398 2399 msk->first_pending = rtx_head; 2400 msk->snd_burst = 0; 2401 2402 /* be sure to clear the "sent status" on all re-injected fragments */ 2403 list_for_each_entry(cur, &msk->rtx_queue, list) { 2404 if (!cur->already_sent) 2405 break; 2406 cur->already_sent = 0; 2407 } 2408 2409 return true; 2410 } 2411 2412 /* flags for __mptcp_close_ssk() */ 2413 #define MPTCP_CF_PUSH BIT(1) 2414 2415 /* be sure to send a reset only if the caller asked for it, also 2416 * clean completely the subflow status when the subflow reaches 2417 * TCP_CLOSE state 2418 */ 2419 static void __mptcp_subflow_disconnect(struct sock *ssk, 2420 struct mptcp_subflow_context *subflow, 2421 unsigned int flags) 2422 { 2423 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 2424 subflow->send_fastclose) { 2425 /* The MPTCP code never wait on the subflow sockets, TCP-level 2426 * disconnect should never fail 2427 */ 2428 WARN_ON_ONCE(tcp_disconnect(ssk, 0)); 2429 mptcp_subflow_ctx_reset(subflow); 2430 } else { 2431 tcp_shutdown(ssk, SEND_SHUTDOWN); 2432 } 2433 } 2434 2435 /* subflow sockets can be either outgoing (connect) or incoming 2436 * (accept). 2437 * 2438 * Outgoing subflows use in-kernel sockets. 2439 * Incoming subflows do not have their own 'struct socket' allocated, 2440 * so we need to use tcp_close() after detaching them from the mptcp 2441 * parent socket. 2442 */ 2443 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2444 struct mptcp_subflow_context *subflow, 2445 unsigned int flags) 2446 { 2447 struct mptcp_sock *msk = mptcp_sk(sk); 2448 bool dispose_it, need_push = false; 2449 2450 /* If the first subflow moved to a close state before accept, e.g. due 2451 * to an incoming reset or listener shutdown, the subflow socket is 2452 * already deleted by inet_child_forget() and the mptcp socket can't 2453 * survive too. 2454 */ 2455 if (msk->in_accept_queue && msk->first == ssk && 2456 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) { 2457 /* ensure later check in mptcp_worker() will dispose the msk */ 2458 sock_set_flag(sk, SOCK_DEAD); 2459 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1)); 2460 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2461 mptcp_subflow_drop_ctx(ssk); 2462 goto out_release; 2463 } 2464 2465 dispose_it = msk->free_first || ssk != msk->first; 2466 if (dispose_it) 2467 list_del(&subflow->node); 2468 2469 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2470 2471 if (subflow->send_fastclose && ssk->sk_state != TCP_CLOSE) 2472 tcp_set_state(ssk, TCP_CLOSE); 2473 2474 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2475 if (!dispose_it) { 2476 __mptcp_subflow_disconnect(ssk, subflow, flags); 2477 release_sock(ssk); 2478 2479 goto out; 2480 } 2481 2482 subflow->disposable = 1; 2483 2484 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2485 * the ssk has been already destroyed, we just need to release the 2486 * reference owned by msk; 2487 */ 2488 if (!inet_csk(ssk)->icsk_ulp_ops) { 2489 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD)); 2490 kfree_rcu(subflow, rcu); 2491 } else { 2492 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2493 __tcp_close(ssk, 0); 2494 2495 /* close acquired an extra ref */ 2496 __sock_put(ssk); 2497 } 2498 2499 out_release: 2500 __mptcp_subflow_error_report(sk, ssk); 2501 release_sock(ssk); 2502 2503 sock_put(ssk); 2504 2505 if (ssk == msk->first) 2506 WRITE_ONCE(msk->first, NULL); 2507 2508 out: 2509 __mptcp_sync_sndbuf(sk); 2510 if (need_push) 2511 __mptcp_push_pending(sk, 0); 2512 2513 /* Catch every 'all subflows closed' scenario, including peers silently 2514 * closing them, e.g. due to timeout. 2515 * For established sockets, allow an additional timeout before closing, 2516 * as the protocol can still create more subflows. 2517 */ 2518 if (list_is_singular(&msk->conn_list) && msk->first && 2519 inet_sk_state_load(msk->first) == TCP_CLOSE) { 2520 if (sk->sk_state != TCP_ESTABLISHED || 2521 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) { 2522 mptcp_set_state(sk, TCP_CLOSE); 2523 mptcp_close_wake_up(sk); 2524 } else { 2525 mptcp_start_tout_timer(sk); 2526 } 2527 } 2528 } 2529 2530 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2531 struct mptcp_subflow_context *subflow) 2532 { 2533 /* The first subflow can already be closed and still in the list */ 2534 if (subflow->close_event_done) 2535 return; 2536 2537 subflow->close_event_done = true; 2538 2539 if (sk->sk_state == TCP_ESTABLISHED) 2540 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2541 2542 /* subflow aborted before reaching the fully_established status 2543 * attempt the creation of the next subflow 2544 */ 2545 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow); 2546 2547 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2548 } 2549 2550 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2551 { 2552 return 0; 2553 } 2554 2555 static void __mptcp_close_subflow(struct sock *sk) 2556 { 2557 struct mptcp_subflow_context *subflow, *tmp; 2558 struct mptcp_sock *msk = mptcp_sk(sk); 2559 2560 might_sleep(); 2561 2562 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2563 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2564 int ssk_state = inet_sk_state_load(ssk); 2565 2566 if (ssk_state != TCP_CLOSE && 2567 (ssk_state != TCP_CLOSE_WAIT || 2568 inet_sk_state_load(sk) != TCP_ESTABLISHED || 2569 __mptcp_check_fallback(msk))) 2570 continue; 2571 2572 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2573 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2574 continue; 2575 2576 mptcp_close_ssk(sk, ssk, subflow); 2577 } 2578 2579 } 2580 2581 static bool mptcp_close_tout_expired(const struct sock *sk) 2582 { 2583 if (!inet_csk(sk)->icsk_mtup.probe_timestamp || 2584 sk->sk_state == TCP_CLOSE) 2585 return false; 2586 2587 return time_after32(tcp_jiffies32, 2588 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk)); 2589 } 2590 2591 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2592 { 2593 struct mptcp_subflow_context *subflow, *tmp; 2594 struct sock *sk = (struct sock *)msk; 2595 2596 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2597 return; 2598 2599 mptcp_token_destroy(msk); 2600 2601 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2602 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2603 bool slow; 2604 2605 slow = lock_sock_fast(tcp_sk); 2606 if (tcp_sk->sk_state != TCP_CLOSE) { 2607 mptcp_send_active_reset_reason(tcp_sk); 2608 tcp_set_state(tcp_sk, TCP_CLOSE); 2609 } 2610 unlock_sock_fast(tcp_sk, slow); 2611 } 2612 2613 /* Mirror the tcp_reset() error propagation */ 2614 switch (sk->sk_state) { 2615 case TCP_SYN_SENT: 2616 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 2617 break; 2618 case TCP_CLOSE_WAIT: 2619 WRITE_ONCE(sk->sk_err, EPIPE); 2620 break; 2621 case TCP_CLOSE: 2622 return; 2623 default: 2624 WRITE_ONCE(sk->sk_err, ECONNRESET); 2625 } 2626 2627 mptcp_set_state(sk, TCP_CLOSE); 2628 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2629 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2630 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2631 2632 /* the calling mptcp_worker will properly destroy the socket */ 2633 if (sock_flag(sk, SOCK_DEAD)) 2634 return; 2635 2636 sk->sk_state_change(sk); 2637 sk_error_report(sk); 2638 } 2639 2640 static void __mptcp_retrans(struct sock *sk) 2641 { 2642 struct mptcp_sendmsg_info info = { .data_lock_held = true, }; 2643 struct mptcp_sock *msk = mptcp_sk(sk); 2644 struct mptcp_subflow_context *subflow; 2645 struct mptcp_data_frag *dfrag; 2646 struct sock *ssk; 2647 int ret, err; 2648 u16 len = 0; 2649 2650 mptcp_clean_una_wakeup(sk); 2651 2652 /* first check ssk: need to kick "stale" logic */ 2653 err = mptcp_sched_get_retrans(msk); 2654 dfrag = mptcp_rtx_head(sk); 2655 if (!dfrag) { 2656 if (mptcp_data_fin_enabled(msk)) { 2657 struct inet_connection_sock *icsk = inet_csk(sk); 2658 2659 WRITE_ONCE(icsk->icsk_retransmits, 2660 icsk->icsk_retransmits + 1); 2661 mptcp_set_datafin_timeout(sk); 2662 mptcp_send_ack(msk); 2663 2664 goto reset_timer; 2665 } 2666 2667 if (!mptcp_send_head(sk)) 2668 return; 2669 2670 goto reset_timer; 2671 } 2672 2673 if (err) 2674 goto reset_timer; 2675 2676 mptcp_for_each_subflow(msk, subflow) { 2677 if (READ_ONCE(subflow->scheduled)) { 2678 u16 copied = 0; 2679 2680 mptcp_subflow_set_scheduled(subflow, false); 2681 2682 ssk = mptcp_subflow_tcp_sock(subflow); 2683 2684 lock_sock(ssk); 2685 2686 /* limit retransmission to the bytes already sent on some subflows */ 2687 info.sent = 0; 2688 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : 2689 dfrag->already_sent; 2690 2691 /* 2692 * make the whole retrans decision, xmit, disallow 2693 * fallback atomic 2694 */ 2695 spin_lock_bh(&msk->fallback_lock); 2696 if (__mptcp_check_fallback(msk)) { 2697 spin_unlock_bh(&msk->fallback_lock); 2698 release_sock(ssk); 2699 return; 2700 } 2701 2702 while (info.sent < info.limit) { 2703 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2704 if (ret <= 0) 2705 break; 2706 2707 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2708 copied += ret; 2709 info.sent += ret; 2710 } 2711 if (copied) { 2712 len = max(copied, len); 2713 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2714 info.size_goal); 2715 msk->allow_infinite_fallback = false; 2716 } 2717 spin_unlock_bh(&msk->fallback_lock); 2718 2719 release_sock(ssk); 2720 } 2721 } 2722 2723 msk->bytes_retrans += len; 2724 dfrag->already_sent = max(dfrag->already_sent, len); 2725 2726 reset_timer: 2727 mptcp_check_and_set_pending(sk); 2728 2729 if (!mptcp_rtx_timer_pending(sk)) 2730 mptcp_reset_rtx_timer(sk); 2731 } 2732 2733 /* schedule the timeout timer for the relevant event: either close timeout 2734 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one 2735 */ 2736 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout) 2737 { 2738 struct sock *sk = (struct sock *)msk; 2739 unsigned long timeout, close_timeout; 2740 2741 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp) 2742 return; 2743 2744 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp - 2745 tcp_jiffies32 + jiffies + mptcp_close_timeout(sk); 2746 2747 /* the close timeout takes precedence on the fail one, and here at least one of 2748 * them is active 2749 */ 2750 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout; 2751 2752 sk_reset_timer(sk, &sk->sk_timer, timeout); 2753 } 2754 2755 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2756 { 2757 struct sock *ssk = msk->first; 2758 bool slow; 2759 2760 if (!ssk) 2761 return; 2762 2763 pr_debug("MP_FAIL doesn't respond, reset the subflow\n"); 2764 2765 slow = lock_sock_fast(ssk); 2766 mptcp_subflow_reset(ssk); 2767 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0); 2768 unlock_sock_fast(ssk, slow); 2769 } 2770 2771 static void mptcp_do_fastclose(struct sock *sk) 2772 { 2773 struct mptcp_subflow_context *subflow, *tmp; 2774 struct mptcp_sock *msk = mptcp_sk(sk); 2775 2776 mptcp_set_state(sk, TCP_CLOSE); 2777 2778 /* Explicitly send the fastclose reset as need */ 2779 if (__mptcp_check_fallback(msk)) 2780 return; 2781 2782 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2783 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2784 2785 lock_sock(ssk); 2786 2787 /* Some subflow socket states don't allow/need a reset.*/ 2788 if ((1 << ssk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) 2789 goto unlock; 2790 2791 subflow->send_fastclose = 1; 2792 tcp_send_active_reset(ssk, ssk->sk_allocation, 2793 SK_RST_REASON_TCP_ABORT_ON_CLOSE); 2794 unlock: 2795 release_sock(ssk); 2796 } 2797 } 2798 2799 static void mptcp_worker(struct work_struct *work) 2800 { 2801 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2802 struct sock *sk = (struct sock *)msk; 2803 unsigned long fail_tout; 2804 int state; 2805 2806 lock_sock(sk); 2807 state = sk->sk_state; 2808 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN))) 2809 goto unlock; 2810 2811 mptcp_check_fastclose(msk); 2812 2813 mptcp_pm_worker(msk); 2814 2815 mptcp_check_send_data_fin(sk); 2816 mptcp_check_data_fin_ack(sk); 2817 mptcp_check_data_fin(sk); 2818 2819 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2820 __mptcp_close_subflow(sk); 2821 2822 if (mptcp_close_tout_expired(sk)) { 2823 struct mptcp_subflow_context *subflow, *tmp; 2824 2825 mptcp_do_fastclose(sk); 2826 mptcp_for_each_subflow_safe(msk, subflow, tmp) 2827 __mptcp_close_ssk(sk, subflow->tcp_sock, subflow, 0); 2828 mptcp_close_wake_up(sk); 2829 } 2830 2831 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) { 2832 __mptcp_destroy_sock(sk); 2833 goto unlock; 2834 } 2835 2836 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2837 __mptcp_retrans(sk); 2838 2839 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0; 2840 if (fail_tout && time_after(jiffies, fail_tout)) 2841 mptcp_mp_fail_no_response(msk); 2842 2843 unlock: 2844 release_sock(sk); 2845 sock_put(sk); 2846 } 2847 2848 static void __mptcp_init_sock(struct sock *sk) 2849 { 2850 struct mptcp_sock *msk = mptcp_sk(sk); 2851 2852 INIT_LIST_HEAD(&msk->conn_list); 2853 INIT_LIST_HEAD(&msk->join_list); 2854 INIT_LIST_HEAD(&msk->rtx_queue); 2855 INIT_WORK(&msk->work, mptcp_worker); 2856 msk->out_of_order_queue = RB_ROOT; 2857 msk->first_pending = NULL; 2858 msk->timer_ival = TCP_RTO_MIN; 2859 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO; 2860 2861 WRITE_ONCE(msk->first, NULL); 2862 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2863 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2864 msk->allow_infinite_fallback = true; 2865 msk->allow_subflows = true; 2866 msk->recovery = false; 2867 msk->subflow_id = 1; 2868 msk->last_data_sent = tcp_jiffies32; 2869 msk->last_data_recv = tcp_jiffies32; 2870 msk->last_ack_recv = tcp_jiffies32; 2871 2872 mptcp_pm_data_init(msk); 2873 spin_lock_init(&msk->fallback_lock); 2874 2875 /* re-use the csk retrans timer for MPTCP-level retrans */ 2876 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2877 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0); 2878 } 2879 2880 static void mptcp_ca_reset(struct sock *sk) 2881 { 2882 struct inet_connection_sock *icsk = inet_csk(sk); 2883 2884 tcp_assign_congestion_control(sk); 2885 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name, 2886 sizeof(mptcp_sk(sk)->ca_name)); 2887 2888 /* no need to keep a reference to the ops, the name will suffice */ 2889 tcp_cleanup_congestion_control(sk); 2890 icsk->icsk_ca_ops = NULL; 2891 } 2892 2893 static int mptcp_init_sock(struct sock *sk) 2894 { 2895 struct net *net = sock_net(sk); 2896 int ret; 2897 2898 __mptcp_init_sock(sk); 2899 2900 if (!mptcp_is_enabled(net)) 2901 return -ENOPROTOOPT; 2902 2903 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2904 return -ENOMEM; 2905 2906 rcu_read_lock(); 2907 ret = mptcp_init_sched(mptcp_sk(sk), 2908 mptcp_sched_find(mptcp_get_scheduler(net))); 2909 rcu_read_unlock(); 2910 if (ret) 2911 return ret; 2912 2913 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); 2914 2915 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2916 * propagate the correct value 2917 */ 2918 mptcp_ca_reset(sk); 2919 2920 sk_sockets_allocated_inc(sk); 2921 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]); 2922 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]); 2923 2924 return 0; 2925 } 2926 2927 static void __mptcp_clear_xmit(struct sock *sk) 2928 { 2929 struct mptcp_sock *msk = mptcp_sk(sk); 2930 struct mptcp_data_frag *dtmp, *dfrag; 2931 2932 msk->first_pending = NULL; 2933 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2934 dfrag_clear(sk, dfrag); 2935 } 2936 2937 void mptcp_cancel_work(struct sock *sk) 2938 { 2939 struct mptcp_sock *msk = mptcp_sk(sk); 2940 2941 if (cancel_work_sync(&msk->work)) 2942 __sock_put(sk); 2943 } 2944 2945 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2946 { 2947 lock_sock(ssk); 2948 2949 switch (ssk->sk_state) { 2950 case TCP_LISTEN: 2951 if (!(how & RCV_SHUTDOWN)) 2952 break; 2953 fallthrough; 2954 case TCP_SYN_SENT: 2955 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK)); 2956 break; 2957 default: 2958 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2959 pr_debug("Fallback\n"); 2960 ssk->sk_shutdown |= how; 2961 tcp_shutdown(ssk, how); 2962 2963 /* simulate the data_fin ack reception to let the state 2964 * machine move forward 2965 */ 2966 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt); 2967 mptcp_schedule_work(sk); 2968 } else { 2969 pr_debug("Sending DATA_FIN on subflow %p\n", ssk); 2970 tcp_send_ack(ssk); 2971 if (!mptcp_rtx_timer_pending(sk)) 2972 mptcp_reset_rtx_timer(sk); 2973 } 2974 break; 2975 } 2976 2977 release_sock(ssk); 2978 } 2979 2980 void mptcp_set_state(struct sock *sk, int state) 2981 { 2982 int oldstate = sk->sk_state; 2983 2984 switch (state) { 2985 case TCP_ESTABLISHED: 2986 if (oldstate != TCP_ESTABLISHED) 2987 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2988 break; 2989 case TCP_CLOSE_WAIT: 2990 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state: 2991 * MPTCP "accepted" sockets will be created later on. So no 2992 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT. 2993 */ 2994 break; 2995 default: 2996 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 2997 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2998 } 2999 3000 inet_sk_state_store(sk, state); 3001 } 3002 3003 static const unsigned char new_state[16] = { 3004 /* current state: new state: action: */ 3005 [0 /* (Invalid) */] = TCP_CLOSE, 3006 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3007 [TCP_SYN_SENT] = TCP_CLOSE, 3008 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3009 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 3010 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 3011 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 3012 [TCP_CLOSE] = TCP_CLOSE, 3013 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 3014 [TCP_LAST_ACK] = TCP_LAST_ACK, 3015 [TCP_LISTEN] = TCP_CLOSE, 3016 [TCP_CLOSING] = TCP_CLOSING, 3017 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 3018 }; 3019 3020 static int mptcp_close_state(struct sock *sk) 3021 { 3022 int next = (int)new_state[sk->sk_state]; 3023 int ns = next & TCP_STATE_MASK; 3024 3025 mptcp_set_state(sk, ns); 3026 3027 return next & TCP_ACTION_FIN; 3028 } 3029 3030 static void mptcp_check_send_data_fin(struct sock *sk) 3031 { 3032 struct mptcp_subflow_context *subflow; 3033 struct mptcp_sock *msk = mptcp_sk(sk); 3034 3035 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n", 3036 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 3037 msk->snd_nxt, msk->write_seq); 3038 3039 /* we still need to enqueue subflows or not really shutting down, 3040 * skip this 3041 */ 3042 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 3043 mptcp_send_head(sk)) 3044 return; 3045 3046 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3047 3048 mptcp_for_each_subflow(msk, subflow) { 3049 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 3050 3051 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 3052 } 3053 } 3054 3055 static void __mptcp_wr_shutdown(struct sock *sk) 3056 { 3057 struct mptcp_sock *msk = mptcp_sk(sk); 3058 3059 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n", 3060 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 3061 !!mptcp_send_head(sk)); 3062 3063 /* will be ignored by fallback sockets */ 3064 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 3065 WRITE_ONCE(msk->snd_data_fin_enable, 1); 3066 3067 mptcp_check_send_data_fin(sk); 3068 } 3069 3070 static void __mptcp_destroy_sock(struct sock *sk) 3071 { 3072 struct mptcp_sock *msk = mptcp_sk(sk); 3073 3074 pr_debug("msk=%p\n", msk); 3075 3076 might_sleep(); 3077 3078 mptcp_stop_rtx_timer(sk); 3079 sk_stop_timer(sk, &sk->sk_timer); 3080 msk->pm.status = 0; 3081 mptcp_release_sched(msk); 3082 3083 sk->sk_prot->destroy(sk); 3084 3085 sk_stream_kill_queues(sk); 3086 xfrm_sk_free_policy(sk); 3087 3088 sock_put(sk); 3089 } 3090 3091 void __mptcp_unaccepted_force_close(struct sock *sk) 3092 { 3093 sock_set_flag(sk, SOCK_DEAD); 3094 mptcp_do_fastclose(sk); 3095 __mptcp_destroy_sock(sk); 3096 } 3097 3098 static __poll_t mptcp_check_readable(struct sock *sk) 3099 { 3100 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0; 3101 } 3102 3103 static void mptcp_check_listen_stop(struct sock *sk) 3104 { 3105 struct sock *ssk; 3106 3107 if (inet_sk_state_load(sk) != TCP_LISTEN) 3108 return; 3109 3110 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 3111 ssk = mptcp_sk(sk)->first; 3112 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN)) 3113 return; 3114 3115 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 3116 tcp_set_state(ssk, TCP_CLOSE); 3117 mptcp_subflow_queue_clean(sk, ssk); 3118 inet_csk_listen_stop(ssk); 3119 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED); 3120 release_sock(ssk); 3121 } 3122 3123 bool __mptcp_close(struct sock *sk, long timeout) 3124 { 3125 struct mptcp_subflow_context *subflow; 3126 struct mptcp_sock *msk = mptcp_sk(sk); 3127 bool do_cancel_work = false; 3128 int subflows_alive = 0; 3129 3130 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3131 3132 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 3133 mptcp_check_listen_stop(sk); 3134 mptcp_set_state(sk, TCP_CLOSE); 3135 goto cleanup; 3136 } 3137 3138 if (mptcp_data_avail(msk) || timeout < 0) { 3139 /* If the msk has read data, or the caller explicitly ask it, 3140 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose 3141 */ 3142 mptcp_do_fastclose(sk); 3143 timeout = 0; 3144 } else if (mptcp_close_state(sk)) { 3145 __mptcp_wr_shutdown(sk); 3146 } 3147 3148 sk_stream_wait_close(sk, timeout); 3149 3150 cleanup: 3151 /* orphan all the subflows */ 3152 mptcp_for_each_subflow(msk, subflow) { 3153 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3154 bool slow = lock_sock_fast_nested(ssk); 3155 3156 subflows_alive += ssk->sk_state != TCP_CLOSE; 3157 3158 /* since the close timeout takes precedence on the fail one, 3159 * cancel the latter 3160 */ 3161 if (ssk == msk->first) 3162 subflow->fail_tout = 0; 3163 3164 /* detach from the parent socket, but allow data_ready to 3165 * push incoming data into the mptcp stack, to properly ack it 3166 */ 3167 ssk->sk_socket = NULL; 3168 ssk->sk_wq = NULL; 3169 unlock_sock_fast(ssk, slow); 3170 } 3171 sock_orphan(sk); 3172 3173 /* all the subflows are closed, only timeout can change the msk 3174 * state, let's not keep resources busy for no reasons 3175 */ 3176 if (subflows_alive == 0) 3177 mptcp_set_state(sk, TCP_CLOSE); 3178 3179 sock_hold(sk); 3180 pr_debug("msk=%p state=%d\n", sk, sk->sk_state); 3181 mptcp_pm_connection_closed(msk); 3182 3183 if (sk->sk_state == TCP_CLOSE) { 3184 __mptcp_destroy_sock(sk); 3185 do_cancel_work = true; 3186 } else { 3187 mptcp_start_tout_timer(sk); 3188 } 3189 3190 return do_cancel_work; 3191 } 3192 3193 static void mptcp_close(struct sock *sk, long timeout) 3194 { 3195 bool do_cancel_work; 3196 3197 lock_sock(sk); 3198 3199 do_cancel_work = __mptcp_close(sk, timeout); 3200 release_sock(sk); 3201 if (do_cancel_work) 3202 mptcp_cancel_work(sk); 3203 3204 sock_put(sk); 3205 } 3206 3207 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 3208 { 3209 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3210 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 3211 struct ipv6_pinfo *msk6 = inet6_sk(msk); 3212 3213 msk->sk_v6_daddr = ssk->sk_v6_daddr; 3214 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 3215 3216 if (msk6 && ssk6) { 3217 msk6->saddr = ssk6->saddr; 3218 msk6->flow_label = ssk6->flow_label; 3219 } 3220 #endif 3221 3222 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 3223 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 3224 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 3225 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 3226 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 3227 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 3228 } 3229 3230 static int mptcp_disconnect(struct sock *sk, int flags) 3231 { 3232 struct mptcp_sock *msk = mptcp_sk(sk); 3233 3234 /* We are on the fastopen error path. We can't call straight into the 3235 * subflows cleanup code due to lock nesting (we are already under 3236 * msk->firstsocket lock). 3237 */ 3238 if (msk->fastopening) 3239 return -EBUSY; 3240 3241 mptcp_check_listen_stop(sk); 3242 mptcp_set_state(sk, TCP_CLOSE); 3243 3244 mptcp_stop_rtx_timer(sk); 3245 mptcp_stop_tout_timer(sk); 3246 3247 mptcp_pm_connection_closed(msk); 3248 3249 /* msk->subflow is still intact, the following will not free the first 3250 * subflow 3251 */ 3252 mptcp_do_fastclose(sk); 3253 mptcp_destroy_common(msk); 3254 3255 /* The first subflow is already in TCP_CLOSE status, the following 3256 * can't overlap with a fallback anymore 3257 */ 3258 spin_lock_bh(&msk->fallback_lock); 3259 msk->allow_subflows = true; 3260 msk->allow_infinite_fallback = true; 3261 WRITE_ONCE(msk->flags, 0); 3262 spin_unlock_bh(&msk->fallback_lock); 3263 3264 msk->cb_flags = 0; 3265 msk->recovery = false; 3266 WRITE_ONCE(msk->can_ack, false); 3267 WRITE_ONCE(msk->fully_established, false); 3268 WRITE_ONCE(msk->rcv_data_fin, false); 3269 WRITE_ONCE(msk->snd_data_fin_enable, false); 3270 WRITE_ONCE(msk->rcv_fastclose, false); 3271 WRITE_ONCE(msk->use_64bit_ack, false); 3272 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 3273 mptcp_pm_data_reset(msk); 3274 mptcp_ca_reset(sk); 3275 msk->bytes_consumed = 0; 3276 msk->bytes_acked = 0; 3277 msk->bytes_received = 0; 3278 msk->bytes_sent = 0; 3279 msk->bytes_retrans = 0; 3280 msk->rcvspace_init = 0; 3281 3282 WRITE_ONCE(sk->sk_shutdown, 0); 3283 sk_error_report(sk); 3284 return 0; 3285 } 3286 3287 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3288 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 3289 { 3290 struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk); 3291 3292 return &msk6->np; 3293 } 3294 3295 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk) 3296 { 3297 const struct ipv6_pinfo *np = inet6_sk(sk); 3298 struct ipv6_txoptions *opt; 3299 struct ipv6_pinfo *newnp; 3300 3301 newnp = inet6_sk(newsk); 3302 3303 rcu_read_lock(); 3304 opt = rcu_dereference(np->opt); 3305 if (opt) { 3306 opt = ipv6_dup_options(newsk, opt); 3307 if (!opt) 3308 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__); 3309 } 3310 RCU_INIT_POINTER(newnp->opt, opt); 3311 rcu_read_unlock(); 3312 } 3313 #endif 3314 3315 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk) 3316 { 3317 struct ip_options_rcu *inet_opt, *newopt = NULL; 3318 const struct inet_sock *inet = inet_sk(sk); 3319 struct inet_sock *newinet; 3320 3321 newinet = inet_sk(newsk); 3322 3323 rcu_read_lock(); 3324 inet_opt = rcu_dereference(inet->inet_opt); 3325 if (inet_opt) { 3326 newopt = sock_kmemdup(newsk, inet_opt, sizeof(*inet_opt) + 3327 inet_opt->opt.optlen, GFP_ATOMIC); 3328 if (!newopt) 3329 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__); 3330 } 3331 RCU_INIT_POINTER(newinet->inet_opt, newopt); 3332 rcu_read_unlock(); 3333 } 3334 3335 struct sock *mptcp_sk_clone_init(const struct sock *sk, 3336 const struct mptcp_options_received *mp_opt, 3337 struct sock *ssk, 3338 struct request_sock *req) 3339 { 3340 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 3341 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 3342 struct mptcp_subflow_context *subflow; 3343 struct mptcp_sock *msk; 3344 3345 if (!nsk) 3346 return NULL; 3347 3348 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3349 if (nsk->sk_family == AF_INET6) 3350 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 3351 #endif 3352 3353 __mptcp_init_sock(nsk); 3354 3355 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3356 if (nsk->sk_family == AF_INET6) 3357 mptcp_copy_ip6_options(nsk, sk); 3358 else 3359 #endif 3360 mptcp_copy_ip_options(nsk, sk); 3361 3362 msk = mptcp_sk(nsk); 3363 WRITE_ONCE(msk->local_key, subflow_req->local_key); 3364 WRITE_ONCE(msk->token, subflow_req->token); 3365 msk->in_accept_queue = 1; 3366 WRITE_ONCE(msk->fully_established, false); 3367 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 3368 WRITE_ONCE(msk->csum_enabled, true); 3369 3370 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1); 3371 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3372 WRITE_ONCE(msk->snd_una, msk->write_seq); 3373 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 3374 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 3375 mptcp_init_sched(msk, mptcp_sk(sk)->sched); 3376 3377 /* passive msk is created after the first/MPC subflow */ 3378 msk->subflow_id = 2; 3379 3380 sock_reset_flag(nsk, SOCK_RCU_FREE); 3381 security_inet_csk_clone(nsk, req); 3382 3383 /* this can't race with mptcp_close(), as the msk is 3384 * not yet exposted to user-space 3385 */ 3386 mptcp_set_state(nsk, TCP_ESTABLISHED); 3387 3388 /* The msk maintain a ref to each subflow in the connections list */ 3389 WRITE_ONCE(msk->first, ssk); 3390 subflow = mptcp_subflow_ctx(ssk); 3391 list_add(&subflow->node, &msk->conn_list); 3392 sock_hold(ssk); 3393 3394 /* new mpc subflow takes ownership of the newly 3395 * created mptcp socket 3396 */ 3397 mptcp_token_accept(subflow_req, msk); 3398 3399 /* set msk addresses early to ensure mptcp_pm_get_local_id() 3400 * uses the correct data 3401 */ 3402 mptcp_copy_inaddrs(nsk, ssk); 3403 __mptcp_propagate_sndbuf(nsk, ssk); 3404 3405 mptcp_rcv_space_init(msk, ssk); 3406 3407 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 3408 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 3409 bh_unlock_sock(nsk); 3410 3411 /* note: the newly allocated socket refcount is 2 now */ 3412 return nsk; 3413 } 3414 3415 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3416 { 3417 const struct tcp_sock *tp = tcp_sk(ssk); 3418 3419 msk->rcvspace_init = 1; 3420 msk->rcvq_space.copied = 0; 3421 msk->rcvq_space.rtt_us = 0; 3422 3423 msk->rcvq_space.time = tp->tcp_mstamp; 3424 3425 /* initial rcv_space offering made to peer */ 3426 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3427 TCP_INIT_CWND * tp->advmss); 3428 if (msk->rcvq_space.space == 0) 3429 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3430 } 3431 3432 void mptcp_destroy_common(struct mptcp_sock *msk) 3433 { 3434 struct mptcp_subflow_context *subflow, *tmp; 3435 struct sock *sk = (struct sock *)msk; 3436 3437 __mptcp_clear_xmit(sk); 3438 3439 /* join list will be eventually flushed (with rst) at sock lock release time */ 3440 mptcp_for_each_subflow_safe(msk, subflow, tmp) 3441 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, 0); 3442 3443 __skb_queue_purge(&sk->sk_receive_queue); 3444 skb_rbtree_purge(&msk->out_of_order_queue); 3445 3446 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3447 * inet_sock_destruct() will dispose it 3448 */ 3449 mptcp_token_destroy(msk); 3450 mptcp_pm_destroy(msk); 3451 } 3452 3453 static void mptcp_destroy(struct sock *sk) 3454 { 3455 struct mptcp_sock *msk = mptcp_sk(sk); 3456 3457 /* allow the following to close even the initial subflow */ 3458 msk->free_first = 1; 3459 mptcp_destroy_common(msk); 3460 sk_sockets_allocated_dec(sk); 3461 } 3462 3463 void __mptcp_data_acked(struct sock *sk) 3464 { 3465 if (!sock_owned_by_user(sk)) 3466 __mptcp_clean_una(sk); 3467 else 3468 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3469 } 3470 3471 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3472 { 3473 if (!sock_owned_by_user(sk)) 3474 __mptcp_subflow_push_pending(sk, ssk, false); 3475 else 3476 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3477 } 3478 3479 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3480 BIT(MPTCP_RETRANSMIT) | \ 3481 BIT(MPTCP_FLUSH_JOIN_LIST) | \ 3482 BIT(MPTCP_DEQUEUE)) 3483 3484 /* processes deferred events and flush wmem */ 3485 static void mptcp_release_cb(struct sock *sk) 3486 __must_hold(&sk->sk_lock.slock) 3487 { 3488 struct mptcp_sock *msk = mptcp_sk(sk); 3489 3490 for (;;) { 3491 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED); 3492 struct list_head join_list; 3493 3494 if (!flags) 3495 break; 3496 3497 INIT_LIST_HEAD(&join_list); 3498 list_splice_init(&msk->join_list, &join_list); 3499 3500 /* the following actions acquire the subflow socket lock 3501 * 3502 * 1) can't be invoked in atomic scope 3503 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3504 * datapath acquires the msk socket spinlock while helding 3505 * the subflow socket lock 3506 */ 3507 msk->cb_flags &= ~flags; 3508 spin_unlock_bh(&sk->sk_lock.slock); 3509 3510 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3511 __mptcp_flush_join_list(sk, &join_list); 3512 if (flags & BIT(MPTCP_PUSH_PENDING)) 3513 __mptcp_push_pending(sk, 0); 3514 if (flags & BIT(MPTCP_RETRANSMIT)) 3515 __mptcp_retrans(sk); 3516 if ((flags & BIT(MPTCP_DEQUEUE)) && __mptcp_move_skbs(sk)) { 3517 /* notify ack seq update */ 3518 mptcp_cleanup_rbuf(msk, 0); 3519 sk->sk_data_ready(sk); 3520 } 3521 3522 cond_resched(); 3523 spin_lock_bh(&sk->sk_lock.slock); 3524 } 3525 3526 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3527 __mptcp_clean_una_wakeup(sk); 3528 if (unlikely(msk->cb_flags)) { 3529 /* be sure to sync the msk state before taking actions 3530 * depending on sk_state (MPTCP_ERROR_REPORT) 3531 * On sk release avoid actions depending on the first subflow 3532 */ 3533 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first) 3534 __mptcp_sync_state(sk, msk->pending_state); 3535 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3536 __mptcp_error_report(sk); 3537 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags)) 3538 __mptcp_sync_sndbuf(sk); 3539 } 3540 } 3541 3542 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3543 * TCP can't schedule delack timer before the subflow is fully established. 3544 * MPTCP uses the delack timer to do 3rd ack retransmissions 3545 */ 3546 static void schedule_3rdack_retransmission(struct sock *ssk) 3547 { 3548 struct inet_connection_sock *icsk = inet_csk(ssk); 3549 struct tcp_sock *tp = tcp_sk(ssk); 3550 unsigned long timeout; 3551 3552 if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established)) 3553 return; 3554 3555 /* reschedule with a timeout above RTT, as we must look only for drop */ 3556 if (tp->srtt_us) 3557 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3558 else 3559 timeout = TCP_TIMEOUT_INIT; 3560 timeout += jiffies; 3561 3562 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3563 smp_store_release(&icsk->icsk_ack.pending, 3564 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER); 3565 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3566 } 3567 3568 void mptcp_subflow_process_delegated(struct sock *ssk, long status) 3569 { 3570 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3571 struct sock *sk = subflow->conn; 3572 3573 if (status & BIT(MPTCP_DELEGATE_SEND)) { 3574 mptcp_data_lock(sk); 3575 if (!sock_owned_by_user(sk)) 3576 __mptcp_subflow_push_pending(sk, ssk, true); 3577 else 3578 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3579 mptcp_data_unlock(sk); 3580 } 3581 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) { 3582 mptcp_data_lock(sk); 3583 if (!sock_owned_by_user(sk)) 3584 __mptcp_sync_sndbuf(sk); 3585 else 3586 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags); 3587 mptcp_data_unlock(sk); 3588 } 3589 if (status & BIT(MPTCP_DELEGATE_ACK)) 3590 schedule_3rdack_retransmission(ssk); 3591 } 3592 3593 static int mptcp_hash(struct sock *sk) 3594 { 3595 /* should never be called, 3596 * we hash the TCP subflows not the MPTCP socket 3597 */ 3598 WARN_ON_ONCE(1); 3599 return 0; 3600 } 3601 3602 static void mptcp_unhash(struct sock *sk) 3603 { 3604 /* called from sk_common_release(), but nothing to do here */ 3605 } 3606 3607 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3608 { 3609 struct mptcp_sock *msk = mptcp_sk(sk); 3610 3611 pr_debug("msk=%p, ssk=%p\n", msk, msk->first); 3612 if (WARN_ON_ONCE(!msk->first)) 3613 return -EINVAL; 3614 3615 return inet_csk_get_port(msk->first, snum); 3616 } 3617 3618 void mptcp_finish_connect(struct sock *ssk) 3619 { 3620 struct mptcp_subflow_context *subflow; 3621 struct mptcp_sock *msk; 3622 struct sock *sk; 3623 3624 subflow = mptcp_subflow_ctx(ssk); 3625 sk = subflow->conn; 3626 msk = mptcp_sk(sk); 3627 3628 pr_debug("msk=%p, token=%u\n", sk, subflow->token); 3629 3630 subflow->map_seq = subflow->iasn; 3631 subflow->map_subflow_seq = 1; 3632 3633 /* the socket is not connected yet, no msk/subflow ops can access/race 3634 * accessing the field below 3635 */ 3636 WRITE_ONCE(msk->local_key, subflow->local_key); 3637 3638 mptcp_pm_new_connection(msk, ssk, 0); 3639 } 3640 3641 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3642 { 3643 write_lock_bh(&sk->sk_callback_lock); 3644 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3645 sk_set_socket(sk, parent); 3646 write_unlock_bh(&sk->sk_callback_lock); 3647 } 3648 3649 bool mptcp_finish_join(struct sock *ssk) 3650 { 3651 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3652 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3653 struct sock *parent = (void *)msk; 3654 bool ret = true; 3655 3656 pr_debug("msk=%p, subflow=%p\n", msk, subflow); 3657 3658 /* mptcp socket already closing? */ 3659 if (!mptcp_is_fully_established(parent)) { 3660 subflow->reset_reason = MPTCP_RST_EMPTCP; 3661 return false; 3662 } 3663 3664 /* active subflow, already present inside the conn_list */ 3665 if (!list_empty(&subflow->node)) { 3666 spin_lock_bh(&msk->fallback_lock); 3667 if (!msk->allow_subflows) { 3668 spin_unlock_bh(&msk->fallback_lock); 3669 return false; 3670 } 3671 mptcp_subflow_joined(msk, ssk); 3672 spin_unlock_bh(&msk->fallback_lock); 3673 mptcp_propagate_sndbuf(parent, ssk); 3674 return true; 3675 } 3676 3677 if (!mptcp_pm_allow_new_subflow(msk)) { 3678 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_JOINREJECTED); 3679 goto err_prohibited; 3680 } 3681 3682 /* If we can't acquire msk socket lock here, let the release callback 3683 * handle it 3684 */ 3685 mptcp_data_lock(parent); 3686 if (!sock_owned_by_user(parent)) { 3687 ret = __mptcp_finish_join(msk, ssk); 3688 if (ret) { 3689 sock_hold(ssk); 3690 list_add_tail(&subflow->node, &msk->conn_list); 3691 } 3692 } else { 3693 sock_hold(ssk); 3694 list_add_tail(&subflow->node, &msk->join_list); 3695 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3696 } 3697 mptcp_data_unlock(parent); 3698 3699 if (!ret) { 3700 err_prohibited: 3701 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3702 return false; 3703 } 3704 3705 return true; 3706 } 3707 3708 static void mptcp_shutdown(struct sock *sk, int how) 3709 { 3710 pr_debug("sk=%p, how=%d\n", sk, how); 3711 3712 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3713 __mptcp_wr_shutdown(sk); 3714 } 3715 3716 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3717 { 3718 const struct sock *sk = (void *)msk; 3719 u64 delta; 3720 3721 if (sk->sk_state == TCP_LISTEN) 3722 return -EINVAL; 3723 3724 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3725 return 0; 3726 3727 delta = msk->write_seq - v; 3728 if (__mptcp_check_fallback(msk) && msk->first) { 3729 struct tcp_sock *tp = tcp_sk(msk->first); 3730 3731 /* the first subflow is disconnected after close - see 3732 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3733 * so ignore that status, too. 3734 */ 3735 if (!((1 << msk->first->sk_state) & 3736 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3737 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3738 } 3739 if (delta > INT_MAX) 3740 delta = INT_MAX; 3741 3742 return (int)delta; 3743 } 3744 3745 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg) 3746 { 3747 struct mptcp_sock *msk = mptcp_sk(sk); 3748 bool slow; 3749 3750 switch (cmd) { 3751 case SIOCINQ: 3752 if (sk->sk_state == TCP_LISTEN) 3753 return -EINVAL; 3754 3755 lock_sock(sk); 3756 if (__mptcp_move_skbs(sk)) 3757 mptcp_cleanup_rbuf(msk, 0); 3758 *karg = mptcp_inq_hint(sk); 3759 release_sock(sk); 3760 break; 3761 case SIOCOUTQ: 3762 slow = lock_sock_fast(sk); 3763 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3764 unlock_sock_fast(sk, slow); 3765 break; 3766 case SIOCOUTQNSD: 3767 slow = lock_sock_fast(sk); 3768 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt); 3769 unlock_sock_fast(sk, slow); 3770 break; 3771 default: 3772 return -ENOIOCTLCMD; 3773 } 3774 3775 return 0; 3776 } 3777 3778 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) 3779 { 3780 struct mptcp_subflow_context *subflow; 3781 struct mptcp_sock *msk = mptcp_sk(sk); 3782 int err = -EINVAL; 3783 struct sock *ssk; 3784 3785 ssk = __mptcp_nmpc_sk(msk); 3786 if (IS_ERR(ssk)) 3787 return PTR_ERR(ssk); 3788 3789 mptcp_set_state(sk, TCP_SYN_SENT); 3790 subflow = mptcp_subflow_ctx(ssk); 3791 #ifdef CONFIG_TCP_MD5SIG 3792 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3793 * TCP option space. 3794 */ 3795 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info)) 3796 mptcp_early_fallback(msk, subflow, MPTCP_MIB_MD5SIGFALLBACK); 3797 #endif 3798 if (subflow->request_mptcp) { 3799 if (mptcp_active_should_disable(sk)) 3800 mptcp_early_fallback(msk, subflow, 3801 MPTCP_MIB_MPCAPABLEACTIVEDISABLED); 3802 else if (mptcp_token_new_connect(ssk) < 0) 3803 mptcp_early_fallback(msk, subflow, 3804 MPTCP_MIB_TOKENFALLBACKINIT); 3805 } 3806 3807 WRITE_ONCE(msk->write_seq, subflow->idsn); 3808 WRITE_ONCE(msk->snd_nxt, subflow->idsn); 3809 WRITE_ONCE(msk->snd_una, subflow->idsn); 3810 if (likely(!__mptcp_check_fallback(msk))) 3811 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE); 3812 3813 /* if reaching here via the fastopen/sendmsg path, the caller already 3814 * acquired the subflow socket lock, too. 3815 */ 3816 if (!msk->fastopening) 3817 lock_sock(ssk); 3818 3819 /* the following mirrors closely a very small chunk of code from 3820 * __inet_stream_connect() 3821 */ 3822 if (ssk->sk_state != TCP_CLOSE) 3823 goto out; 3824 3825 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) { 3826 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len); 3827 if (err) 3828 goto out; 3829 } 3830 3831 err = ssk->sk_prot->connect(ssk, uaddr, addr_len); 3832 if (err < 0) 3833 goto out; 3834 3835 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk)); 3836 3837 out: 3838 if (!msk->fastopening) 3839 release_sock(ssk); 3840 3841 /* on successful connect, the msk state will be moved to established by 3842 * subflow_finish_connect() 3843 */ 3844 if (unlikely(err)) { 3845 /* avoid leaving a dangling token in an unconnected socket */ 3846 mptcp_token_destroy(msk); 3847 mptcp_set_state(sk, TCP_CLOSE); 3848 return err; 3849 } 3850 3851 mptcp_copy_inaddrs(sk, ssk); 3852 return 0; 3853 } 3854 3855 static struct proto mptcp_prot = { 3856 .name = "MPTCP", 3857 .owner = THIS_MODULE, 3858 .init = mptcp_init_sock, 3859 .connect = mptcp_connect, 3860 .disconnect = mptcp_disconnect, 3861 .close = mptcp_close, 3862 .setsockopt = mptcp_setsockopt, 3863 .getsockopt = mptcp_getsockopt, 3864 .shutdown = mptcp_shutdown, 3865 .destroy = mptcp_destroy, 3866 .sendmsg = mptcp_sendmsg, 3867 .ioctl = mptcp_ioctl, 3868 .recvmsg = mptcp_recvmsg, 3869 .release_cb = mptcp_release_cb, 3870 .hash = mptcp_hash, 3871 .unhash = mptcp_unhash, 3872 .get_port = mptcp_get_port, 3873 .stream_memory_free = mptcp_stream_memory_free, 3874 .sockets_allocated = &mptcp_sockets_allocated, 3875 3876 .memory_allocated = &net_aligned_data.tcp_memory_allocated, 3877 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc, 3878 3879 .memory_pressure = &tcp_memory_pressure, 3880 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3881 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3882 .sysctl_mem = sysctl_tcp_mem, 3883 .obj_size = sizeof(struct mptcp_sock), 3884 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3885 .no_autobind = true, 3886 }; 3887 3888 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) 3889 { 3890 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3891 struct sock *ssk, *sk = sock->sk; 3892 int err = -EINVAL; 3893 3894 lock_sock(sk); 3895 ssk = __mptcp_nmpc_sk(msk); 3896 if (IS_ERR(ssk)) { 3897 err = PTR_ERR(ssk); 3898 goto unlock; 3899 } 3900 3901 if (sk->sk_family == AF_INET) 3902 err = inet_bind_sk(ssk, uaddr, addr_len); 3903 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3904 else if (sk->sk_family == AF_INET6) 3905 err = inet6_bind_sk(ssk, uaddr, addr_len); 3906 #endif 3907 if (!err) 3908 mptcp_copy_inaddrs(sk, ssk); 3909 3910 unlock: 3911 release_sock(sk); 3912 return err; 3913 } 3914 3915 static int mptcp_listen(struct socket *sock, int backlog) 3916 { 3917 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3918 struct sock *sk = sock->sk; 3919 struct sock *ssk; 3920 int err; 3921 3922 pr_debug("msk=%p\n", msk); 3923 3924 lock_sock(sk); 3925 3926 err = -EINVAL; 3927 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM) 3928 goto unlock; 3929 3930 ssk = __mptcp_nmpc_sk(msk); 3931 if (IS_ERR(ssk)) { 3932 err = PTR_ERR(ssk); 3933 goto unlock; 3934 } 3935 3936 mptcp_set_state(sk, TCP_LISTEN); 3937 sock_set_flag(sk, SOCK_RCU_FREE); 3938 3939 lock_sock(ssk); 3940 err = __inet_listen_sk(ssk, backlog); 3941 release_sock(ssk); 3942 mptcp_set_state(sk, inet_sk_state_load(ssk)); 3943 3944 if (!err) { 3945 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 3946 mptcp_copy_inaddrs(sk, ssk); 3947 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED); 3948 } 3949 3950 unlock: 3951 release_sock(sk); 3952 return err; 3953 } 3954 3955 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3956 struct proto_accept_arg *arg) 3957 { 3958 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3959 struct sock *ssk, *newsk; 3960 3961 pr_debug("msk=%p\n", msk); 3962 3963 /* Buggy applications can call accept on socket states other then LISTEN 3964 * but no need to allocate the first subflow just to error out. 3965 */ 3966 ssk = READ_ONCE(msk->first); 3967 if (!ssk) 3968 return -EINVAL; 3969 3970 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk)); 3971 newsk = inet_csk_accept(ssk, arg); 3972 if (!newsk) 3973 return arg->err; 3974 3975 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk)); 3976 if (sk_is_mptcp(newsk)) { 3977 struct mptcp_subflow_context *subflow; 3978 struct sock *new_mptcp_sock; 3979 3980 subflow = mptcp_subflow_ctx(newsk); 3981 new_mptcp_sock = subflow->conn; 3982 3983 /* is_mptcp should be false if subflow->conn is missing, see 3984 * subflow_syn_recv_sock() 3985 */ 3986 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3987 tcp_sk(newsk)->is_mptcp = 0; 3988 goto tcpfallback; 3989 } 3990 3991 newsk = new_mptcp_sock; 3992 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3993 3994 newsk->sk_kern_sock = arg->kern; 3995 lock_sock(newsk); 3996 __inet_accept(sock, newsock, newsk); 3997 3998 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags); 3999 msk = mptcp_sk(newsk); 4000 msk->in_accept_queue = 0; 4001 4002 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 4003 * This is needed so NOSPACE flag can be set from tcp stack. 4004 */ 4005 mptcp_for_each_subflow(msk, subflow) { 4006 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4007 4008 if (!ssk->sk_socket) 4009 mptcp_sock_graft(ssk, newsock); 4010 } 4011 4012 mptcp_rps_record_subflows(msk); 4013 4014 /* Do late cleanup for the first subflow as necessary. Also 4015 * deal with bad peers not doing a complete shutdown. 4016 */ 4017 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) { 4018 __mptcp_close_ssk(newsk, msk->first, 4019 mptcp_subflow_ctx(msk->first), 0); 4020 if (unlikely(list_is_singular(&msk->conn_list))) 4021 mptcp_set_state(newsk, TCP_CLOSE); 4022 } 4023 } else { 4024 tcpfallback: 4025 newsk->sk_kern_sock = arg->kern; 4026 lock_sock(newsk); 4027 __inet_accept(sock, newsock, newsk); 4028 /* we are being invoked after accepting a non-mp-capable 4029 * flow: sk is a tcp_sk, not an mptcp one. 4030 * 4031 * Hand the socket over to tcp so all further socket ops 4032 * bypass mptcp. 4033 */ 4034 WRITE_ONCE(newsock->sk->sk_socket->ops, 4035 mptcp_fallback_tcp_ops(newsock->sk)); 4036 } 4037 release_sock(newsk); 4038 4039 return 0; 4040 } 4041 4042 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 4043 { 4044 struct sock *sk = (struct sock *)msk; 4045 4046 if (__mptcp_stream_is_writeable(sk, 1)) 4047 return EPOLLOUT | EPOLLWRNORM; 4048 4049 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 4050 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */ 4051 if (__mptcp_stream_is_writeable(sk, 1)) 4052 return EPOLLOUT | EPOLLWRNORM; 4053 4054 return 0; 4055 } 4056 4057 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 4058 struct poll_table_struct *wait) 4059 { 4060 struct sock *sk = sock->sk; 4061 struct mptcp_sock *msk; 4062 __poll_t mask = 0; 4063 u8 shutdown; 4064 int state; 4065 4066 msk = mptcp_sk(sk); 4067 sock_poll_wait(file, sock, wait); 4068 4069 state = inet_sk_state_load(sk); 4070 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags); 4071 if (state == TCP_LISTEN) { 4072 struct sock *ssk = READ_ONCE(msk->first); 4073 4074 if (WARN_ON_ONCE(!ssk)) 4075 return 0; 4076 4077 return inet_csk_listen_poll(ssk); 4078 } 4079 4080 shutdown = READ_ONCE(sk->sk_shutdown); 4081 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 4082 mask |= EPOLLHUP; 4083 if (shutdown & RCV_SHUTDOWN) 4084 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 4085 4086 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 4087 mask |= mptcp_check_readable(sk); 4088 if (shutdown & SEND_SHUTDOWN) 4089 mask |= EPOLLOUT | EPOLLWRNORM; 4090 else 4091 mask |= mptcp_check_writeable(msk); 4092 } else if (state == TCP_SYN_SENT && 4093 inet_test_bit(DEFER_CONNECT, sk)) { 4094 /* cf tcp_poll() note about TFO */ 4095 mask |= EPOLLOUT | EPOLLWRNORM; 4096 } 4097 4098 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */ 4099 smp_rmb(); 4100 if (READ_ONCE(sk->sk_err)) 4101 mask |= EPOLLERR; 4102 4103 return mask; 4104 } 4105 4106 static const struct proto_ops mptcp_stream_ops = { 4107 .family = PF_INET, 4108 .owner = THIS_MODULE, 4109 .release = inet_release, 4110 .bind = mptcp_bind, 4111 .connect = inet_stream_connect, 4112 .socketpair = sock_no_socketpair, 4113 .accept = mptcp_stream_accept, 4114 .getname = inet_getname, 4115 .poll = mptcp_poll, 4116 .ioctl = inet_ioctl, 4117 .gettstamp = sock_gettstamp, 4118 .listen = mptcp_listen, 4119 .shutdown = inet_shutdown, 4120 .setsockopt = sock_common_setsockopt, 4121 .getsockopt = sock_common_getsockopt, 4122 .sendmsg = inet_sendmsg, 4123 .recvmsg = inet_recvmsg, 4124 .mmap = sock_no_mmap, 4125 .set_rcvlowat = mptcp_set_rcvlowat, 4126 }; 4127 4128 static struct inet_protosw mptcp_protosw = { 4129 .type = SOCK_STREAM, 4130 .protocol = IPPROTO_MPTCP, 4131 .prot = &mptcp_prot, 4132 .ops = &mptcp_stream_ops, 4133 .flags = INET_PROTOSW_ICSK, 4134 }; 4135 4136 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 4137 { 4138 struct mptcp_delegated_action *delegated; 4139 struct mptcp_subflow_context *subflow; 4140 int work_done = 0; 4141 4142 delegated = container_of(napi, struct mptcp_delegated_action, napi); 4143 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 4144 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4145 4146 bh_lock_sock_nested(ssk); 4147 if (!sock_owned_by_user(ssk)) { 4148 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0)); 4149 } else { 4150 /* tcp_release_cb_override already processed 4151 * the action or will do at next release_sock(). 4152 * In both case must dequeue the subflow here - on the same 4153 * CPU that scheduled it. 4154 */ 4155 smp_wmb(); 4156 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status); 4157 } 4158 bh_unlock_sock(ssk); 4159 sock_put(ssk); 4160 4161 if (++work_done == budget) 4162 return budget; 4163 } 4164 4165 /* always provide a 0 'work_done' argument, so that napi_complete_done 4166 * will not try accessing the NULL napi->dev ptr 4167 */ 4168 napi_complete_done(napi, 0); 4169 return work_done; 4170 } 4171 4172 void __init mptcp_proto_init(void) 4173 { 4174 struct mptcp_delegated_action *delegated; 4175 int cpu; 4176 4177 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 4178 4179 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 4180 panic("Failed to allocate MPTCP pcpu counter\n"); 4181 4182 mptcp_napi_dev = alloc_netdev_dummy(0); 4183 if (!mptcp_napi_dev) 4184 panic("Failed to allocate MPTCP dummy netdev\n"); 4185 for_each_possible_cpu(cpu) { 4186 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 4187 INIT_LIST_HEAD(&delegated->head); 4188 netif_napi_add_tx(mptcp_napi_dev, &delegated->napi, 4189 mptcp_napi_poll); 4190 napi_enable(&delegated->napi); 4191 } 4192 4193 mptcp_subflow_init(); 4194 mptcp_pm_init(); 4195 mptcp_sched_init(); 4196 mptcp_token_init(); 4197 4198 if (proto_register(&mptcp_prot, 1) != 0) 4199 panic("Failed to register MPTCP proto.\n"); 4200 4201 inet_register_protosw(&mptcp_protosw); 4202 4203 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 4204 } 4205 4206 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 4207 static const struct proto_ops mptcp_v6_stream_ops = { 4208 .family = PF_INET6, 4209 .owner = THIS_MODULE, 4210 .release = inet6_release, 4211 .bind = mptcp_bind, 4212 .connect = inet_stream_connect, 4213 .socketpair = sock_no_socketpair, 4214 .accept = mptcp_stream_accept, 4215 .getname = inet6_getname, 4216 .poll = mptcp_poll, 4217 .ioctl = inet6_ioctl, 4218 .gettstamp = sock_gettstamp, 4219 .listen = mptcp_listen, 4220 .shutdown = inet_shutdown, 4221 .setsockopt = sock_common_setsockopt, 4222 .getsockopt = sock_common_getsockopt, 4223 .sendmsg = inet6_sendmsg, 4224 .recvmsg = inet6_recvmsg, 4225 .mmap = sock_no_mmap, 4226 #ifdef CONFIG_COMPAT 4227 .compat_ioctl = inet6_compat_ioctl, 4228 #endif 4229 .set_rcvlowat = mptcp_set_rcvlowat, 4230 }; 4231 4232 static struct proto mptcp_v6_prot; 4233 4234 static struct inet_protosw mptcp_v6_protosw = { 4235 .type = SOCK_STREAM, 4236 .protocol = IPPROTO_MPTCP, 4237 .prot = &mptcp_v6_prot, 4238 .ops = &mptcp_v6_stream_ops, 4239 .flags = INET_PROTOSW_ICSK, 4240 }; 4241 4242 int __init mptcp_proto_v6_init(void) 4243 { 4244 int err; 4245 4246 mptcp_v6_prot = mptcp_prot; 4247 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name)); 4248 mptcp_v6_prot.slab = NULL; 4249 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 4250 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np); 4251 4252 err = proto_register(&mptcp_v6_prot, 1); 4253 if (err) 4254 return err; 4255 4256 err = inet6_register_protosw(&mptcp_v6_protosw); 4257 if (err) 4258 proto_unregister(&mptcp_v6_prot); 4259 4260 return err; 4261 } 4262 #endif 4263