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 && !ssk->sk_bypass_prot_mem) { 1092 tcp_enter_memory_pressure(ssk); 1093 first = false; 1094 } 1095 1096 sk_stream_moderate_sndbuf(ssk); 1097 } 1098 __mptcp_sync_sndbuf(sk); 1099 } 1100 1101 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of 1102 * data 1103 */ 1104 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag) 1105 { 1106 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag), 1107 pfrag, sk->sk_allocation))) 1108 return true; 1109 1110 mptcp_enter_memory_pressure(sk); 1111 return false; 1112 } 1113 1114 static struct mptcp_data_frag * 1115 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag, 1116 int orig_offset) 1117 { 1118 int offset = ALIGN(orig_offset, sizeof(long)); 1119 struct mptcp_data_frag *dfrag; 1120 1121 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset); 1122 dfrag->data_len = 0; 1123 dfrag->data_seq = msk->write_seq; 1124 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag); 1125 dfrag->offset = offset + sizeof(struct mptcp_data_frag); 1126 dfrag->already_sent = 0; 1127 dfrag->page = pfrag->page; 1128 1129 return dfrag; 1130 } 1131 1132 struct mptcp_sendmsg_info { 1133 int mss_now; 1134 int size_goal; 1135 u16 limit; 1136 u16 sent; 1137 unsigned int flags; 1138 bool data_lock_held; 1139 }; 1140 1141 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk, 1142 u64 data_seq, int avail_size) 1143 { 1144 u64 window_end = mptcp_wnd_end(msk); 1145 u64 mptcp_snd_wnd; 1146 1147 if (__mptcp_check_fallback(msk)) 1148 return avail_size; 1149 1150 mptcp_snd_wnd = window_end - data_seq; 1151 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size); 1152 1153 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) { 1154 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd); 1155 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED); 1156 } 1157 1158 return avail_size; 1159 } 1160 1161 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp) 1162 { 1163 struct skb_ext *mpext = __skb_ext_alloc(gfp); 1164 1165 if (!mpext) 1166 return false; 1167 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext); 1168 return true; 1169 } 1170 1171 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp) 1172 { 1173 struct sk_buff *skb; 1174 1175 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp); 1176 if (likely(skb)) { 1177 if (likely(__mptcp_add_ext(skb, gfp))) { 1178 skb_reserve(skb, MAX_TCP_HEADER); 1179 skb->ip_summed = CHECKSUM_PARTIAL; 1180 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 1181 return skb; 1182 } 1183 __kfree_skb(skb); 1184 } else { 1185 mptcp_enter_memory_pressure(sk); 1186 } 1187 return NULL; 1188 } 1189 1190 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp) 1191 { 1192 struct sk_buff *skb; 1193 1194 skb = __mptcp_do_alloc_tx_skb(sk, gfp); 1195 if (!skb) 1196 return NULL; 1197 1198 if (likely(sk_wmem_schedule(ssk, skb->truesize))) { 1199 tcp_skb_entail(ssk, skb); 1200 return skb; 1201 } 1202 tcp_skb_tsorted_anchor_cleanup(skb); 1203 kfree_skb(skb); 1204 return NULL; 1205 } 1206 1207 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held) 1208 { 1209 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation; 1210 1211 return __mptcp_alloc_tx_skb(sk, ssk, gfp); 1212 } 1213 1214 /* note: this always recompute the csum on the whole skb, even 1215 * if we just appended a single frag. More status info needed 1216 */ 1217 static void mptcp_update_data_checksum(struct sk_buff *skb, int added) 1218 { 1219 struct mptcp_ext *mpext = mptcp_get_ext(skb); 1220 __wsum csum = ~csum_unfold(mpext->csum); 1221 int offset = skb->len - added; 1222 1223 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset)); 1224 } 1225 1226 static void mptcp_update_infinite_map(struct mptcp_sock *msk, 1227 struct sock *ssk, 1228 struct mptcp_ext *mpext) 1229 { 1230 if (!mpext) 1231 return; 1232 1233 mpext->infinite_map = 1; 1234 mpext->data_len = 0; 1235 1236 if (!mptcp_try_fallback(ssk, MPTCP_MIB_INFINITEMAPTX)) { 1237 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED); 1238 mptcp_subflow_reset(ssk); 1239 return; 1240 } 1241 1242 mptcp_subflow_ctx(ssk)->send_infinite_map = 0; 1243 } 1244 1245 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1)) 1246 1247 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk, 1248 struct mptcp_data_frag *dfrag, 1249 struct mptcp_sendmsg_info *info) 1250 { 1251 u64 data_seq = dfrag->data_seq + info->sent; 1252 int offset = dfrag->offset + info->sent; 1253 struct mptcp_sock *msk = mptcp_sk(sk); 1254 bool zero_window_probe = false; 1255 struct mptcp_ext *mpext = NULL; 1256 bool can_coalesce = false; 1257 bool reuse_skb = true; 1258 struct sk_buff *skb; 1259 size_t copy; 1260 int i; 1261 1262 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n", 1263 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent); 1264 1265 if (WARN_ON_ONCE(info->sent > info->limit || 1266 info->limit > dfrag->data_len)) 1267 return 0; 1268 1269 if (unlikely(!__tcp_can_send(ssk))) 1270 return -EAGAIN; 1271 1272 /* compute send limit */ 1273 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE)) 1274 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE; 1275 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags); 1276 copy = info->size_goal; 1277 1278 skb = tcp_write_queue_tail(ssk); 1279 if (skb && copy > skb->len) { 1280 /* Limit the write to the size available in the 1281 * current skb, if any, so that we create at most a new skb. 1282 * Explicitly tells TCP internals to avoid collapsing on later 1283 * queue management operation, to avoid breaking the ext <-> 1284 * SSN association set here 1285 */ 1286 mpext = mptcp_get_ext(skb); 1287 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) { 1288 TCP_SKB_CB(skb)->eor = 1; 1289 tcp_mark_push(tcp_sk(ssk), skb); 1290 goto alloc_skb; 1291 } 1292 1293 i = skb_shinfo(skb)->nr_frags; 1294 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset); 1295 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) { 1296 tcp_mark_push(tcp_sk(ssk), skb); 1297 goto alloc_skb; 1298 } 1299 1300 copy -= skb->len; 1301 } else { 1302 alloc_skb: 1303 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held); 1304 if (!skb) 1305 return -ENOMEM; 1306 1307 i = skb_shinfo(skb)->nr_frags; 1308 reuse_skb = false; 1309 mpext = mptcp_get_ext(skb); 1310 } 1311 1312 /* Zero window and all data acked? Probe. */ 1313 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy); 1314 if (copy == 0) { 1315 u64 snd_una = READ_ONCE(msk->snd_una); 1316 1317 /* No need for zero probe if there are any data pending 1318 * either at the msk or ssk level; skb is the current write 1319 * queue tail and can be empty at this point. 1320 */ 1321 if (snd_una != msk->snd_nxt || skb->len || 1322 skb != tcp_send_head(ssk)) { 1323 tcp_remove_empty_skb(ssk); 1324 return 0; 1325 } 1326 1327 zero_window_probe = true; 1328 data_seq = snd_una - 1; 1329 copy = 1; 1330 } 1331 1332 copy = min_t(size_t, copy, info->limit - info->sent); 1333 if (!sk_wmem_schedule(ssk, copy)) { 1334 tcp_remove_empty_skb(ssk); 1335 return -ENOMEM; 1336 } 1337 1338 if (can_coalesce) { 1339 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy); 1340 } else { 1341 get_page(dfrag->page); 1342 skb_fill_page_desc(skb, i, dfrag->page, offset, copy); 1343 } 1344 1345 skb->len += copy; 1346 skb->data_len += copy; 1347 skb->truesize += copy; 1348 sk_wmem_queued_add(ssk, copy); 1349 sk_mem_charge(ssk, copy); 1350 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy); 1351 TCP_SKB_CB(skb)->end_seq += copy; 1352 tcp_skb_pcount_set(skb, 0); 1353 1354 /* on skb reuse we just need to update the DSS len */ 1355 if (reuse_skb) { 1356 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH; 1357 mpext->data_len += copy; 1358 goto out; 1359 } 1360 1361 memset(mpext, 0, sizeof(*mpext)); 1362 mpext->data_seq = data_seq; 1363 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq; 1364 mpext->data_len = copy; 1365 mpext->use_map = 1; 1366 mpext->dsn64 = 1; 1367 1368 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n", 1369 mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1370 mpext->dsn64); 1371 1372 if (zero_window_probe) { 1373 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_WINPROBE); 1374 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1375 mpext->frozen = 1; 1376 if (READ_ONCE(msk->csum_enabled)) 1377 mptcp_update_data_checksum(skb, copy); 1378 tcp_push_pending_frames(ssk); 1379 return 0; 1380 } 1381 out: 1382 if (READ_ONCE(msk->csum_enabled)) 1383 mptcp_update_data_checksum(skb, copy); 1384 if (mptcp_subflow_ctx(ssk)->send_infinite_map) 1385 mptcp_update_infinite_map(msk, ssk, mpext); 1386 trace_mptcp_sendmsg_frag(mpext); 1387 mptcp_subflow_ctx(ssk)->rel_write_seq += copy; 1388 return copy; 1389 } 1390 1391 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \ 1392 sizeof(struct tcphdr) - \ 1393 MAX_TCP_OPTION_SPACE - \ 1394 sizeof(struct ipv6hdr) - \ 1395 sizeof(struct frag_hdr)) 1396 1397 struct subflow_send_info { 1398 struct sock *ssk; 1399 u64 linger_time; 1400 }; 1401 1402 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow) 1403 { 1404 if (!subflow->stale) 1405 return; 1406 1407 subflow->stale = 0; 1408 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER); 1409 } 1410 1411 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow) 1412 { 1413 if (unlikely(subflow->stale)) { 1414 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp); 1415 1416 if (subflow->stale_rcv_tstamp == rcv_tstamp) 1417 return false; 1418 1419 mptcp_subflow_set_active(subflow); 1420 } 1421 return __mptcp_subflow_active(subflow); 1422 } 1423 1424 #define SSK_MODE_ACTIVE 0 1425 #define SSK_MODE_BACKUP 1 1426 #define SSK_MODE_MAX 2 1427 1428 /* implement the mptcp packet scheduler; 1429 * returns the subflow that will transmit the next DSS 1430 * additionally updates the rtx timeout 1431 */ 1432 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk) 1433 { 1434 struct subflow_send_info send_info[SSK_MODE_MAX]; 1435 struct mptcp_subflow_context *subflow; 1436 struct sock *sk = (struct sock *)msk; 1437 u32 pace, burst, wmem; 1438 int i, nr_active = 0; 1439 struct sock *ssk; 1440 u64 linger_time; 1441 long tout = 0; 1442 1443 /* pick the subflow with the lower wmem/wspace ratio */ 1444 for (i = 0; i < SSK_MODE_MAX; ++i) { 1445 send_info[i].ssk = NULL; 1446 send_info[i].linger_time = -1; 1447 } 1448 1449 mptcp_for_each_subflow(msk, subflow) { 1450 bool backup = subflow->backup || subflow->request_bkup; 1451 1452 trace_mptcp_subflow_get_send(subflow); 1453 ssk = mptcp_subflow_tcp_sock(subflow); 1454 if (!mptcp_subflow_active(subflow)) 1455 continue; 1456 1457 tout = max(tout, mptcp_timeout_from_subflow(subflow)); 1458 nr_active += !backup; 1459 pace = subflow->avg_pacing_rate; 1460 if (unlikely(!pace)) { 1461 /* init pacing rate from socket */ 1462 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate); 1463 pace = subflow->avg_pacing_rate; 1464 if (!pace) 1465 continue; 1466 } 1467 1468 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace); 1469 if (linger_time < send_info[backup].linger_time) { 1470 send_info[backup].ssk = ssk; 1471 send_info[backup].linger_time = linger_time; 1472 } 1473 } 1474 __mptcp_set_timeout(sk, tout); 1475 1476 /* pick the best backup if no other subflow is active */ 1477 if (!nr_active) 1478 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk; 1479 1480 /* According to the blest algorithm, to avoid HoL blocking for the 1481 * faster flow, we need to: 1482 * - estimate the faster flow linger time 1483 * - use the above to estimate the amount of byte transferred 1484 * by the faster flow 1485 * - check that the amount of queued data is greater than the above, 1486 * otherwise do not use the picked, slower, subflow 1487 * We select the subflow with the shorter estimated time to flush 1488 * the queued mem, which basically ensure the above. We just need 1489 * to check that subflow has a non empty cwin. 1490 */ 1491 ssk = send_info[SSK_MODE_ACTIVE].ssk; 1492 if (!ssk || !sk_stream_memory_free(ssk)) 1493 return NULL; 1494 1495 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt); 1496 wmem = READ_ONCE(ssk->sk_wmem_queued); 1497 if (!burst) 1498 return ssk; 1499 1500 subflow = mptcp_subflow_ctx(ssk); 1501 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem + 1502 READ_ONCE(ssk->sk_pacing_rate) * burst, 1503 burst + wmem); 1504 msk->snd_burst = burst; 1505 return ssk; 1506 } 1507 1508 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info) 1509 { 1510 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal); 1511 release_sock(ssk); 1512 } 1513 1514 static void mptcp_update_post_push(struct mptcp_sock *msk, 1515 struct mptcp_data_frag *dfrag, 1516 u32 sent) 1517 { 1518 u64 snd_nxt_new = dfrag->data_seq; 1519 1520 dfrag->already_sent += sent; 1521 1522 msk->snd_burst -= sent; 1523 1524 snd_nxt_new += dfrag->already_sent; 1525 1526 /* snd_nxt_new can be smaller than snd_nxt in case mptcp 1527 * is recovering after a failover. In that event, this re-sends 1528 * old segments. 1529 * 1530 * Thus compute snd_nxt_new candidate based on 1531 * the dfrag->data_seq that was sent and the data 1532 * that has been handed to the subflow for transmission 1533 * and skip update in case it was old dfrag. 1534 */ 1535 if (likely(after64(snd_nxt_new, msk->snd_nxt))) { 1536 msk->bytes_sent += snd_nxt_new - msk->snd_nxt; 1537 WRITE_ONCE(msk->snd_nxt, snd_nxt_new); 1538 } 1539 } 1540 1541 void mptcp_check_and_set_pending(struct sock *sk) 1542 { 1543 if (mptcp_send_head(sk)) { 1544 mptcp_data_lock(sk); 1545 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING); 1546 mptcp_data_unlock(sk); 1547 } 1548 } 1549 1550 static int __subflow_push_pending(struct sock *sk, struct sock *ssk, 1551 struct mptcp_sendmsg_info *info) 1552 { 1553 struct mptcp_sock *msk = mptcp_sk(sk); 1554 struct mptcp_data_frag *dfrag; 1555 int len, copied = 0, err = 0; 1556 1557 while ((dfrag = mptcp_send_head(sk))) { 1558 info->sent = dfrag->already_sent; 1559 info->limit = dfrag->data_len; 1560 len = dfrag->data_len - dfrag->already_sent; 1561 while (len > 0) { 1562 int ret = 0; 1563 1564 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info); 1565 if (ret <= 0) { 1566 err = copied ? : ret; 1567 goto out; 1568 } 1569 1570 info->sent += ret; 1571 copied += ret; 1572 len -= ret; 1573 1574 mptcp_update_post_push(msk, dfrag, ret); 1575 } 1576 msk->first_pending = mptcp_send_next(sk); 1577 1578 if (msk->snd_burst <= 0 || 1579 !sk_stream_memory_free(ssk) || 1580 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) { 1581 err = copied; 1582 goto out; 1583 } 1584 mptcp_set_timeout(sk); 1585 } 1586 err = copied; 1587 1588 out: 1589 if (err > 0) 1590 msk->last_data_sent = tcp_jiffies32; 1591 return err; 1592 } 1593 1594 void __mptcp_push_pending(struct sock *sk, unsigned int flags) 1595 { 1596 struct sock *prev_ssk = NULL, *ssk = NULL; 1597 struct mptcp_sock *msk = mptcp_sk(sk); 1598 struct mptcp_sendmsg_info info = { 1599 .flags = flags, 1600 }; 1601 bool do_check_data_fin = false; 1602 int push_count = 1; 1603 1604 while (mptcp_send_head(sk) && (push_count > 0)) { 1605 struct mptcp_subflow_context *subflow; 1606 int ret = 0; 1607 1608 if (mptcp_sched_get_send(msk)) 1609 break; 1610 1611 push_count = 0; 1612 1613 mptcp_for_each_subflow(msk, subflow) { 1614 if (READ_ONCE(subflow->scheduled)) { 1615 mptcp_subflow_set_scheduled(subflow, false); 1616 1617 prev_ssk = ssk; 1618 ssk = mptcp_subflow_tcp_sock(subflow); 1619 if (ssk != prev_ssk) { 1620 /* First check. If the ssk has changed since 1621 * the last round, release prev_ssk 1622 */ 1623 if (prev_ssk) 1624 mptcp_push_release(prev_ssk, &info); 1625 1626 /* Need to lock the new subflow only if different 1627 * from the previous one, otherwise we are still 1628 * helding the relevant lock 1629 */ 1630 lock_sock(ssk); 1631 } 1632 1633 push_count++; 1634 1635 ret = __subflow_push_pending(sk, ssk, &info); 1636 if (ret <= 0) { 1637 if (ret != -EAGAIN || 1638 (1 << ssk->sk_state) & 1639 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE)) 1640 push_count--; 1641 continue; 1642 } 1643 do_check_data_fin = true; 1644 } 1645 } 1646 } 1647 1648 /* at this point we held the socket lock for the last subflow we used */ 1649 if (ssk) 1650 mptcp_push_release(ssk, &info); 1651 1652 /* ensure the rtx timer is running */ 1653 if (!mptcp_rtx_timer_pending(sk)) 1654 mptcp_reset_rtx_timer(sk); 1655 if (do_check_data_fin) 1656 mptcp_check_send_data_fin(sk); 1657 } 1658 1659 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first) 1660 { 1661 struct mptcp_sock *msk = mptcp_sk(sk); 1662 struct mptcp_sendmsg_info info = { 1663 .data_lock_held = true, 1664 }; 1665 bool keep_pushing = true; 1666 struct sock *xmit_ssk; 1667 int copied = 0; 1668 1669 info.flags = 0; 1670 while (mptcp_send_head(sk) && keep_pushing) { 1671 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1672 int ret = 0; 1673 1674 /* check for a different subflow usage only after 1675 * spooling the first chunk of data 1676 */ 1677 if (first) { 1678 mptcp_subflow_set_scheduled(subflow, false); 1679 ret = __subflow_push_pending(sk, ssk, &info); 1680 first = false; 1681 if (ret <= 0) 1682 break; 1683 copied += ret; 1684 continue; 1685 } 1686 1687 if (mptcp_sched_get_send(msk)) 1688 goto out; 1689 1690 if (READ_ONCE(subflow->scheduled)) { 1691 mptcp_subflow_set_scheduled(subflow, false); 1692 ret = __subflow_push_pending(sk, ssk, &info); 1693 if (ret <= 0) 1694 keep_pushing = false; 1695 copied += ret; 1696 } 1697 1698 mptcp_for_each_subflow(msk, subflow) { 1699 if (READ_ONCE(subflow->scheduled)) { 1700 xmit_ssk = mptcp_subflow_tcp_sock(subflow); 1701 if (xmit_ssk != ssk) { 1702 mptcp_subflow_delegate(subflow, 1703 MPTCP_DELEGATE_SEND); 1704 keep_pushing = false; 1705 } 1706 } 1707 } 1708 } 1709 1710 out: 1711 /* __mptcp_alloc_tx_skb could have released some wmem and we are 1712 * not going to flush it via release_sock() 1713 */ 1714 if (copied) { 1715 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 1716 info.size_goal); 1717 if (!mptcp_rtx_timer_pending(sk)) 1718 mptcp_reset_rtx_timer(sk); 1719 1720 if (msk->snd_data_fin_enable && 1721 msk->snd_nxt + 1 == msk->write_seq) 1722 mptcp_schedule_work(sk); 1723 } 1724 } 1725 1726 static int mptcp_disconnect(struct sock *sk, int flags); 1727 1728 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, 1729 size_t len, int *copied_syn) 1730 { 1731 unsigned int saved_flags = msg->msg_flags; 1732 struct mptcp_sock *msk = mptcp_sk(sk); 1733 struct sock *ssk; 1734 int ret; 1735 1736 /* on flags based fastopen the mptcp is supposed to create the 1737 * first subflow right now. Otherwise we are in the defer_connect 1738 * path, and the first subflow must be already present. 1739 * Since the defer_connect flag is cleared after the first succsful 1740 * fastopen attempt, no need to check for additional subflow status. 1741 */ 1742 if (msg->msg_flags & MSG_FASTOPEN) { 1743 ssk = __mptcp_nmpc_sk(msk); 1744 if (IS_ERR(ssk)) 1745 return PTR_ERR(ssk); 1746 } 1747 if (!msk->first) 1748 return -EINVAL; 1749 1750 ssk = msk->first; 1751 1752 lock_sock(ssk); 1753 msg->msg_flags |= MSG_DONTWAIT; 1754 msk->fastopening = 1; 1755 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL); 1756 msk->fastopening = 0; 1757 msg->msg_flags = saved_flags; 1758 release_sock(ssk); 1759 1760 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */ 1761 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) { 1762 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name, 1763 msg->msg_namelen, msg->msg_flags, 1); 1764 1765 /* Keep the same behaviour of plain TCP: zero the copied bytes in 1766 * case of any error, except timeout or signal 1767 */ 1768 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR) 1769 *copied_syn = 0; 1770 } else if (ret && ret != -EINPROGRESS) { 1771 /* The disconnect() op called by tcp_sendmsg_fastopen()/ 1772 * __inet_stream_connect() can fail, due to looking check, 1773 * see mptcp_disconnect(). 1774 * Attempt it again outside the problematic scope. 1775 */ 1776 if (!mptcp_disconnect(sk, 0)) { 1777 sk->sk_disconnects++; 1778 sk->sk_socket->state = SS_UNCONNECTED; 1779 } 1780 } 1781 inet_clear_bit(DEFER_CONNECT, sk); 1782 1783 return ret; 1784 } 1785 1786 static int do_copy_data_nocache(struct sock *sk, int copy, 1787 struct iov_iter *from, char *to) 1788 { 1789 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) { 1790 if (!copy_from_iter_full_nocache(to, copy, from)) 1791 return -EFAULT; 1792 } else if (!copy_from_iter_full(to, copy, from)) { 1793 return -EFAULT; 1794 } 1795 return 0; 1796 } 1797 1798 /* open-code sk_stream_memory_free() plus sent limit computation to 1799 * avoid indirect calls in fast-path. 1800 * Called under the msk socket lock, so we can avoid a bunch of ONCE 1801 * annotations. 1802 */ 1803 static u32 mptcp_send_limit(const struct sock *sk) 1804 { 1805 const struct mptcp_sock *msk = mptcp_sk(sk); 1806 u32 limit, not_sent; 1807 1808 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf)) 1809 return 0; 1810 1811 limit = mptcp_notsent_lowat(sk); 1812 if (limit == UINT_MAX) 1813 return UINT_MAX; 1814 1815 not_sent = msk->write_seq - msk->snd_nxt; 1816 if (not_sent >= limit) 1817 return 0; 1818 1819 return limit - not_sent; 1820 } 1821 1822 static void mptcp_rps_record_subflows(const struct mptcp_sock *msk) 1823 { 1824 struct mptcp_subflow_context *subflow; 1825 1826 if (!rfs_is_needed()) 1827 return; 1828 1829 mptcp_for_each_subflow(msk, subflow) { 1830 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 1831 1832 sock_rps_record_flow(ssk); 1833 } 1834 } 1835 1836 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) 1837 { 1838 struct mptcp_sock *msk = mptcp_sk(sk); 1839 struct page_frag *pfrag; 1840 size_t copied = 0; 1841 int ret = 0; 1842 long timeo; 1843 1844 /* silently ignore everything else */ 1845 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN; 1846 1847 lock_sock(sk); 1848 1849 mptcp_rps_record_subflows(msk); 1850 1851 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) || 1852 msg->msg_flags & MSG_FASTOPEN)) { 1853 int copied_syn = 0; 1854 1855 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn); 1856 copied += copied_syn; 1857 if (ret == -EINPROGRESS && copied_syn > 0) 1858 goto out; 1859 else if (ret) 1860 goto do_error; 1861 } 1862 1863 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT); 1864 1865 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) { 1866 ret = sk_stream_wait_connect(sk, &timeo); 1867 if (ret) 1868 goto do_error; 1869 } 1870 1871 ret = -EPIPE; 1872 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))) 1873 goto do_error; 1874 1875 pfrag = sk_page_frag(sk); 1876 1877 while (msg_data_left(msg)) { 1878 int total_ts, frag_truesize = 0; 1879 struct mptcp_data_frag *dfrag; 1880 bool dfrag_collapsed; 1881 size_t psize, offset; 1882 u32 copy_limit; 1883 1884 /* ensure fitting the notsent_lowat() constraint */ 1885 copy_limit = mptcp_send_limit(sk); 1886 if (!copy_limit) 1887 goto wait_for_memory; 1888 1889 /* reuse tail pfrag, if possible, or carve a new one from the 1890 * page allocator 1891 */ 1892 dfrag = mptcp_pending_tail(sk); 1893 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag); 1894 if (!dfrag_collapsed) { 1895 if (!mptcp_page_frag_refill(sk, pfrag)) 1896 goto wait_for_memory; 1897 1898 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset); 1899 frag_truesize = dfrag->overhead; 1900 } 1901 1902 /* we do not bound vs wspace, to allow a single packet. 1903 * memory accounting will prevent execessive memory usage 1904 * anyway 1905 */ 1906 offset = dfrag->offset + dfrag->data_len; 1907 psize = pfrag->size - offset; 1908 psize = min_t(size_t, psize, msg_data_left(msg)); 1909 psize = min_t(size_t, psize, copy_limit); 1910 total_ts = psize + frag_truesize; 1911 1912 if (!sk_wmem_schedule(sk, total_ts)) 1913 goto wait_for_memory; 1914 1915 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter, 1916 page_address(dfrag->page) + offset); 1917 if (ret) 1918 goto do_error; 1919 1920 /* data successfully copied into the write queue */ 1921 sk_forward_alloc_add(sk, -total_ts); 1922 copied += psize; 1923 dfrag->data_len += psize; 1924 frag_truesize += psize; 1925 pfrag->offset += frag_truesize; 1926 WRITE_ONCE(msk->write_seq, msk->write_seq + psize); 1927 1928 /* charge data on mptcp pending queue to the msk socket 1929 * Note: we charge such data both to sk and ssk 1930 */ 1931 sk_wmem_queued_add(sk, frag_truesize); 1932 if (!dfrag_collapsed) { 1933 get_page(dfrag->page); 1934 list_add_tail(&dfrag->list, &msk->rtx_queue); 1935 if (!msk->first_pending) 1936 msk->first_pending = dfrag; 1937 } 1938 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk, 1939 dfrag->data_seq, dfrag->data_len, dfrag->already_sent, 1940 !dfrag_collapsed); 1941 1942 continue; 1943 1944 wait_for_memory: 1945 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 1946 __mptcp_push_pending(sk, msg->msg_flags); 1947 ret = sk_stream_wait_memory(sk, &timeo); 1948 if (ret) 1949 goto do_error; 1950 } 1951 1952 if (copied) 1953 __mptcp_push_pending(sk, msg->msg_flags); 1954 1955 out: 1956 release_sock(sk); 1957 return copied; 1958 1959 do_error: 1960 if (copied) 1961 goto out; 1962 1963 copied = sk_stream_error(sk, msg->msg_flags, ret); 1964 goto out; 1965 } 1966 1967 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied); 1968 1969 static int __mptcp_recvmsg_mskq(struct sock *sk, struct msghdr *msg, 1970 size_t len, int flags, int copied_total, 1971 struct scm_timestamping_internal *tss, 1972 int *cmsg_flags) 1973 { 1974 struct mptcp_sock *msk = mptcp_sk(sk); 1975 struct sk_buff *skb, *tmp; 1976 int total_data_len = 0; 1977 int copied = 0; 1978 1979 skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) { 1980 u32 delta, offset = MPTCP_SKB_CB(skb)->offset; 1981 u32 data_len = skb->len - offset; 1982 u32 count; 1983 int err; 1984 1985 if (flags & MSG_PEEK) { 1986 /* skip already peeked skbs */ 1987 if (total_data_len + data_len <= copied_total) { 1988 total_data_len += data_len; 1989 continue; 1990 } 1991 1992 /* skip the already peeked data in the current skb */ 1993 delta = copied_total - total_data_len; 1994 offset += delta; 1995 data_len -= delta; 1996 } 1997 1998 count = min_t(size_t, len - copied, data_len); 1999 if (!(flags & MSG_TRUNC)) { 2000 err = skb_copy_datagram_msg(skb, offset, msg, count); 2001 if (unlikely(err < 0)) { 2002 if (!copied) 2003 return err; 2004 break; 2005 } 2006 } 2007 2008 if (MPTCP_SKB_CB(skb)->has_rxtstamp) { 2009 tcp_update_recv_tstamps(skb, tss); 2010 *cmsg_flags |= MPTCP_CMSG_TS; 2011 } 2012 2013 copied += count; 2014 2015 if (!(flags & MSG_PEEK)) { 2016 msk->bytes_consumed += count; 2017 if (count < data_len) { 2018 MPTCP_SKB_CB(skb)->offset += count; 2019 MPTCP_SKB_CB(skb)->map_seq += count; 2020 break; 2021 } 2022 2023 /* avoid the indirect call, we know the destructor is sock_rfree */ 2024 skb->destructor = NULL; 2025 skb->sk = NULL; 2026 atomic_sub(skb->truesize, &sk->sk_rmem_alloc); 2027 sk_mem_uncharge(sk, skb->truesize); 2028 __skb_unlink(skb, &sk->sk_receive_queue); 2029 skb_attempt_defer_free(skb); 2030 } 2031 2032 if (copied >= len) 2033 break; 2034 } 2035 2036 mptcp_rcv_space_adjust(msk, copied); 2037 return copied; 2038 } 2039 2040 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information. 2041 * 2042 * Only difference: Use highest rtt estimate of the subflows in use. 2043 */ 2044 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied) 2045 { 2046 struct mptcp_subflow_context *subflow; 2047 struct sock *sk = (struct sock *)msk; 2048 u8 scaling_ratio = U8_MAX; 2049 u32 time, advmss = 1; 2050 u64 rtt_us, mstamp; 2051 2052 msk_owned_by_me(msk); 2053 2054 if (copied <= 0) 2055 return; 2056 2057 if (!msk->rcvspace_init) 2058 mptcp_rcv_space_init(msk, msk->first); 2059 2060 msk->rcvq_space.copied += copied; 2061 2062 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC); 2063 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time); 2064 2065 rtt_us = msk->rcvq_space.rtt_us; 2066 if (rtt_us && time < (rtt_us >> 3)) 2067 return; 2068 2069 rtt_us = 0; 2070 mptcp_for_each_subflow(msk, subflow) { 2071 const struct tcp_sock *tp; 2072 u64 sf_rtt_us; 2073 u32 sf_advmss; 2074 2075 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow)); 2076 2077 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us); 2078 sf_advmss = READ_ONCE(tp->advmss); 2079 2080 rtt_us = max(sf_rtt_us, rtt_us); 2081 advmss = max(sf_advmss, advmss); 2082 scaling_ratio = min(tp->scaling_ratio, scaling_ratio); 2083 } 2084 2085 msk->rcvq_space.rtt_us = rtt_us; 2086 msk->scaling_ratio = scaling_ratio; 2087 if (time < (rtt_us >> 3) || rtt_us == 0) 2088 return; 2089 2090 if (msk->rcvq_space.copied <= msk->rcvq_space.space) 2091 goto new_measure; 2092 2093 if (mptcp_rcvbuf_grow(sk, msk->rcvq_space.copied)) { 2094 /* Make subflows follow along. If we do not do this, we 2095 * get drops at subflow level if skbs can't be moved to 2096 * the mptcp rx queue fast enough (announced rcv_win can 2097 * exceed ssk->sk_rcvbuf). 2098 */ 2099 mptcp_for_each_subflow(msk, subflow) { 2100 struct sock *ssk; 2101 bool slow; 2102 2103 ssk = mptcp_subflow_tcp_sock(subflow); 2104 slow = lock_sock_fast(ssk); 2105 /* subflows can be added before tcp_init_transfer() */ 2106 if (tcp_sk(ssk)->rcvq_space.space) 2107 tcp_rcvbuf_grow(ssk, msk->rcvq_space.copied); 2108 unlock_sock_fast(ssk, slow); 2109 } 2110 } 2111 2112 new_measure: 2113 msk->rcvq_space.copied = 0; 2114 msk->rcvq_space.time = mstamp; 2115 } 2116 2117 static struct mptcp_subflow_context * 2118 __mptcp_first_ready_from(struct mptcp_sock *msk, 2119 struct mptcp_subflow_context *subflow) 2120 { 2121 struct mptcp_subflow_context *start_subflow = subflow; 2122 2123 while (!READ_ONCE(subflow->data_avail)) { 2124 subflow = mptcp_next_subflow(msk, subflow); 2125 if (subflow == start_subflow) 2126 return NULL; 2127 } 2128 return subflow; 2129 } 2130 2131 static bool __mptcp_move_skbs(struct sock *sk) 2132 { 2133 struct mptcp_subflow_context *subflow; 2134 struct mptcp_sock *msk = mptcp_sk(sk); 2135 bool ret = false; 2136 2137 if (list_empty(&msk->conn_list)) 2138 return false; 2139 2140 subflow = list_first_entry(&msk->conn_list, 2141 struct mptcp_subflow_context, node); 2142 for (;;) { 2143 struct sock *ssk; 2144 bool slowpath; 2145 2146 /* 2147 * As an optimization avoid traversing the subflows list 2148 * and ev. acquiring the subflow socket lock before baling out 2149 */ 2150 if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf) 2151 break; 2152 2153 subflow = __mptcp_first_ready_from(msk, subflow); 2154 if (!subflow) 2155 break; 2156 2157 ssk = mptcp_subflow_tcp_sock(subflow); 2158 slowpath = lock_sock_fast(ssk); 2159 ret = __mptcp_move_skbs_from_subflow(msk, ssk) || ret; 2160 if (unlikely(ssk->sk_err)) 2161 __mptcp_error_report(sk); 2162 unlock_sock_fast(ssk, slowpath); 2163 2164 subflow = mptcp_next_subflow(msk, subflow); 2165 } 2166 2167 __mptcp_ofo_queue(msk); 2168 if (ret) 2169 mptcp_check_data_fin((struct sock *)msk); 2170 return ret; 2171 } 2172 2173 static unsigned int mptcp_inq_hint(const struct sock *sk) 2174 { 2175 const struct mptcp_sock *msk = mptcp_sk(sk); 2176 const struct sk_buff *skb; 2177 2178 skb = skb_peek(&sk->sk_receive_queue); 2179 if (skb) { 2180 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq; 2181 2182 if (hint_val >= INT_MAX) 2183 return INT_MAX; 2184 2185 return (unsigned int)hint_val; 2186 } 2187 2188 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN)) 2189 return 1; 2190 2191 return 0; 2192 } 2193 2194 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, 2195 int flags, int *addr_len) 2196 { 2197 struct mptcp_sock *msk = mptcp_sk(sk); 2198 struct scm_timestamping_internal tss; 2199 int copied = 0, cmsg_flags = 0; 2200 int target; 2201 long timeo; 2202 2203 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */ 2204 if (unlikely(flags & MSG_ERRQUEUE)) 2205 return inet_recv_error(sk, msg, len, addr_len); 2206 2207 lock_sock(sk); 2208 if (unlikely(sk->sk_state == TCP_LISTEN)) { 2209 copied = -ENOTCONN; 2210 goto out_err; 2211 } 2212 2213 mptcp_rps_record_subflows(msk); 2214 2215 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT); 2216 2217 len = min_t(size_t, len, INT_MAX); 2218 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len); 2219 2220 if (unlikely(msk->recvmsg_inq)) 2221 cmsg_flags = MPTCP_CMSG_INQ; 2222 2223 while (copied < len) { 2224 int err, bytes_read; 2225 2226 bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags, 2227 copied, &tss, &cmsg_flags); 2228 if (unlikely(bytes_read < 0)) { 2229 if (!copied) 2230 copied = bytes_read; 2231 goto out_err; 2232 } 2233 2234 copied += bytes_read; 2235 2236 if (skb_queue_empty(&sk->sk_receive_queue) && __mptcp_move_skbs(sk)) 2237 continue; 2238 2239 /* only the MPTCP socket status is relevant here. The exit 2240 * conditions mirror closely tcp_recvmsg() 2241 */ 2242 if (copied >= target) 2243 break; 2244 2245 if (copied) { 2246 if (sk->sk_err || 2247 sk->sk_state == TCP_CLOSE || 2248 (sk->sk_shutdown & RCV_SHUTDOWN) || 2249 !timeo || 2250 signal_pending(current)) 2251 break; 2252 } else { 2253 if (sk->sk_err) { 2254 copied = sock_error(sk); 2255 break; 2256 } 2257 2258 if (sk->sk_shutdown & RCV_SHUTDOWN) 2259 break; 2260 2261 if (sk->sk_state == TCP_CLOSE) { 2262 copied = -ENOTCONN; 2263 break; 2264 } 2265 2266 if (!timeo) { 2267 copied = -EAGAIN; 2268 break; 2269 } 2270 2271 if (signal_pending(current)) { 2272 copied = sock_intr_errno(timeo); 2273 break; 2274 } 2275 } 2276 2277 pr_debug("block timeout %ld\n", timeo); 2278 mptcp_cleanup_rbuf(msk, copied); 2279 err = sk_wait_data(sk, &timeo, NULL); 2280 if (err < 0) { 2281 err = copied ? : err; 2282 goto out_err; 2283 } 2284 } 2285 2286 mptcp_cleanup_rbuf(msk, copied); 2287 2288 out_err: 2289 if (cmsg_flags && copied >= 0) { 2290 if (cmsg_flags & MPTCP_CMSG_TS) 2291 tcp_recv_timestamp(msg, sk, &tss); 2292 2293 if (cmsg_flags & MPTCP_CMSG_INQ) { 2294 unsigned int inq = mptcp_inq_hint(sk); 2295 2296 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq); 2297 } 2298 } 2299 2300 pr_debug("msk=%p rx queue empty=%d copied=%d\n", 2301 msk, skb_queue_empty(&sk->sk_receive_queue), copied); 2302 2303 release_sock(sk); 2304 return copied; 2305 } 2306 2307 static void mptcp_retransmit_timer(struct timer_list *t) 2308 { 2309 struct inet_connection_sock *icsk = timer_container_of(icsk, t, 2310 icsk_retransmit_timer); 2311 struct sock *sk = &icsk->icsk_inet.sk; 2312 struct mptcp_sock *msk = mptcp_sk(sk); 2313 2314 bh_lock_sock(sk); 2315 if (!sock_owned_by_user(sk)) { 2316 /* we need a process context to retransmit */ 2317 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags)) 2318 mptcp_schedule_work(sk); 2319 } else { 2320 /* delegate our work to tcp_release_cb() */ 2321 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags); 2322 } 2323 bh_unlock_sock(sk); 2324 sock_put(sk); 2325 } 2326 2327 static void mptcp_tout_timer(struct timer_list *t) 2328 { 2329 struct sock *sk = timer_container_of(sk, t, sk_timer); 2330 2331 mptcp_schedule_work(sk); 2332 sock_put(sk); 2333 } 2334 2335 /* Find an idle subflow. Return NULL if there is unacked data at tcp 2336 * level. 2337 * 2338 * A backup subflow is returned only if that is the only kind available. 2339 */ 2340 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk) 2341 { 2342 struct sock *backup = NULL, *pick = NULL; 2343 struct mptcp_subflow_context *subflow; 2344 int min_stale_count = INT_MAX; 2345 2346 mptcp_for_each_subflow(msk, subflow) { 2347 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2348 2349 if (!__mptcp_subflow_active(subflow)) 2350 continue; 2351 2352 /* still data outstanding at TCP level? skip this */ 2353 if (!tcp_rtx_and_write_queues_empty(ssk)) { 2354 mptcp_pm_subflow_chk_stale(msk, ssk); 2355 min_stale_count = min_t(int, min_stale_count, subflow->stale_count); 2356 continue; 2357 } 2358 2359 if (subflow->backup || subflow->request_bkup) { 2360 if (!backup) 2361 backup = ssk; 2362 continue; 2363 } 2364 2365 if (!pick) 2366 pick = ssk; 2367 } 2368 2369 if (pick) 2370 return pick; 2371 2372 /* use backup only if there are no progresses anywhere */ 2373 return min_stale_count > 1 ? backup : NULL; 2374 } 2375 2376 bool __mptcp_retransmit_pending_data(struct sock *sk) 2377 { 2378 struct mptcp_data_frag *cur, *rtx_head; 2379 struct mptcp_sock *msk = mptcp_sk(sk); 2380 2381 if (__mptcp_check_fallback(msk)) 2382 return false; 2383 2384 /* the closing socket has some data untransmitted and/or unacked: 2385 * some data in the mptcp rtx queue has not really xmitted yet. 2386 * keep it simple and re-inject the whole mptcp level rtx queue 2387 */ 2388 mptcp_data_lock(sk); 2389 __mptcp_clean_una_wakeup(sk); 2390 rtx_head = mptcp_rtx_head(sk); 2391 if (!rtx_head) { 2392 mptcp_data_unlock(sk); 2393 return false; 2394 } 2395 2396 msk->recovery_snd_nxt = msk->snd_nxt; 2397 msk->recovery = true; 2398 mptcp_data_unlock(sk); 2399 2400 msk->first_pending = rtx_head; 2401 msk->snd_burst = 0; 2402 2403 /* be sure to clear the "sent status" on all re-injected fragments */ 2404 list_for_each_entry(cur, &msk->rtx_queue, list) { 2405 if (!cur->already_sent) 2406 break; 2407 cur->already_sent = 0; 2408 } 2409 2410 return true; 2411 } 2412 2413 /* flags for __mptcp_close_ssk() */ 2414 #define MPTCP_CF_PUSH BIT(1) 2415 2416 /* be sure to send a reset only if the caller asked for it, also 2417 * clean completely the subflow status when the subflow reaches 2418 * TCP_CLOSE state 2419 */ 2420 static void __mptcp_subflow_disconnect(struct sock *ssk, 2421 struct mptcp_subflow_context *subflow, 2422 unsigned int flags) 2423 { 2424 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 2425 subflow->send_fastclose) { 2426 /* The MPTCP code never wait on the subflow sockets, TCP-level 2427 * disconnect should never fail 2428 */ 2429 WARN_ON_ONCE(tcp_disconnect(ssk, 0)); 2430 mptcp_subflow_ctx_reset(subflow); 2431 } else { 2432 tcp_shutdown(ssk, SEND_SHUTDOWN); 2433 } 2434 } 2435 2436 /* subflow sockets can be either outgoing (connect) or incoming 2437 * (accept). 2438 * 2439 * Outgoing subflows use in-kernel sockets. 2440 * Incoming subflows do not have their own 'struct socket' allocated, 2441 * so we need to use tcp_close() after detaching them from the mptcp 2442 * parent socket. 2443 */ 2444 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2445 struct mptcp_subflow_context *subflow, 2446 unsigned int flags) 2447 { 2448 struct mptcp_sock *msk = mptcp_sk(sk); 2449 bool dispose_it, need_push = false; 2450 2451 /* If the first subflow moved to a close state before accept, e.g. due 2452 * to an incoming reset or listener shutdown, the subflow socket is 2453 * already deleted by inet_child_forget() and the mptcp socket can't 2454 * survive too. 2455 */ 2456 if (msk->in_accept_queue && msk->first == ssk && 2457 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) { 2458 /* ensure later check in mptcp_worker() will dispose the msk */ 2459 sock_set_flag(sk, SOCK_DEAD); 2460 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1)); 2461 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2462 mptcp_subflow_drop_ctx(ssk); 2463 goto out_release; 2464 } 2465 2466 dispose_it = msk->free_first || ssk != msk->first; 2467 if (dispose_it) 2468 list_del(&subflow->node); 2469 2470 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 2471 2472 if (subflow->send_fastclose && ssk->sk_state != TCP_CLOSE) 2473 tcp_set_state(ssk, TCP_CLOSE); 2474 2475 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk); 2476 if (!dispose_it) { 2477 __mptcp_subflow_disconnect(ssk, subflow, flags); 2478 release_sock(ssk); 2479 2480 goto out; 2481 } 2482 2483 subflow->disposable = 1; 2484 2485 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops 2486 * the ssk has been already destroyed, we just need to release the 2487 * reference owned by msk; 2488 */ 2489 if (!inet_csk(ssk)->icsk_ulp_ops) { 2490 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD)); 2491 kfree_rcu(subflow, rcu); 2492 } else { 2493 /* otherwise tcp will dispose of the ssk and subflow ctx */ 2494 __tcp_close(ssk, 0); 2495 2496 /* close acquired an extra ref */ 2497 __sock_put(ssk); 2498 } 2499 2500 out_release: 2501 __mptcp_subflow_error_report(sk, ssk); 2502 release_sock(ssk); 2503 2504 sock_put(ssk); 2505 2506 if (ssk == msk->first) 2507 WRITE_ONCE(msk->first, NULL); 2508 2509 out: 2510 __mptcp_sync_sndbuf(sk); 2511 if (need_push) 2512 __mptcp_push_pending(sk, 0); 2513 2514 /* Catch every 'all subflows closed' scenario, including peers silently 2515 * closing them, e.g. due to timeout. 2516 * For established sockets, allow an additional timeout before closing, 2517 * as the protocol can still create more subflows. 2518 */ 2519 if (list_is_singular(&msk->conn_list) && msk->first && 2520 inet_sk_state_load(msk->first) == TCP_CLOSE) { 2521 if (sk->sk_state != TCP_ESTABLISHED || 2522 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) { 2523 mptcp_set_state(sk, TCP_CLOSE); 2524 mptcp_close_wake_up(sk); 2525 } else { 2526 mptcp_start_tout_timer(sk); 2527 } 2528 } 2529 } 2530 2531 void mptcp_close_ssk(struct sock *sk, struct sock *ssk, 2532 struct mptcp_subflow_context *subflow) 2533 { 2534 /* The first subflow can already be closed and still in the list */ 2535 if (subflow->close_event_done) 2536 return; 2537 2538 subflow->close_event_done = true; 2539 2540 if (sk->sk_state == TCP_ESTABLISHED) 2541 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL); 2542 2543 /* subflow aborted before reaching the fully_established status 2544 * attempt the creation of the next subflow 2545 */ 2546 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow); 2547 2548 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH); 2549 } 2550 2551 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu) 2552 { 2553 return 0; 2554 } 2555 2556 static void __mptcp_close_subflow(struct sock *sk) 2557 { 2558 struct mptcp_subflow_context *subflow, *tmp; 2559 struct mptcp_sock *msk = mptcp_sk(sk); 2560 2561 might_sleep(); 2562 2563 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2564 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2565 int ssk_state = inet_sk_state_load(ssk); 2566 2567 if (ssk_state != TCP_CLOSE && 2568 (ssk_state != TCP_CLOSE_WAIT || 2569 inet_sk_state_load(sk) != TCP_ESTABLISHED || 2570 __mptcp_check_fallback(msk))) 2571 continue; 2572 2573 /* 'subflow_data_ready' will re-sched once rx queue is empty */ 2574 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue)) 2575 continue; 2576 2577 mptcp_close_ssk(sk, ssk, subflow); 2578 } 2579 2580 } 2581 2582 static bool mptcp_close_tout_expired(const struct sock *sk) 2583 { 2584 if (!inet_csk(sk)->icsk_mtup.probe_timestamp || 2585 sk->sk_state == TCP_CLOSE) 2586 return false; 2587 2588 return time_after32(tcp_jiffies32, 2589 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk)); 2590 } 2591 2592 static void mptcp_check_fastclose(struct mptcp_sock *msk) 2593 { 2594 struct mptcp_subflow_context *subflow, *tmp; 2595 struct sock *sk = (struct sock *)msk; 2596 2597 if (likely(!READ_ONCE(msk->rcv_fastclose))) 2598 return; 2599 2600 mptcp_token_destroy(msk); 2601 2602 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2603 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 2604 bool slow; 2605 2606 slow = lock_sock_fast(tcp_sk); 2607 if (tcp_sk->sk_state != TCP_CLOSE) { 2608 mptcp_send_active_reset_reason(tcp_sk); 2609 tcp_set_state(tcp_sk, TCP_CLOSE); 2610 } 2611 unlock_sock_fast(tcp_sk, slow); 2612 } 2613 2614 /* Mirror the tcp_reset() error propagation */ 2615 switch (sk->sk_state) { 2616 case TCP_SYN_SENT: 2617 WRITE_ONCE(sk->sk_err, ECONNREFUSED); 2618 break; 2619 case TCP_CLOSE_WAIT: 2620 WRITE_ONCE(sk->sk_err, EPIPE); 2621 break; 2622 case TCP_CLOSE: 2623 return; 2624 default: 2625 WRITE_ONCE(sk->sk_err, ECONNRESET); 2626 } 2627 2628 mptcp_set_state(sk, TCP_CLOSE); 2629 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 2630 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */ 2631 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags); 2632 2633 /* the calling mptcp_worker will properly destroy the socket */ 2634 if (sock_flag(sk, SOCK_DEAD)) 2635 return; 2636 2637 sk->sk_state_change(sk); 2638 sk_error_report(sk); 2639 } 2640 2641 static void __mptcp_retrans(struct sock *sk) 2642 { 2643 struct mptcp_sendmsg_info info = { .data_lock_held = true, }; 2644 struct mptcp_sock *msk = mptcp_sk(sk); 2645 struct mptcp_subflow_context *subflow; 2646 struct mptcp_data_frag *dfrag; 2647 struct sock *ssk; 2648 int ret, err; 2649 u16 len = 0; 2650 2651 mptcp_clean_una_wakeup(sk); 2652 2653 /* first check ssk: need to kick "stale" logic */ 2654 err = mptcp_sched_get_retrans(msk); 2655 dfrag = mptcp_rtx_head(sk); 2656 if (!dfrag) { 2657 if (mptcp_data_fin_enabled(msk)) { 2658 struct inet_connection_sock *icsk = inet_csk(sk); 2659 2660 WRITE_ONCE(icsk->icsk_retransmits, 2661 icsk->icsk_retransmits + 1); 2662 mptcp_set_datafin_timeout(sk); 2663 mptcp_send_ack(msk); 2664 2665 goto reset_timer; 2666 } 2667 2668 if (!mptcp_send_head(sk)) 2669 return; 2670 2671 goto reset_timer; 2672 } 2673 2674 if (err) 2675 goto reset_timer; 2676 2677 mptcp_for_each_subflow(msk, subflow) { 2678 if (READ_ONCE(subflow->scheduled)) { 2679 u16 copied = 0; 2680 2681 mptcp_subflow_set_scheduled(subflow, false); 2682 2683 ssk = mptcp_subflow_tcp_sock(subflow); 2684 2685 lock_sock(ssk); 2686 2687 /* limit retransmission to the bytes already sent on some subflows */ 2688 info.sent = 0; 2689 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : 2690 dfrag->already_sent; 2691 2692 /* 2693 * make the whole retrans decision, xmit, disallow 2694 * fallback atomic 2695 */ 2696 spin_lock_bh(&msk->fallback_lock); 2697 if (__mptcp_check_fallback(msk)) { 2698 spin_unlock_bh(&msk->fallback_lock); 2699 release_sock(ssk); 2700 return; 2701 } 2702 2703 while (info.sent < info.limit) { 2704 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info); 2705 if (ret <= 0) 2706 break; 2707 2708 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS); 2709 copied += ret; 2710 info.sent += ret; 2711 } 2712 if (copied) { 2713 len = max(copied, len); 2714 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle, 2715 info.size_goal); 2716 msk->allow_infinite_fallback = false; 2717 } 2718 spin_unlock_bh(&msk->fallback_lock); 2719 2720 release_sock(ssk); 2721 } 2722 } 2723 2724 msk->bytes_retrans += len; 2725 dfrag->already_sent = max(dfrag->already_sent, len); 2726 2727 reset_timer: 2728 mptcp_check_and_set_pending(sk); 2729 2730 if (!mptcp_rtx_timer_pending(sk)) 2731 mptcp_reset_rtx_timer(sk); 2732 } 2733 2734 /* schedule the timeout timer for the relevant event: either close timeout 2735 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one 2736 */ 2737 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout) 2738 { 2739 struct sock *sk = (struct sock *)msk; 2740 unsigned long timeout, close_timeout; 2741 2742 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp) 2743 return; 2744 2745 close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp - 2746 tcp_jiffies32 + jiffies + mptcp_close_timeout(sk); 2747 2748 /* the close timeout takes precedence on the fail one, and here at least one of 2749 * them is active 2750 */ 2751 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout; 2752 2753 sk_reset_timer(sk, &sk->sk_timer, timeout); 2754 } 2755 2756 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk) 2757 { 2758 struct sock *ssk = msk->first; 2759 bool slow; 2760 2761 if (!ssk) 2762 return; 2763 2764 pr_debug("MP_FAIL doesn't respond, reset the subflow\n"); 2765 2766 slow = lock_sock_fast(ssk); 2767 mptcp_subflow_reset(ssk); 2768 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0); 2769 unlock_sock_fast(ssk, slow); 2770 } 2771 2772 static void mptcp_do_fastclose(struct sock *sk) 2773 { 2774 struct mptcp_subflow_context *subflow, *tmp; 2775 struct mptcp_sock *msk = mptcp_sk(sk); 2776 2777 mptcp_set_state(sk, TCP_CLOSE); 2778 2779 /* Explicitly send the fastclose reset as need */ 2780 if (__mptcp_check_fallback(msk)) 2781 return; 2782 2783 mptcp_for_each_subflow_safe(msk, subflow, tmp) { 2784 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 2785 2786 lock_sock(ssk); 2787 2788 /* Some subflow socket states don't allow/need a reset.*/ 2789 if ((1 << ssk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) 2790 goto unlock; 2791 2792 subflow->send_fastclose = 1; 2793 tcp_send_active_reset(ssk, ssk->sk_allocation, 2794 SK_RST_REASON_TCP_ABORT_ON_CLOSE); 2795 unlock: 2796 release_sock(ssk); 2797 } 2798 } 2799 2800 static void mptcp_worker(struct work_struct *work) 2801 { 2802 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work); 2803 struct sock *sk = (struct sock *)msk; 2804 unsigned long fail_tout; 2805 int state; 2806 2807 lock_sock(sk); 2808 state = sk->sk_state; 2809 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN))) 2810 goto unlock; 2811 2812 mptcp_check_fastclose(msk); 2813 2814 mptcp_pm_worker(msk); 2815 2816 mptcp_check_send_data_fin(sk); 2817 mptcp_check_data_fin_ack(sk); 2818 mptcp_check_data_fin(sk); 2819 2820 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags)) 2821 __mptcp_close_subflow(sk); 2822 2823 if (mptcp_close_tout_expired(sk)) { 2824 struct mptcp_subflow_context *subflow, *tmp; 2825 2826 mptcp_do_fastclose(sk); 2827 mptcp_for_each_subflow_safe(msk, subflow, tmp) 2828 __mptcp_close_ssk(sk, subflow->tcp_sock, subflow, 0); 2829 mptcp_close_wake_up(sk); 2830 } 2831 2832 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) { 2833 __mptcp_destroy_sock(sk); 2834 goto unlock; 2835 } 2836 2837 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags)) 2838 __mptcp_retrans(sk); 2839 2840 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0; 2841 if (fail_tout && time_after(jiffies, fail_tout)) 2842 mptcp_mp_fail_no_response(msk); 2843 2844 unlock: 2845 release_sock(sk); 2846 sock_put(sk); 2847 } 2848 2849 static void __mptcp_init_sock(struct sock *sk) 2850 { 2851 struct mptcp_sock *msk = mptcp_sk(sk); 2852 2853 INIT_LIST_HEAD(&msk->conn_list); 2854 INIT_LIST_HEAD(&msk->join_list); 2855 INIT_LIST_HEAD(&msk->rtx_queue); 2856 INIT_WORK(&msk->work, mptcp_worker); 2857 msk->out_of_order_queue = RB_ROOT; 2858 msk->first_pending = NULL; 2859 msk->timer_ival = TCP_RTO_MIN; 2860 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO; 2861 2862 WRITE_ONCE(msk->first, NULL); 2863 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss; 2864 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 2865 msk->allow_infinite_fallback = true; 2866 msk->allow_subflows = true; 2867 msk->recovery = false; 2868 msk->subflow_id = 1; 2869 msk->last_data_sent = tcp_jiffies32; 2870 msk->last_data_recv = tcp_jiffies32; 2871 msk->last_ack_recv = tcp_jiffies32; 2872 2873 mptcp_pm_data_init(msk); 2874 spin_lock_init(&msk->fallback_lock); 2875 2876 /* re-use the csk retrans timer for MPTCP-level retrans */ 2877 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0); 2878 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0); 2879 } 2880 2881 static void mptcp_ca_reset(struct sock *sk) 2882 { 2883 struct inet_connection_sock *icsk = inet_csk(sk); 2884 2885 tcp_assign_congestion_control(sk); 2886 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name, 2887 sizeof(mptcp_sk(sk)->ca_name)); 2888 2889 /* no need to keep a reference to the ops, the name will suffice */ 2890 tcp_cleanup_congestion_control(sk); 2891 icsk->icsk_ca_ops = NULL; 2892 } 2893 2894 static int mptcp_init_sock(struct sock *sk) 2895 { 2896 struct net *net = sock_net(sk); 2897 int ret; 2898 2899 __mptcp_init_sock(sk); 2900 2901 if (!mptcp_is_enabled(net)) 2902 return -ENOPROTOOPT; 2903 2904 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net)) 2905 return -ENOMEM; 2906 2907 rcu_read_lock(); 2908 ret = mptcp_init_sched(mptcp_sk(sk), 2909 mptcp_sched_find(mptcp_get_scheduler(net))); 2910 rcu_read_unlock(); 2911 if (ret) 2912 return ret; 2913 2914 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags); 2915 2916 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will 2917 * propagate the correct value 2918 */ 2919 mptcp_ca_reset(sk); 2920 2921 sk_sockets_allocated_inc(sk); 2922 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]); 2923 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]); 2924 2925 return 0; 2926 } 2927 2928 static void __mptcp_clear_xmit(struct sock *sk) 2929 { 2930 struct mptcp_sock *msk = mptcp_sk(sk); 2931 struct mptcp_data_frag *dtmp, *dfrag; 2932 2933 msk->first_pending = NULL; 2934 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) 2935 dfrag_clear(sk, dfrag); 2936 } 2937 2938 void mptcp_cancel_work(struct sock *sk) 2939 { 2940 struct mptcp_sock *msk = mptcp_sk(sk); 2941 2942 if (cancel_work_sync(&msk->work)) 2943 __sock_put(sk); 2944 } 2945 2946 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how) 2947 { 2948 lock_sock(ssk); 2949 2950 switch (ssk->sk_state) { 2951 case TCP_LISTEN: 2952 if (!(how & RCV_SHUTDOWN)) 2953 break; 2954 fallthrough; 2955 case TCP_SYN_SENT: 2956 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK)); 2957 break; 2958 default: 2959 if (__mptcp_check_fallback(mptcp_sk(sk))) { 2960 pr_debug("Fallback\n"); 2961 ssk->sk_shutdown |= how; 2962 tcp_shutdown(ssk, how); 2963 2964 /* simulate the data_fin ack reception to let the state 2965 * machine move forward 2966 */ 2967 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt); 2968 mptcp_schedule_work(sk); 2969 } else { 2970 pr_debug("Sending DATA_FIN on subflow %p\n", ssk); 2971 tcp_send_ack(ssk); 2972 if (!mptcp_rtx_timer_pending(sk)) 2973 mptcp_reset_rtx_timer(sk); 2974 } 2975 break; 2976 } 2977 2978 release_sock(ssk); 2979 } 2980 2981 void mptcp_set_state(struct sock *sk, int state) 2982 { 2983 int oldstate = sk->sk_state; 2984 2985 switch (state) { 2986 case TCP_ESTABLISHED: 2987 if (oldstate != TCP_ESTABLISHED) 2988 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2989 break; 2990 case TCP_CLOSE_WAIT: 2991 /* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state: 2992 * MPTCP "accepted" sockets will be created later on. So no 2993 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT. 2994 */ 2995 break; 2996 default: 2997 if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT) 2998 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB); 2999 } 3000 3001 inet_sk_state_store(sk, state); 3002 } 3003 3004 static const unsigned char new_state[16] = { 3005 /* current state: new state: action: */ 3006 [0 /* (Invalid) */] = TCP_CLOSE, 3007 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3008 [TCP_SYN_SENT] = TCP_CLOSE, 3009 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN, 3010 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1, 3011 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2, 3012 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */ 3013 [TCP_CLOSE] = TCP_CLOSE, 3014 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN, 3015 [TCP_LAST_ACK] = TCP_LAST_ACK, 3016 [TCP_LISTEN] = TCP_CLOSE, 3017 [TCP_CLOSING] = TCP_CLOSING, 3018 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */ 3019 }; 3020 3021 static int mptcp_close_state(struct sock *sk) 3022 { 3023 int next = (int)new_state[sk->sk_state]; 3024 int ns = next & TCP_STATE_MASK; 3025 3026 mptcp_set_state(sk, ns); 3027 3028 return next & TCP_ACTION_FIN; 3029 } 3030 3031 static void mptcp_check_send_data_fin(struct sock *sk) 3032 { 3033 struct mptcp_subflow_context *subflow; 3034 struct mptcp_sock *msk = mptcp_sk(sk); 3035 3036 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n", 3037 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk), 3038 msk->snd_nxt, msk->write_seq); 3039 3040 /* we still need to enqueue subflows or not really shutting down, 3041 * skip this 3042 */ 3043 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq || 3044 mptcp_send_head(sk)) 3045 return; 3046 3047 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3048 3049 mptcp_for_each_subflow(msk, subflow) { 3050 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow); 3051 3052 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN); 3053 } 3054 } 3055 3056 static void __mptcp_wr_shutdown(struct sock *sk) 3057 { 3058 struct mptcp_sock *msk = mptcp_sk(sk); 3059 3060 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n", 3061 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state, 3062 !!mptcp_send_head(sk)); 3063 3064 /* will be ignored by fallback sockets */ 3065 WRITE_ONCE(msk->write_seq, msk->write_seq + 1); 3066 WRITE_ONCE(msk->snd_data_fin_enable, 1); 3067 3068 mptcp_check_send_data_fin(sk); 3069 } 3070 3071 static void __mptcp_destroy_sock(struct sock *sk) 3072 { 3073 struct mptcp_sock *msk = mptcp_sk(sk); 3074 3075 pr_debug("msk=%p\n", msk); 3076 3077 might_sleep(); 3078 3079 mptcp_stop_rtx_timer(sk); 3080 sk_stop_timer(sk, &sk->sk_timer); 3081 msk->pm.status = 0; 3082 mptcp_release_sched(msk); 3083 3084 sk->sk_prot->destroy(sk); 3085 3086 sk_stream_kill_queues(sk); 3087 xfrm_sk_free_policy(sk); 3088 3089 sock_put(sk); 3090 } 3091 3092 void __mptcp_unaccepted_force_close(struct sock *sk) 3093 { 3094 sock_set_flag(sk, SOCK_DEAD); 3095 mptcp_do_fastclose(sk); 3096 __mptcp_destroy_sock(sk); 3097 } 3098 3099 static __poll_t mptcp_check_readable(struct sock *sk) 3100 { 3101 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0; 3102 } 3103 3104 static void mptcp_check_listen_stop(struct sock *sk) 3105 { 3106 struct sock *ssk; 3107 3108 if (inet_sk_state_load(sk) != TCP_LISTEN) 3109 return; 3110 3111 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1); 3112 ssk = mptcp_sk(sk)->first; 3113 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN)) 3114 return; 3115 3116 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING); 3117 tcp_set_state(ssk, TCP_CLOSE); 3118 mptcp_subflow_queue_clean(sk, ssk); 3119 inet_csk_listen_stop(ssk); 3120 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED); 3121 release_sock(ssk); 3122 } 3123 3124 bool __mptcp_close(struct sock *sk, long timeout) 3125 { 3126 struct mptcp_subflow_context *subflow; 3127 struct mptcp_sock *msk = mptcp_sk(sk); 3128 bool do_cancel_work = false; 3129 int subflows_alive = 0; 3130 3131 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK); 3132 3133 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) { 3134 mptcp_check_listen_stop(sk); 3135 mptcp_set_state(sk, TCP_CLOSE); 3136 goto cleanup; 3137 } 3138 3139 if (mptcp_data_avail(msk) || timeout < 0) { 3140 /* If the msk has read data, or the caller explicitly ask it, 3141 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose 3142 */ 3143 mptcp_do_fastclose(sk); 3144 timeout = 0; 3145 } else if (mptcp_close_state(sk)) { 3146 __mptcp_wr_shutdown(sk); 3147 } 3148 3149 sk_stream_wait_close(sk, timeout); 3150 3151 cleanup: 3152 /* orphan all the subflows */ 3153 mptcp_for_each_subflow(msk, subflow) { 3154 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 3155 bool slow = lock_sock_fast_nested(ssk); 3156 3157 subflows_alive += ssk->sk_state != TCP_CLOSE; 3158 3159 /* since the close timeout takes precedence on the fail one, 3160 * cancel the latter 3161 */ 3162 if (ssk == msk->first) 3163 subflow->fail_tout = 0; 3164 3165 /* detach from the parent socket, but allow data_ready to 3166 * push incoming data into the mptcp stack, to properly ack it 3167 */ 3168 ssk->sk_socket = NULL; 3169 ssk->sk_wq = NULL; 3170 unlock_sock_fast(ssk, slow); 3171 } 3172 sock_orphan(sk); 3173 3174 /* all the subflows are closed, only timeout can change the msk 3175 * state, let's not keep resources busy for no reasons 3176 */ 3177 if (subflows_alive == 0) 3178 mptcp_set_state(sk, TCP_CLOSE); 3179 3180 sock_hold(sk); 3181 pr_debug("msk=%p state=%d\n", sk, sk->sk_state); 3182 mptcp_pm_connection_closed(msk); 3183 3184 if (sk->sk_state == TCP_CLOSE) { 3185 __mptcp_destroy_sock(sk); 3186 do_cancel_work = true; 3187 } else { 3188 mptcp_start_tout_timer(sk); 3189 } 3190 3191 return do_cancel_work; 3192 } 3193 3194 static void mptcp_close(struct sock *sk, long timeout) 3195 { 3196 bool do_cancel_work; 3197 3198 lock_sock(sk); 3199 3200 do_cancel_work = __mptcp_close(sk, timeout); 3201 release_sock(sk); 3202 if (do_cancel_work) 3203 mptcp_cancel_work(sk); 3204 3205 sock_put(sk); 3206 } 3207 3208 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk) 3209 { 3210 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3211 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk); 3212 struct ipv6_pinfo *msk6 = inet6_sk(msk); 3213 3214 msk->sk_v6_daddr = ssk->sk_v6_daddr; 3215 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr; 3216 3217 if (msk6 && ssk6) { 3218 msk6->saddr = ssk6->saddr; 3219 msk6->flow_label = ssk6->flow_label; 3220 } 3221 #endif 3222 3223 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num; 3224 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport; 3225 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport; 3226 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr; 3227 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr; 3228 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr; 3229 } 3230 3231 static int mptcp_disconnect(struct sock *sk, int flags) 3232 { 3233 struct mptcp_sock *msk = mptcp_sk(sk); 3234 3235 /* We are on the fastopen error path. We can't call straight into the 3236 * subflows cleanup code due to lock nesting (we are already under 3237 * msk->firstsocket lock). 3238 */ 3239 if (msk->fastopening) 3240 return -EBUSY; 3241 3242 mptcp_check_listen_stop(sk); 3243 mptcp_set_state(sk, TCP_CLOSE); 3244 3245 mptcp_stop_rtx_timer(sk); 3246 mptcp_stop_tout_timer(sk); 3247 3248 mptcp_pm_connection_closed(msk); 3249 3250 /* msk->subflow is still intact, the following will not free the first 3251 * subflow 3252 */ 3253 mptcp_do_fastclose(sk); 3254 mptcp_destroy_common(msk); 3255 3256 /* The first subflow is already in TCP_CLOSE status, the following 3257 * can't overlap with a fallback anymore 3258 */ 3259 spin_lock_bh(&msk->fallback_lock); 3260 msk->allow_subflows = true; 3261 msk->allow_infinite_fallback = true; 3262 WRITE_ONCE(msk->flags, 0); 3263 spin_unlock_bh(&msk->fallback_lock); 3264 3265 msk->cb_flags = 0; 3266 msk->recovery = false; 3267 WRITE_ONCE(msk->can_ack, false); 3268 WRITE_ONCE(msk->fully_established, false); 3269 WRITE_ONCE(msk->rcv_data_fin, false); 3270 WRITE_ONCE(msk->snd_data_fin_enable, false); 3271 WRITE_ONCE(msk->rcv_fastclose, false); 3272 WRITE_ONCE(msk->use_64bit_ack, false); 3273 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk))); 3274 mptcp_pm_data_reset(msk); 3275 mptcp_ca_reset(sk); 3276 msk->bytes_consumed = 0; 3277 msk->bytes_acked = 0; 3278 msk->bytes_received = 0; 3279 msk->bytes_sent = 0; 3280 msk->bytes_retrans = 0; 3281 msk->rcvspace_init = 0; 3282 3283 WRITE_ONCE(sk->sk_shutdown, 0); 3284 sk_error_report(sk); 3285 return 0; 3286 } 3287 3288 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3289 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk) 3290 { 3291 struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk); 3292 3293 return &msk6->np; 3294 } 3295 3296 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk) 3297 { 3298 const struct ipv6_pinfo *np = inet6_sk(sk); 3299 struct ipv6_txoptions *opt; 3300 struct ipv6_pinfo *newnp; 3301 3302 newnp = inet6_sk(newsk); 3303 3304 rcu_read_lock(); 3305 opt = rcu_dereference(np->opt); 3306 if (opt) { 3307 opt = ipv6_dup_options(newsk, opt); 3308 if (!opt) 3309 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__); 3310 } 3311 RCU_INIT_POINTER(newnp->opt, opt); 3312 rcu_read_unlock(); 3313 } 3314 #endif 3315 3316 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk) 3317 { 3318 struct ip_options_rcu *inet_opt, *newopt = NULL; 3319 const struct inet_sock *inet = inet_sk(sk); 3320 struct inet_sock *newinet; 3321 3322 newinet = inet_sk(newsk); 3323 3324 rcu_read_lock(); 3325 inet_opt = rcu_dereference(inet->inet_opt); 3326 if (inet_opt) { 3327 newopt = sock_kmemdup(newsk, inet_opt, sizeof(*inet_opt) + 3328 inet_opt->opt.optlen, GFP_ATOMIC); 3329 if (!newopt) 3330 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__); 3331 } 3332 RCU_INIT_POINTER(newinet->inet_opt, newopt); 3333 rcu_read_unlock(); 3334 } 3335 3336 struct sock *mptcp_sk_clone_init(const struct sock *sk, 3337 const struct mptcp_options_received *mp_opt, 3338 struct sock *ssk, 3339 struct request_sock *req) 3340 { 3341 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 3342 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC); 3343 struct mptcp_subflow_context *subflow; 3344 struct mptcp_sock *msk; 3345 3346 if (!nsk) 3347 return NULL; 3348 3349 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3350 if (nsk->sk_family == AF_INET6) 3351 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk); 3352 #endif 3353 3354 __mptcp_init_sock(nsk); 3355 3356 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3357 if (nsk->sk_family == AF_INET6) 3358 mptcp_copy_ip6_options(nsk, sk); 3359 else 3360 #endif 3361 mptcp_copy_ip_options(nsk, sk); 3362 3363 msk = mptcp_sk(nsk); 3364 WRITE_ONCE(msk->local_key, subflow_req->local_key); 3365 WRITE_ONCE(msk->token, subflow_req->token); 3366 msk->in_accept_queue = 1; 3367 WRITE_ONCE(msk->fully_established, false); 3368 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD) 3369 WRITE_ONCE(msk->csum_enabled, true); 3370 3371 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1); 3372 WRITE_ONCE(msk->snd_nxt, msk->write_seq); 3373 WRITE_ONCE(msk->snd_una, msk->write_seq); 3374 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd); 3375 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq; 3376 mptcp_init_sched(msk, mptcp_sk(sk)->sched); 3377 3378 /* passive msk is created after the first/MPC subflow */ 3379 msk->subflow_id = 2; 3380 3381 sock_reset_flag(nsk, SOCK_RCU_FREE); 3382 security_inet_csk_clone(nsk, req); 3383 3384 /* this can't race with mptcp_close(), as the msk is 3385 * not yet exposted to user-space 3386 */ 3387 mptcp_set_state(nsk, TCP_ESTABLISHED); 3388 3389 /* The msk maintain a ref to each subflow in the connections list */ 3390 WRITE_ONCE(msk->first, ssk); 3391 subflow = mptcp_subflow_ctx(ssk); 3392 list_add(&subflow->node, &msk->conn_list); 3393 sock_hold(ssk); 3394 3395 /* new mpc subflow takes ownership of the newly 3396 * created mptcp socket 3397 */ 3398 mptcp_token_accept(subflow_req, msk); 3399 3400 /* set msk addresses early to ensure mptcp_pm_get_local_id() 3401 * uses the correct data 3402 */ 3403 mptcp_copy_inaddrs(nsk, ssk); 3404 __mptcp_propagate_sndbuf(nsk, ssk); 3405 3406 mptcp_rcv_space_init(msk, ssk); 3407 3408 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 3409 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 3410 bh_unlock_sock(nsk); 3411 3412 /* note: the newly allocated socket refcount is 2 now */ 3413 return nsk; 3414 } 3415 3416 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk) 3417 { 3418 const struct tcp_sock *tp = tcp_sk(ssk); 3419 3420 msk->rcvspace_init = 1; 3421 msk->rcvq_space.copied = 0; 3422 msk->rcvq_space.rtt_us = 0; 3423 3424 msk->rcvq_space.time = tp->tcp_mstamp; 3425 3426 /* initial rcv_space offering made to peer */ 3427 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd, 3428 TCP_INIT_CWND * tp->advmss); 3429 if (msk->rcvq_space.space == 0) 3430 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT; 3431 } 3432 3433 void mptcp_destroy_common(struct mptcp_sock *msk) 3434 { 3435 struct mptcp_subflow_context *subflow, *tmp; 3436 struct sock *sk = (struct sock *)msk; 3437 3438 __mptcp_clear_xmit(sk); 3439 3440 /* join list will be eventually flushed (with rst) at sock lock release time */ 3441 mptcp_for_each_subflow_safe(msk, subflow, tmp) 3442 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, 0); 3443 3444 __skb_queue_purge(&sk->sk_receive_queue); 3445 skb_rbtree_purge(&msk->out_of_order_queue); 3446 3447 /* move all the rx fwd alloc into the sk_mem_reclaim_final in 3448 * inet_sock_destruct() will dispose it 3449 */ 3450 mptcp_token_destroy(msk); 3451 mptcp_pm_destroy(msk); 3452 } 3453 3454 static void mptcp_destroy(struct sock *sk) 3455 { 3456 struct mptcp_sock *msk = mptcp_sk(sk); 3457 3458 /* allow the following to close even the initial subflow */ 3459 msk->free_first = 1; 3460 mptcp_destroy_common(msk); 3461 sk_sockets_allocated_dec(sk); 3462 } 3463 3464 void __mptcp_data_acked(struct sock *sk) 3465 { 3466 if (!sock_owned_by_user(sk)) 3467 __mptcp_clean_una(sk); 3468 else 3469 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags); 3470 } 3471 3472 void __mptcp_check_push(struct sock *sk, struct sock *ssk) 3473 { 3474 if (!sock_owned_by_user(sk)) 3475 __mptcp_subflow_push_pending(sk, ssk, false); 3476 else 3477 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3478 } 3479 3480 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \ 3481 BIT(MPTCP_RETRANSMIT) | \ 3482 BIT(MPTCP_FLUSH_JOIN_LIST) | \ 3483 BIT(MPTCP_DEQUEUE)) 3484 3485 /* processes deferred events and flush wmem */ 3486 static void mptcp_release_cb(struct sock *sk) 3487 __must_hold(&sk->sk_lock.slock) 3488 { 3489 struct mptcp_sock *msk = mptcp_sk(sk); 3490 3491 for (;;) { 3492 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED); 3493 struct list_head join_list; 3494 3495 if (!flags) 3496 break; 3497 3498 INIT_LIST_HEAD(&join_list); 3499 list_splice_init(&msk->join_list, &join_list); 3500 3501 /* the following actions acquire the subflow socket lock 3502 * 3503 * 1) can't be invoked in atomic scope 3504 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX 3505 * datapath acquires the msk socket spinlock while helding 3506 * the subflow socket lock 3507 */ 3508 msk->cb_flags &= ~flags; 3509 spin_unlock_bh(&sk->sk_lock.slock); 3510 3511 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST)) 3512 __mptcp_flush_join_list(sk, &join_list); 3513 if (flags & BIT(MPTCP_PUSH_PENDING)) 3514 __mptcp_push_pending(sk, 0); 3515 if (flags & BIT(MPTCP_RETRANSMIT)) 3516 __mptcp_retrans(sk); 3517 if ((flags & BIT(MPTCP_DEQUEUE)) && __mptcp_move_skbs(sk)) { 3518 /* notify ack seq update */ 3519 mptcp_cleanup_rbuf(msk, 0); 3520 sk->sk_data_ready(sk); 3521 } 3522 3523 cond_resched(); 3524 spin_lock_bh(&sk->sk_lock.slock); 3525 } 3526 3527 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags)) 3528 __mptcp_clean_una_wakeup(sk); 3529 if (unlikely(msk->cb_flags)) { 3530 /* be sure to sync the msk state before taking actions 3531 * depending on sk_state (MPTCP_ERROR_REPORT) 3532 * On sk release avoid actions depending on the first subflow 3533 */ 3534 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first) 3535 __mptcp_sync_state(sk, msk->pending_state); 3536 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags)) 3537 __mptcp_error_report(sk); 3538 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags)) 3539 __mptcp_sync_sndbuf(sk); 3540 } 3541 } 3542 3543 /* MP_JOIN client subflow must wait for 4th ack before sending any data: 3544 * TCP can't schedule delack timer before the subflow is fully established. 3545 * MPTCP uses the delack timer to do 3rd ack retransmissions 3546 */ 3547 static void schedule_3rdack_retransmission(struct sock *ssk) 3548 { 3549 struct inet_connection_sock *icsk = inet_csk(ssk); 3550 struct tcp_sock *tp = tcp_sk(ssk); 3551 unsigned long timeout; 3552 3553 if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established)) 3554 return; 3555 3556 /* reschedule with a timeout above RTT, as we must look only for drop */ 3557 if (tp->srtt_us) 3558 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1)); 3559 else 3560 timeout = TCP_TIMEOUT_INIT; 3561 timeout += jiffies; 3562 3563 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER); 3564 smp_store_release(&icsk->icsk_ack.pending, 3565 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER); 3566 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout); 3567 } 3568 3569 void mptcp_subflow_process_delegated(struct sock *ssk, long status) 3570 { 3571 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3572 struct sock *sk = subflow->conn; 3573 3574 if (status & BIT(MPTCP_DELEGATE_SEND)) { 3575 mptcp_data_lock(sk); 3576 if (!sock_owned_by_user(sk)) 3577 __mptcp_subflow_push_pending(sk, ssk, true); 3578 else 3579 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags); 3580 mptcp_data_unlock(sk); 3581 } 3582 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) { 3583 mptcp_data_lock(sk); 3584 if (!sock_owned_by_user(sk)) 3585 __mptcp_sync_sndbuf(sk); 3586 else 3587 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags); 3588 mptcp_data_unlock(sk); 3589 } 3590 if (status & BIT(MPTCP_DELEGATE_ACK)) 3591 schedule_3rdack_retransmission(ssk); 3592 } 3593 3594 static int mptcp_hash(struct sock *sk) 3595 { 3596 /* should never be called, 3597 * we hash the TCP subflows not the MPTCP socket 3598 */ 3599 WARN_ON_ONCE(1); 3600 return 0; 3601 } 3602 3603 static void mptcp_unhash(struct sock *sk) 3604 { 3605 /* called from sk_common_release(), but nothing to do here */ 3606 } 3607 3608 static int mptcp_get_port(struct sock *sk, unsigned short snum) 3609 { 3610 struct mptcp_sock *msk = mptcp_sk(sk); 3611 3612 pr_debug("msk=%p, ssk=%p\n", msk, msk->first); 3613 if (WARN_ON_ONCE(!msk->first)) 3614 return -EINVAL; 3615 3616 return inet_csk_get_port(msk->first, snum); 3617 } 3618 3619 void mptcp_finish_connect(struct sock *ssk) 3620 { 3621 struct mptcp_subflow_context *subflow; 3622 struct mptcp_sock *msk; 3623 struct sock *sk; 3624 3625 subflow = mptcp_subflow_ctx(ssk); 3626 sk = subflow->conn; 3627 msk = mptcp_sk(sk); 3628 3629 pr_debug("msk=%p, token=%u\n", sk, subflow->token); 3630 3631 subflow->map_seq = subflow->iasn; 3632 subflow->map_subflow_seq = 1; 3633 3634 /* the socket is not connected yet, no msk/subflow ops can access/race 3635 * accessing the field below 3636 */ 3637 WRITE_ONCE(msk->local_key, subflow->local_key); 3638 3639 mptcp_pm_new_connection(msk, ssk, 0); 3640 } 3641 3642 void mptcp_sock_graft(struct sock *sk, struct socket *parent) 3643 { 3644 write_lock_bh(&sk->sk_callback_lock); 3645 rcu_assign_pointer(sk->sk_wq, &parent->wq); 3646 sk_set_socket(sk, parent); 3647 write_unlock_bh(&sk->sk_callback_lock); 3648 } 3649 3650 bool mptcp_finish_join(struct sock *ssk) 3651 { 3652 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 3653 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 3654 struct sock *parent = (void *)msk; 3655 bool ret = true; 3656 3657 pr_debug("msk=%p, subflow=%p\n", msk, subflow); 3658 3659 /* mptcp socket already closing? */ 3660 if (!mptcp_is_fully_established(parent)) { 3661 subflow->reset_reason = MPTCP_RST_EMPTCP; 3662 return false; 3663 } 3664 3665 /* active subflow, already present inside the conn_list */ 3666 if (!list_empty(&subflow->node)) { 3667 spin_lock_bh(&msk->fallback_lock); 3668 if (!msk->allow_subflows) { 3669 spin_unlock_bh(&msk->fallback_lock); 3670 return false; 3671 } 3672 mptcp_subflow_joined(msk, ssk); 3673 spin_unlock_bh(&msk->fallback_lock); 3674 mptcp_propagate_sndbuf(parent, ssk); 3675 return true; 3676 } 3677 3678 if (!mptcp_pm_allow_new_subflow(msk)) { 3679 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_JOINREJECTED); 3680 goto err_prohibited; 3681 } 3682 3683 /* If we can't acquire msk socket lock here, let the release callback 3684 * handle it 3685 */ 3686 mptcp_data_lock(parent); 3687 if (!sock_owned_by_user(parent)) { 3688 ret = __mptcp_finish_join(msk, ssk); 3689 if (ret) { 3690 sock_hold(ssk); 3691 list_add_tail(&subflow->node, &msk->conn_list); 3692 } 3693 } else { 3694 sock_hold(ssk); 3695 list_add_tail(&subflow->node, &msk->join_list); 3696 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags); 3697 } 3698 mptcp_data_unlock(parent); 3699 3700 if (!ret) { 3701 err_prohibited: 3702 subflow->reset_reason = MPTCP_RST_EPROHIBIT; 3703 return false; 3704 } 3705 3706 return true; 3707 } 3708 3709 static void mptcp_shutdown(struct sock *sk, int how) 3710 { 3711 pr_debug("sk=%p, how=%d\n", sk, how); 3712 3713 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk)) 3714 __mptcp_wr_shutdown(sk); 3715 } 3716 3717 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v) 3718 { 3719 const struct sock *sk = (void *)msk; 3720 u64 delta; 3721 3722 if (sk->sk_state == TCP_LISTEN) 3723 return -EINVAL; 3724 3725 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 3726 return 0; 3727 3728 delta = msk->write_seq - v; 3729 if (__mptcp_check_fallback(msk) && msk->first) { 3730 struct tcp_sock *tp = tcp_sk(msk->first); 3731 3732 /* the first subflow is disconnected after close - see 3733 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq 3734 * so ignore that status, too. 3735 */ 3736 if (!((1 << msk->first->sk_state) & 3737 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))) 3738 delta += READ_ONCE(tp->write_seq) - tp->snd_una; 3739 } 3740 if (delta > INT_MAX) 3741 delta = INT_MAX; 3742 3743 return (int)delta; 3744 } 3745 3746 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg) 3747 { 3748 struct mptcp_sock *msk = mptcp_sk(sk); 3749 bool slow; 3750 3751 switch (cmd) { 3752 case SIOCINQ: 3753 if (sk->sk_state == TCP_LISTEN) 3754 return -EINVAL; 3755 3756 lock_sock(sk); 3757 if (__mptcp_move_skbs(sk)) 3758 mptcp_cleanup_rbuf(msk, 0); 3759 *karg = mptcp_inq_hint(sk); 3760 release_sock(sk); 3761 break; 3762 case SIOCOUTQ: 3763 slow = lock_sock_fast(sk); 3764 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una)); 3765 unlock_sock_fast(sk, slow); 3766 break; 3767 case SIOCOUTQNSD: 3768 slow = lock_sock_fast(sk); 3769 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt); 3770 unlock_sock_fast(sk, slow); 3771 break; 3772 default: 3773 return -ENOIOCTLCMD; 3774 } 3775 3776 return 0; 3777 } 3778 3779 static int mptcp_connect(struct sock *sk, struct sockaddr_unsized *uaddr, 3780 int addr_len) 3781 { 3782 struct mptcp_subflow_context *subflow; 3783 struct mptcp_sock *msk = mptcp_sk(sk); 3784 int err = -EINVAL; 3785 struct sock *ssk; 3786 3787 ssk = __mptcp_nmpc_sk(msk); 3788 if (IS_ERR(ssk)) 3789 return PTR_ERR(ssk); 3790 3791 mptcp_set_state(sk, TCP_SYN_SENT); 3792 subflow = mptcp_subflow_ctx(ssk); 3793 #ifdef CONFIG_TCP_MD5SIG 3794 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of 3795 * TCP option space. 3796 */ 3797 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info)) 3798 mptcp_early_fallback(msk, subflow, MPTCP_MIB_MD5SIGFALLBACK); 3799 #endif 3800 if (subflow->request_mptcp) { 3801 if (mptcp_active_should_disable(sk)) 3802 mptcp_early_fallback(msk, subflow, 3803 MPTCP_MIB_MPCAPABLEACTIVEDISABLED); 3804 else if (mptcp_token_new_connect(ssk) < 0) 3805 mptcp_early_fallback(msk, subflow, 3806 MPTCP_MIB_TOKENFALLBACKINIT); 3807 } 3808 3809 WRITE_ONCE(msk->write_seq, subflow->idsn); 3810 WRITE_ONCE(msk->snd_nxt, subflow->idsn); 3811 WRITE_ONCE(msk->snd_una, subflow->idsn); 3812 if (likely(!__mptcp_check_fallback(msk))) 3813 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE); 3814 3815 /* if reaching here via the fastopen/sendmsg path, the caller already 3816 * acquired the subflow socket lock, too. 3817 */ 3818 if (!msk->fastopening) 3819 lock_sock(ssk); 3820 3821 /* the following mirrors closely a very small chunk of code from 3822 * __inet_stream_connect() 3823 */ 3824 if (ssk->sk_state != TCP_CLOSE) 3825 goto out; 3826 3827 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) { 3828 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len); 3829 if (err) 3830 goto out; 3831 } 3832 3833 err = ssk->sk_prot->connect(ssk, uaddr, addr_len); 3834 if (err < 0) 3835 goto out; 3836 3837 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk)); 3838 3839 out: 3840 if (!msk->fastopening) 3841 release_sock(ssk); 3842 3843 /* on successful connect, the msk state will be moved to established by 3844 * subflow_finish_connect() 3845 */ 3846 if (unlikely(err)) { 3847 /* avoid leaving a dangling token in an unconnected socket */ 3848 mptcp_token_destroy(msk); 3849 mptcp_set_state(sk, TCP_CLOSE); 3850 return err; 3851 } 3852 3853 mptcp_copy_inaddrs(sk, ssk); 3854 return 0; 3855 } 3856 3857 static struct proto mptcp_prot = { 3858 .name = "MPTCP", 3859 .owner = THIS_MODULE, 3860 .init = mptcp_init_sock, 3861 .connect = mptcp_connect, 3862 .disconnect = mptcp_disconnect, 3863 .close = mptcp_close, 3864 .setsockopt = mptcp_setsockopt, 3865 .getsockopt = mptcp_getsockopt, 3866 .shutdown = mptcp_shutdown, 3867 .destroy = mptcp_destroy, 3868 .sendmsg = mptcp_sendmsg, 3869 .ioctl = mptcp_ioctl, 3870 .recvmsg = mptcp_recvmsg, 3871 .release_cb = mptcp_release_cb, 3872 .hash = mptcp_hash, 3873 .unhash = mptcp_unhash, 3874 .get_port = mptcp_get_port, 3875 .stream_memory_free = mptcp_stream_memory_free, 3876 .sockets_allocated = &mptcp_sockets_allocated, 3877 3878 .memory_allocated = &net_aligned_data.tcp_memory_allocated, 3879 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc, 3880 3881 .memory_pressure = &tcp_memory_pressure, 3882 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem), 3883 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem), 3884 .sysctl_mem = sysctl_tcp_mem, 3885 .obj_size = sizeof(struct mptcp_sock), 3886 .slab_flags = SLAB_TYPESAFE_BY_RCU, 3887 .no_autobind = true, 3888 }; 3889 3890 static int mptcp_bind(struct socket *sock, struct sockaddr_unsized *uaddr, int addr_len) 3891 { 3892 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3893 struct sock *ssk, *sk = sock->sk; 3894 int err = -EINVAL; 3895 3896 lock_sock(sk); 3897 ssk = __mptcp_nmpc_sk(msk); 3898 if (IS_ERR(ssk)) { 3899 err = PTR_ERR(ssk); 3900 goto unlock; 3901 } 3902 3903 if (sk->sk_family == AF_INET) 3904 err = inet_bind_sk(ssk, uaddr, addr_len); 3905 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 3906 else if (sk->sk_family == AF_INET6) 3907 err = inet6_bind_sk(ssk, uaddr, addr_len); 3908 #endif 3909 if (!err) 3910 mptcp_copy_inaddrs(sk, ssk); 3911 3912 unlock: 3913 release_sock(sk); 3914 return err; 3915 } 3916 3917 static int mptcp_listen(struct socket *sock, int backlog) 3918 { 3919 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3920 struct sock *sk = sock->sk; 3921 struct sock *ssk; 3922 int err; 3923 3924 pr_debug("msk=%p\n", msk); 3925 3926 lock_sock(sk); 3927 3928 err = -EINVAL; 3929 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM) 3930 goto unlock; 3931 3932 ssk = __mptcp_nmpc_sk(msk); 3933 if (IS_ERR(ssk)) { 3934 err = PTR_ERR(ssk); 3935 goto unlock; 3936 } 3937 3938 mptcp_set_state(sk, TCP_LISTEN); 3939 sock_set_flag(sk, SOCK_RCU_FREE); 3940 3941 lock_sock(ssk); 3942 err = __inet_listen_sk(ssk, backlog); 3943 release_sock(ssk); 3944 mptcp_set_state(sk, inet_sk_state_load(ssk)); 3945 3946 if (!err) { 3947 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1); 3948 mptcp_copy_inaddrs(sk, ssk); 3949 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED); 3950 } 3951 3952 unlock: 3953 release_sock(sk); 3954 return err; 3955 } 3956 3957 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock, 3958 struct proto_accept_arg *arg) 3959 { 3960 struct mptcp_sock *msk = mptcp_sk(sock->sk); 3961 struct sock *ssk, *newsk; 3962 3963 pr_debug("msk=%p\n", msk); 3964 3965 /* Buggy applications can call accept on socket states other then LISTEN 3966 * but no need to allocate the first subflow just to error out. 3967 */ 3968 ssk = READ_ONCE(msk->first); 3969 if (!ssk) 3970 return -EINVAL; 3971 3972 pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk)); 3973 newsk = inet_csk_accept(ssk, arg); 3974 if (!newsk) 3975 return arg->err; 3976 3977 pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk)); 3978 if (sk_is_mptcp(newsk)) { 3979 struct mptcp_subflow_context *subflow; 3980 struct sock *new_mptcp_sock; 3981 3982 subflow = mptcp_subflow_ctx(newsk); 3983 new_mptcp_sock = subflow->conn; 3984 3985 /* is_mptcp should be false if subflow->conn is missing, see 3986 * subflow_syn_recv_sock() 3987 */ 3988 if (WARN_ON_ONCE(!new_mptcp_sock)) { 3989 tcp_sk(newsk)->is_mptcp = 0; 3990 goto tcpfallback; 3991 } 3992 3993 newsk = new_mptcp_sock; 3994 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK); 3995 3996 newsk->sk_kern_sock = arg->kern; 3997 lock_sock(newsk); 3998 __inet_accept(sock, newsock, newsk); 3999 4000 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags); 4001 msk = mptcp_sk(newsk); 4002 msk->in_accept_queue = 0; 4003 4004 /* set ssk->sk_socket of accept()ed flows to mptcp socket. 4005 * This is needed so NOSPACE flag can be set from tcp stack. 4006 */ 4007 mptcp_for_each_subflow(msk, subflow) { 4008 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4009 4010 if (!ssk->sk_socket) 4011 mptcp_sock_graft(ssk, newsock); 4012 } 4013 4014 mptcp_rps_record_subflows(msk); 4015 4016 /* Do late cleanup for the first subflow as necessary. Also 4017 * deal with bad peers not doing a complete shutdown. 4018 */ 4019 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) { 4020 __mptcp_close_ssk(newsk, msk->first, 4021 mptcp_subflow_ctx(msk->first), 0); 4022 if (unlikely(list_is_singular(&msk->conn_list))) 4023 mptcp_set_state(newsk, TCP_CLOSE); 4024 } 4025 } else { 4026 tcpfallback: 4027 newsk->sk_kern_sock = arg->kern; 4028 lock_sock(newsk); 4029 __inet_accept(sock, newsock, newsk); 4030 /* we are being invoked after accepting a non-mp-capable 4031 * flow: sk is a tcp_sk, not an mptcp one. 4032 * 4033 * Hand the socket over to tcp so all further socket ops 4034 * bypass mptcp. 4035 */ 4036 WRITE_ONCE(newsock->sk->sk_socket->ops, 4037 mptcp_fallback_tcp_ops(newsock->sk)); 4038 } 4039 release_sock(newsk); 4040 4041 return 0; 4042 } 4043 4044 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk) 4045 { 4046 struct sock *sk = (struct sock *)msk; 4047 4048 if (__mptcp_stream_is_writeable(sk, 1)) 4049 return EPOLLOUT | EPOLLWRNORM; 4050 4051 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); 4052 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */ 4053 if (__mptcp_stream_is_writeable(sk, 1)) 4054 return EPOLLOUT | EPOLLWRNORM; 4055 4056 return 0; 4057 } 4058 4059 static __poll_t mptcp_poll(struct file *file, struct socket *sock, 4060 struct poll_table_struct *wait) 4061 { 4062 struct sock *sk = sock->sk; 4063 struct mptcp_sock *msk; 4064 __poll_t mask = 0; 4065 u8 shutdown; 4066 int state; 4067 4068 msk = mptcp_sk(sk); 4069 sock_poll_wait(file, sock, wait); 4070 4071 state = inet_sk_state_load(sk); 4072 pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags); 4073 if (state == TCP_LISTEN) { 4074 struct sock *ssk = READ_ONCE(msk->first); 4075 4076 if (WARN_ON_ONCE(!ssk)) 4077 return 0; 4078 4079 return inet_csk_listen_poll(ssk); 4080 } 4081 4082 shutdown = READ_ONCE(sk->sk_shutdown); 4083 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE) 4084 mask |= EPOLLHUP; 4085 if (shutdown & RCV_SHUTDOWN) 4086 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP; 4087 4088 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) { 4089 mask |= mptcp_check_readable(sk); 4090 if (shutdown & SEND_SHUTDOWN) 4091 mask |= EPOLLOUT | EPOLLWRNORM; 4092 else 4093 mask |= mptcp_check_writeable(msk); 4094 } else if (state == TCP_SYN_SENT && 4095 inet_test_bit(DEFER_CONNECT, sk)) { 4096 /* cf tcp_poll() note about TFO */ 4097 mask |= EPOLLOUT | EPOLLWRNORM; 4098 } 4099 4100 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */ 4101 smp_rmb(); 4102 if (READ_ONCE(sk->sk_err)) 4103 mask |= EPOLLERR; 4104 4105 return mask; 4106 } 4107 4108 static const struct proto_ops mptcp_stream_ops = { 4109 .family = PF_INET, 4110 .owner = THIS_MODULE, 4111 .release = inet_release, 4112 .bind = mptcp_bind, 4113 .connect = inet_stream_connect, 4114 .socketpair = sock_no_socketpair, 4115 .accept = mptcp_stream_accept, 4116 .getname = inet_getname, 4117 .poll = mptcp_poll, 4118 .ioctl = inet_ioctl, 4119 .gettstamp = sock_gettstamp, 4120 .listen = mptcp_listen, 4121 .shutdown = inet_shutdown, 4122 .setsockopt = sock_common_setsockopt, 4123 .getsockopt = sock_common_getsockopt, 4124 .sendmsg = inet_sendmsg, 4125 .recvmsg = inet_recvmsg, 4126 .mmap = sock_no_mmap, 4127 .set_rcvlowat = mptcp_set_rcvlowat, 4128 }; 4129 4130 static struct inet_protosw mptcp_protosw = { 4131 .type = SOCK_STREAM, 4132 .protocol = IPPROTO_MPTCP, 4133 .prot = &mptcp_prot, 4134 .ops = &mptcp_stream_ops, 4135 .flags = INET_PROTOSW_ICSK, 4136 }; 4137 4138 static int mptcp_napi_poll(struct napi_struct *napi, int budget) 4139 { 4140 struct mptcp_delegated_action *delegated; 4141 struct mptcp_subflow_context *subflow; 4142 int work_done = 0; 4143 4144 delegated = container_of(napi, struct mptcp_delegated_action, napi); 4145 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) { 4146 struct sock *ssk = mptcp_subflow_tcp_sock(subflow); 4147 4148 bh_lock_sock_nested(ssk); 4149 if (!sock_owned_by_user(ssk)) { 4150 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0)); 4151 } else { 4152 /* tcp_release_cb_override already processed 4153 * the action or will do at next release_sock(). 4154 * In both case must dequeue the subflow here - on the same 4155 * CPU that scheduled it. 4156 */ 4157 smp_wmb(); 4158 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status); 4159 } 4160 bh_unlock_sock(ssk); 4161 sock_put(ssk); 4162 4163 if (++work_done == budget) 4164 return budget; 4165 } 4166 4167 /* always provide a 0 'work_done' argument, so that napi_complete_done 4168 * will not try accessing the NULL napi->dev ptr 4169 */ 4170 napi_complete_done(napi, 0); 4171 return work_done; 4172 } 4173 4174 void __init mptcp_proto_init(void) 4175 { 4176 struct mptcp_delegated_action *delegated; 4177 int cpu; 4178 4179 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo; 4180 4181 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL)) 4182 panic("Failed to allocate MPTCP pcpu counter\n"); 4183 4184 mptcp_napi_dev = alloc_netdev_dummy(0); 4185 if (!mptcp_napi_dev) 4186 panic("Failed to allocate MPTCP dummy netdev\n"); 4187 for_each_possible_cpu(cpu) { 4188 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu); 4189 INIT_LIST_HEAD(&delegated->head); 4190 netif_napi_add_tx(mptcp_napi_dev, &delegated->napi, 4191 mptcp_napi_poll); 4192 napi_enable(&delegated->napi); 4193 } 4194 4195 mptcp_subflow_init(); 4196 mptcp_pm_init(); 4197 mptcp_sched_init(); 4198 mptcp_token_init(); 4199 4200 if (proto_register(&mptcp_prot, 1) != 0) 4201 panic("Failed to register MPTCP proto.\n"); 4202 4203 inet_register_protosw(&mptcp_protosw); 4204 4205 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb)); 4206 } 4207 4208 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 4209 static const struct proto_ops mptcp_v6_stream_ops = { 4210 .family = PF_INET6, 4211 .owner = THIS_MODULE, 4212 .release = inet6_release, 4213 .bind = mptcp_bind, 4214 .connect = inet_stream_connect, 4215 .socketpair = sock_no_socketpair, 4216 .accept = mptcp_stream_accept, 4217 .getname = inet6_getname, 4218 .poll = mptcp_poll, 4219 .ioctl = inet6_ioctl, 4220 .gettstamp = sock_gettstamp, 4221 .listen = mptcp_listen, 4222 .shutdown = inet_shutdown, 4223 .setsockopt = sock_common_setsockopt, 4224 .getsockopt = sock_common_getsockopt, 4225 .sendmsg = inet6_sendmsg, 4226 .recvmsg = inet6_recvmsg, 4227 .mmap = sock_no_mmap, 4228 #ifdef CONFIG_COMPAT 4229 .compat_ioctl = inet6_compat_ioctl, 4230 #endif 4231 .set_rcvlowat = mptcp_set_rcvlowat, 4232 }; 4233 4234 static struct proto mptcp_v6_prot; 4235 4236 static struct inet_protosw mptcp_v6_protosw = { 4237 .type = SOCK_STREAM, 4238 .protocol = IPPROTO_MPTCP, 4239 .prot = &mptcp_v6_prot, 4240 .ops = &mptcp_v6_stream_ops, 4241 .flags = INET_PROTOSW_ICSK, 4242 }; 4243 4244 int __init mptcp_proto_v6_init(void) 4245 { 4246 int err; 4247 4248 mptcp_v6_prot = mptcp_prot; 4249 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name)); 4250 mptcp_v6_prot.slab = NULL; 4251 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock); 4252 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np); 4253 4254 err = proto_register(&mptcp_v6_prot, 1); 4255 if (err) 4256 return err; 4257 4258 err = inet6_register_protosw(&mptcp_v6_protosw); 4259 if (err) 4260 proto_unregister(&mptcp_v6_prot); 4261 4262 return err; 4263 } 4264 #endif 4265