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 <crypto/sha2.h> 11 #include <net/tcp.h> 12 #include <net/mptcp.h> 13 #include "protocol.h" 14 #include "mib.h" 15 16 #include <trace/events/mptcp.h> 17 18 static bool mptcp_cap_flag_sha256(u8 flags) 19 { 20 return (flags & MPTCP_CAP_FLAG_MASK) == MPTCP_CAP_HMAC_SHA256; 21 } 22 23 static void mptcp_parse_option(const struct sk_buff *skb, 24 const unsigned char *ptr, int opsize, 25 struct mptcp_options_received *mp_opt) 26 { 27 u8 subtype = *ptr >> 4; 28 int expected_opsize; 29 u16 subopt; 30 u8 version; 31 u8 flags; 32 u8 i; 33 34 switch (subtype) { 35 case MPTCPOPT_MP_CAPABLE: 36 /* strict size checking */ 37 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 38 if (skb->len > tcp_hdr(skb)->doff << 2) 39 expected_opsize = TCPOLEN_MPTCP_MPC_ACK_DATA; 40 else 41 expected_opsize = TCPOLEN_MPTCP_MPC_ACK; 42 subopt = OPTION_MPTCP_MPC_ACK; 43 } else { 44 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK) { 45 expected_opsize = TCPOLEN_MPTCP_MPC_SYNACK; 46 subopt = OPTION_MPTCP_MPC_SYNACK; 47 } else { 48 expected_opsize = TCPOLEN_MPTCP_MPC_SYN; 49 subopt = OPTION_MPTCP_MPC_SYN; 50 } 51 } 52 53 /* Cfr RFC 8684 Section 3.3.0: 54 * If a checksum is present but its use had 55 * not been negotiated in the MP_CAPABLE handshake, the receiver MUST 56 * close the subflow with a RST, as it is not behaving as negotiated. 57 * If a checksum is not present when its use has been negotiated, the 58 * receiver MUST close the subflow with a RST, as it is considered 59 * broken 60 * We parse even option with mismatching csum presence, so that 61 * later in subflow_data_ready we can trigger the reset. 62 */ 63 if (opsize != expected_opsize && 64 (expected_opsize != TCPOLEN_MPTCP_MPC_ACK_DATA || 65 opsize != TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM)) 66 break; 67 68 /* try to be gentle vs future versions on the initial syn */ 69 version = *ptr++ & MPTCP_VERSION_MASK; 70 if (opsize != TCPOLEN_MPTCP_MPC_SYN) { 71 if (version != MPTCP_SUPPORTED_VERSION) 72 break; 73 } else if (version < MPTCP_SUPPORTED_VERSION) { 74 break; 75 } 76 77 flags = *ptr++; 78 if (!mptcp_cap_flag_sha256(flags) || 79 (flags & MPTCP_CAP_EXTENSIBILITY)) 80 break; 81 82 /* RFC 6824, Section 3.1: 83 * "For the Checksum Required bit (labeled "A"), if either 84 * host requires the use of checksums, checksums MUST be used. 85 * In other words, the only way for checksums not to be used 86 * is if both hosts in their SYNs set A=0." 87 */ 88 if (flags & MPTCP_CAP_CHECKSUM_REQD) 89 mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD; 90 91 mp_opt->deny_join_id0 = !!(flags & MPTCP_CAP_DENY_JOIN_ID0); 92 93 mp_opt->suboptions |= subopt; 94 if (opsize >= TCPOLEN_MPTCP_MPC_SYNACK) { 95 mp_opt->sndr_key = get_unaligned_be64(ptr); 96 ptr += 8; 97 } 98 if (opsize >= TCPOLEN_MPTCP_MPC_ACK) { 99 mp_opt->rcvr_key = get_unaligned_be64(ptr); 100 ptr += 8; 101 } 102 if (opsize >= TCPOLEN_MPTCP_MPC_ACK_DATA) { 103 /* Section 3.1.: 104 * "the data parameters in a MP_CAPABLE are semantically 105 * equivalent to those in a DSS option and can be used 106 * interchangeably." 107 */ 108 mp_opt->suboptions |= OPTION_MPTCP_DSS; 109 mp_opt->use_map = 1; 110 mp_opt->mpc_map = 1; 111 mp_opt->data_len = get_unaligned_be16(ptr); 112 ptr += 2; 113 } 114 if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM) { 115 mp_opt->csum = get_unaligned((__force __sum16 *)ptr); 116 mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD; 117 ptr += 2; 118 } 119 pr_debug("MP_CAPABLE version=%x, flags=%x, optlen=%d sndr=%llu, rcvr=%llu len=%d csum=%u\n", 120 version, flags, opsize, mp_opt->sndr_key, 121 mp_opt->rcvr_key, mp_opt->data_len, mp_opt->csum); 122 break; 123 124 case MPTCPOPT_MP_JOIN: 125 if (opsize == TCPOLEN_MPTCP_MPJ_SYN) { 126 mp_opt->suboptions |= OPTION_MPTCP_MPJ_SYN; 127 mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP; 128 mp_opt->join_id = *ptr++; 129 mp_opt->token = get_unaligned_be32(ptr); 130 ptr += 4; 131 mp_opt->nonce = get_unaligned_be32(ptr); 132 ptr += 4; 133 pr_debug("MP_JOIN bkup=%u, id=%u, token=%u, nonce=%u\n", 134 mp_opt->backup, mp_opt->join_id, 135 mp_opt->token, mp_opt->nonce); 136 } else if (opsize == TCPOLEN_MPTCP_MPJ_SYNACK) { 137 mp_opt->suboptions |= OPTION_MPTCP_MPJ_SYNACK; 138 mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP; 139 mp_opt->join_id = *ptr++; 140 mp_opt->thmac = get_unaligned_be64(ptr); 141 ptr += 8; 142 mp_opt->nonce = get_unaligned_be32(ptr); 143 ptr += 4; 144 pr_debug("MP_JOIN bkup=%u, id=%u, thmac=%llu, nonce=%u\n", 145 mp_opt->backup, mp_opt->join_id, 146 mp_opt->thmac, mp_opt->nonce); 147 } else if (opsize == TCPOLEN_MPTCP_MPJ_ACK) { 148 mp_opt->suboptions |= OPTION_MPTCP_MPJ_ACK; 149 ptr += 2; 150 memcpy(mp_opt->hmac, ptr, MPTCPOPT_HMAC_LEN); 151 pr_debug("MP_JOIN hmac\n"); 152 } 153 break; 154 155 case MPTCPOPT_DSS: 156 pr_debug("DSS\n"); 157 ptr++; 158 159 flags = (*ptr++) & MPTCP_DSS_FLAG_MASK; 160 mp_opt->dsn64 = (flags & MPTCP_DSS_DSN64) != 0; 161 mp_opt->use_map = (flags & MPTCP_DSS_HAS_MAP) != 0; 162 mp_opt->ack64 = (flags & MPTCP_DSS_ACK64) != 0; 163 mp_opt->use_ack = (flags & MPTCP_DSS_HAS_ACK); 164 165 expected_opsize = TCPOLEN_MPTCP_DSS_BASE; 166 167 if (mp_opt->use_ack) { 168 if (mp_opt->ack64) 169 expected_opsize += TCPOLEN_MPTCP_DSS_ACK64; 170 else 171 expected_opsize += TCPOLEN_MPTCP_DSS_ACK32; 172 } 173 174 if (mp_opt->use_map) { 175 mp_opt->data_fin = (flags & MPTCP_DSS_DATA_FIN) != 0; 176 if (mp_opt->dsn64) 177 expected_opsize += TCPOLEN_MPTCP_DSS_MAP64; 178 else 179 expected_opsize += TCPOLEN_MPTCP_DSS_MAP32; 180 } 181 182 pr_debug("data_fin=%d dsn64=%d use_map=%d ack64=%d use_ack=%d\n", 183 mp_opt->data_fin, mp_opt->dsn64, 184 mp_opt->use_map, mp_opt->ack64, 185 mp_opt->use_ack); 186 187 /* Always parse any csum presence combination, we will enforce 188 * RFC 8684 Section 3.3.0 checks later in subflow_data_ready 189 */ 190 if (opsize != expected_opsize && 191 opsize != expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM) 192 break; 193 194 mp_opt->suboptions |= OPTION_MPTCP_DSS; 195 if (mp_opt->use_ack) { 196 if (mp_opt->ack64) { 197 mp_opt->data_ack = get_unaligned_be64(ptr); 198 ptr += 8; 199 } else { 200 mp_opt->data_ack = get_unaligned_be32(ptr); 201 ptr += 4; 202 } 203 204 pr_debug("data_ack=%llu\n", mp_opt->data_ack); 205 } 206 207 if (mp_opt->use_map) { 208 if (mp_opt->dsn64) { 209 mp_opt->data_seq = get_unaligned_be64(ptr); 210 ptr += 8; 211 } else { 212 mp_opt->data_seq = get_unaligned_be32(ptr); 213 ptr += 4; 214 } 215 216 mp_opt->subflow_seq = get_unaligned_be32(ptr); 217 ptr += 4; 218 219 mp_opt->data_len = get_unaligned_be16(ptr); 220 ptr += 2; 221 222 if (opsize == expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM) { 223 mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD; 224 mp_opt->csum = get_unaligned((__force __sum16 *)ptr); 225 ptr += 2; 226 } 227 228 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u csum=%d:%u\n", 229 mp_opt->data_seq, mp_opt->subflow_seq, 230 mp_opt->data_len, !!(mp_opt->suboptions & OPTION_MPTCP_CSUMREQD), 231 mp_opt->csum); 232 } 233 234 break; 235 236 case MPTCPOPT_ADD_ADDR: 237 mp_opt->echo = (*ptr++) & MPTCP_ADDR_ECHO; 238 if (!mp_opt->echo) { 239 if (opsize == TCPOLEN_MPTCP_ADD_ADDR || 240 opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT) 241 mp_opt->addr.family = AF_INET; 242 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 243 else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6 || 244 opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT) 245 mp_opt->addr.family = AF_INET6; 246 #endif 247 else 248 break; 249 } else { 250 if (opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE || 251 opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) 252 mp_opt->addr.family = AF_INET; 253 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 254 else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE || 255 opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) 256 mp_opt->addr.family = AF_INET6; 257 #endif 258 else 259 break; 260 } 261 262 mp_opt->suboptions |= OPTION_MPTCP_ADD_ADDR; 263 mp_opt->addr.id = *ptr++; 264 mp_opt->addr.port = 0; 265 mp_opt->ahmac = 0; 266 if (mp_opt->addr.family == AF_INET) { 267 memcpy((u8 *)&mp_opt->addr.addr.s_addr, (u8 *)ptr, 4); 268 ptr += 4; 269 if (opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT || 270 opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) { 271 mp_opt->addr.port = htons(get_unaligned_be16(ptr)); 272 ptr += 2; 273 } 274 } 275 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 276 else { 277 memcpy(mp_opt->addr.addr6.s6_addr, (u8 *)ptr, 16); 278 ptr += 16; 279 if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT || 280 opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) { 281 mp_opt->addr.port = htons(get_unaligned_be16(ptr)); 282 ptr += 2; 283 } 284 } 285 #endif 286 if (!mp_opt->echo) { 287 mp_opt->ahmac = get_unaligned_be64(ptr); 288 ptr += 8; 289 } 290 pr_debug("ADD_ADDR%s: id=%d, ahmac=%llu, echo=%d, port=%d\n", 291 (mp_opt->addr.family == AF_INET6) ? "6" : "", 292 mp_opt->addr.id, mp_opt->ahmac, mp_opt->echo, ntohs(mp_opt->addr.port)); 293 break; 294 295 case MPTCPOPT_RM_ADDR: 296 if (opsize < TCPOLEN_MPTCP_RM_ADDR_BASE + 1 || 297 opsize > TCPOLEN_MPTCP_RM_ADDR_BASE + MPTCP_RM_IDS_MAX) 298 break; 299 300 ptr++; 301 302 mp_opt->suboptions |= OPTION_MPTCP_RM_ADDR; 303 mp_opt->rm_list.nr = opsize - TCPOLEN_MPTCP_RM_ADDR_BASE; 304 for (i = 0; i < mp_opt->rm_list.nr; i++) 305 mp_opt->rm_list.ids[i] = *ptr++; 306 pr_debug("RM_ADDR: rm_list_nr=%d\n", mp_opt->rm_list.nr); 307 break; 308 309 case MPTCPOPT_MP_PRIO: 310 if (opsize != TCPOLEN_MPTCP_PRIO) 311 break; 312 313 mp_opt->suboptions |= OPTION_MPTCP_PRIO; 314 mp_opt->backup = *ptr++ & MPTCP_PRIO_BKUP; 315 pr_debug("MP_PRIO: prio=%d\n", mp_opt->backup); 316 break; 317 318 case MPTCPOPT_MP_FASTCLOSE: 319 if (opsize != TCPOLEN_MPTCP_FASTCLOSE) 320 break; 321 322 ptr += 2; 323 mp_opt->rcvr_key = get_unaligned_be64(ptr); 324 ptr += 8; 325 mp_opt->suboptions |= OPTION_MPTCP_FASTCLOSE; 326 pr_debug("MP_FASTCLOSE: recv_key=%llu\n", mp_opt->rcvr_key); 327 break; 328 329 case MPTCPOPT_RST: 330 if (opsize != TCPOLEN_MPTCP_RST) 331 break; 332 333 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_RST)) 334 break; 335 336 mp_opt->suboptions |= OPTION_MPTCP_RST; 337 flags = *ptr++; 338 mp_opt->reset_transient = flags & MPTCP_RST_TRANSIENT; 339 mp_opt->reset_reason = *ptr; 340 pr_debug("MP_RST: transient=%u reason=%u\n", 341 mp_opt->reset_transient, mp_opt->reset_reason); 342 break; 343 344 case MPTCPOPT_MP_FAIL: 345 if (opsize != TCPOLEN_MPTCP_FAIL) 346 break; 347 348 ptr += 2; 349 mp_opt->suboptions |= OPTION_MPTCP_FAIL; 350 mp_opt->fail_seq = get_unaligned_be64(ptr); 351 pr_debug("MP_FAIL: data_seq=%llu\n", mp_opt->fail_seq); 352 break; 353 354 default: 355 break; 356 } 357 } 358 359 void mptcp_get_options(const struct sk_buff *skb, 360 struct mptcp_options_received *mp_opt) 361 { 362 const struct tcphdr *th = tcp_hdr(skb); 363 const unsigned char *ptr; 364 int length; 365 366 /* Ensure that casting the whole status to u32 is efficient and safe */ 367 BUILD_BUG_ON(sizeof_field(struct mptcp_options_received, status) != sizeof(u32)); 368 BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct mptcp_options_received, status), 369 sizeof(u32))); 370 *(u32 *)&mp_opt->status = 0; 371 372 length = (th->doff * 4) - sizeof(struct tcphdr); 373 ptr = (const unsigned char *)(th + 1); 374 375 while (length > 0) { 376 int opcode = *ptr++; 377 int opsize; 378 379 switch (opcode) { 380 case TCPOPT_EOL: 381 return; 382 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 383 length--; 384 continue; 385 default: 386 if (length < 2) 387 return; 388 opsize = *ptr++; 389 if (opsize < 2) /* "silly options" */ 390 return; 391 if (opsize > length) 392 return; /* don't parse partial options */ 393 if (opcode == TCPOPT_MPTCP) 394 mptcp_parse_option(skb, ptr, opsize, mp_opt); 395 ptr += opsize - 2; 396 length -= opsize; 397 } 398 } 399 } 400 401 bool mptcp_syn_options(struct sock *sk, const struct sk_buff *skb, 402 unsigned int *size, struct mptcp_out_options *opts) 403 { 404 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 405 406 /* we will use snd_isn to detect first pkt [re]transmission 407 * in mptcp_established_options_mp() 408 */ 409 subflow->snd_isn = TCP_SKB_CB(skb)->end_seq; 410 if (subflow->request_mptcp) { 411 if (unlikely(subflow_simultaneous_connect(sk))) { 412 WARN_ON_ONCE(!mptcp_try_fallback(sk, MPTCP_MIB_SIMULTCONNFALLBACK)); 413 414 /* Ensure mptcp_finish_connect() will not process the 415 * MPC handshake. 416 */ 417 subflow->request_mptcp = 0; 418 return false; 419 } 420 421 opts->suboptions = OPTION_MPTCP_MPC_SYN; 422 opts->csum_reqd = mptcp_is_checksum_enabled(sock_net(sk)); 423 opts->allow_join_id0 = mptcp_allow_join_id0(sock_net(sk)); 424 *size = TCPOLEN_MPTCP_MPC_SYN; 425 return true; 426 } else if (subflow->request_join) { 427 pr_debug("remote_token=%u, nonce=%u\n", subflow->remote_token, 428 subflow->local_nonce); 429 opts->suboptions = OPTION_MPTCP_MPJ_SYN; 430 opts->join_id = subflow->local_id; 431 opts->token = subflow->remote_token; 432 opts->nonce = subflow->local_nonce; 433 opts->backup = subflow->request_bkup; 434 *size = TCPOLEN_MPTCP_MPJ_SYN; 435 return true; 436 } 437 return false; 438 } 439 440 static void clear_3rdack_retransmission(struct sock *sk) 441 { 442 struct inet_connection_sock *icsk = inet_csk(sk); 443 444 sk_stop_timer(sk, &icsk->icsk_delack_timer); 445 icsk->icsk_ack.ato = 0; 446 icsk->icsk_ack.pending &= ~(ICSK_ACK_SCHED | ICSK_ACK_TIMER); 447 } 448 449 static bool mptcp_established_options_mp(struct sock *sk, struct sk_buff *skb, 450 bool snd_data_fin_enable, int *size, 451 struct mptcp_out_options *opts) 452 { 453 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 454 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 455 struct mptcp_ext *mpext; 456 unsigned int data_len; 457 u8 len; 458 459 /* When skb is not available, we better over-estimate the emitted 460 * options len. A full DSS option (28 bytes) is longer than 461 * TCPOLEN_MPTCP_MPC_ACK_DATA(22) or TCPOLEN_MPTCP_MPJ_ACK(24), so 462 * tell the caller to defer the estimate to 463 * mptcp_established_options_dss(), which will reserve enough space. 464 */ 465 if (!skb) 466 return false; 467 468 /* MPC/MPJ needed only on 3rd ack packet, DATA_FIN and TCP shutdown take precedence */ 469 if (READ_ONCE(subflow->fully_established) || snd_data_fin_enable || 470 subflow->snd_isn != TCP_SKB_CB(skb)->seq || 471 sk->sk_state != TCP_ESTABLISHED) 472 return false; 473 474 if (subflow->mp_capable) { 475 mpext = mptcp_get_ext(skb); 476 data_len = mpext ? mpext->data_len : 0; 477 478 /* we will check ops->data_len in mptcp_write_options() to 479 * discriminate between TCPOLEN_MPTCP_MPC_ACK_DATA and 480 * TCPOLEN_MPTCP_MPC_ACK 481 */ 482 opts->data_len = data_len; 483 opts->suboptions = OPTION_MPTCP_MPC_ACK; 484 opts->sndr_key = subflow->local_key; 485 opts->rcvr_key = subflow->remote_key; 486 opts->csum_reqd = READ_ONCE(msk->csum_enabled); 487 opts->allow_join_id0 = mptcp_allow_join_id0(sock_net(sk)); 488 489 /* Section 3.1. 490 * The MP_CAPABLE option is carried on the SYN, SYN/ACK, and ACK 491 * packets that start the first subflow of an MPTCP connection, 492 * as well as the first packet that carries data 493 */ 494 if (data_len > 0) { 495 len = TCPOLEN_MPTCP_MPC_ACK_DATA; 496 if (opts->csum_reqd) { 497 /* we need to propagate more info to csum the pseudo hdr */ 498 opts->data_seq = mpext->data_seq; 499 opts->subflow_seq = mpext->subflow_seq; 500 opts->csum = mpext->csum; 501 len += TCPOLEN_MPTCP_DSS_CHECKSUM; 502 } 503 *size = ALIGN(len, 4); 504 } else { 505 *size = TCPOLEN_MPTCP_MPC_ACK; 506 } 507 508 pr_debug("subflow=%p, local_key=%llu, remote_key=%llu map_len=%d\n", 509 subflow, subflow->local_key, subflow->remote_key, 510 data_len); 511 512 return true; 513 } else if (subflow->mp_join) { 514 opts->suboptions = OPTION_MPTCP_MPJ_ACK; 515 memcpy(opts->hmac, subflow->hmac, MPTCPOPT_HMAC_LEN); 516 *size = TCPOLEN_MPTCP_MPJ_ACK; 517 pr_debug("subflow=%p\n", subflow); 518 519 /* we can use the full delegate action helper only from BH context 520 * If we are in process context - sk is flushing the backlog at 521 * socket lock release time - just set the appropriate flag, will 522 * be handled by the release callback 523 */ 524 if (sock_owned_by_user(sk)) 525 set_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status); 526 else 527 mptcp_subflow_delegate(subflow, MPTCP_DELEGATE_ACK); 528 return true; 529 } 530 return false; 531 } 532 533 static void mptcp_write_data_fin(struct mptcp_subflow_context *subflow, 534 struct sk_buff *skb, struct mptcp_ext *ext) 535 { 536 /* The write_seq value has already been incremented, so the actual 537 * sequence number for the DATA_FIN is one less. 538 */ 539 u64 data_fin_tx_seq = READ_ONCE(mptcp_sk(subflow->conn)->write_seq) - 1; 540 541 if (!ext->use_map || !skb->len) { 542 /* RFC6824 requires a DSS mapping with specific values 543 * if DATA_FIN is set but no data payload is mapped 544 */ 545 ext->data_fin = 1; 546 ext->use_map = 1; 547 ext->dsn64 = 1; 548 ext->data_seq = data_fin_tx_seq; 549 ext->subflow_seq = 0; 550 ext->data_len = 1; 551 } else if (ext->data_seq + ext->data_len == data_fin_tx_seq) { 552 /* If there's an existing DSS mapping and it is the 553 * final mapping, DATA_FIN consumes 1 additional byte of 554 * mapping space. 555 */ 556 ext->data_fin = 1; 557 ext->data_len++; 558 } 559 } 560 561 static bool mptcp_established_options_dss(struct sock *sk, struct sk_buff *skb, 562 bool snd_data_fin_enable, int *size, 563 struct mptcp_out_options *opts) 564 { 565 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 566 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 567 struct tcp_sock *tp = tcp_sk(sk); 568 unsigned int dss_size = 0; 569 struct mptcp_ext *mpext; 570 unsigned int ack_size; 571 bool ret = false; 572 573 /* Zero `use_ack` and `use_map` flags with one shot. */ 574 memset(&opts->ext_copy.flags, 0, sizeof(opts->ext_copy.flags)); 575 opts->csum_reqd = READ_ONCE(msk->csum_enabled); 576 mpext = skb ? mptcp_get_ext(skb) : NULL; 577 578 if (!skb || (mpext && mpext->use_map) || snd_data_fin_enable) { 579 unsigned int map_size = TCPOLEN_MPTCP_DSS_BASE + TCPOLEN_MPTCP_DSS_MAP64; 580 581 if (mpext) { 582 if (opts->csum_reqd) 583 map_size += TCPOLEN_MPTCP_DSS_CHECKSUM; 584 585 opts->ext_copy = *mpext; 586 } 587 588 dss_size = map_size; 589 if (skb && snd_data_fin_enable) 590 mptcp_write_data_fin(subflow, skb, &opts->ext_copy); 591 opts->suboptions = OPTION_MPTCP_DSS; 592 ret = true; 593 } 594 595 /* passive sockets msk will set the 'can_ack' after accept(), even 596 * if the first subflow may have the already the remote key handy 597 */ 598 if (!READ_ONCE(msk->can_ack)) { 599 *size = ALIGN(dss_size, 4); 600 return ret; 601 } 602 603 if (READ_ONCE(msk->use_64bit_ack)) { 604 ack_size = TCPOLEN_MPTCP_DSS_ACK64; 605 opts->ext_copy.ack64 = 1; 606 } else { 607 ack_size = TCPOLEN_MPTCP_DSS_ACK32; 608 opts->ext_copy.ack64 = 0; 609 } 610 opts->ext_copy.use_ack = 1; 611 opts->suboptions = OPTION_MPTCP_DSS; 612 613 /* Add kind/length/subtype/flag overhead if mapping is not populated */ 614 if (dss_size == 0) 615 ack_size += TCPOLEN_MPTCP_DSS_BASE; 616 617 /* The caller is __tcp_transmit_skb(), and will compute the new rcv 618 * wnd soon: ensure that the window can shrink. 619 */ 620 if (skb) 621 tp->rcv_wnd = tp->rcv_nxt - tp->rcv_wup; 622 623 dss_size += ack_size; 624 625 *size = ALIGN(dss_size, 4); 626 return true; 627 } 628 629 static u64 add_addr_generate_hmac(u64 key1, u64 key2, 630 struct mptcp_addr_info *addr) 631 { 632 u16 port = ntohs(addr->port); 633 u8 hmac[SHA256_DIGEST_SIZE]; 634 u8 msg[19]; 635 int i = 0; 636 637 msg[i++] = addr->id; 638 if (addr->family == AF_INET) { 639 memcpy(&msg[i], &addr->addr.s_addr, 4); 640 i += 4; 641 } 642 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 643 else if (addr->family == AF_INET6) { 644 memcpy(&msg[i], &addr->addr6.s6_addr, 16); 645 i += 16; 646 } 647 #endif 648 msg[i++] = port >> 8; 649 msg[i++] = port & 0xFF; 650 651 mptcp_crypto_hmac_sha(key1, key2, msg, i, hmac); 652 653 return get_unaligned_be64(&hmac[SHA256_DIGEST_SIZE - sizeof(u64)]); 654 } 655 656 static bool mptcp_established_options_add_addr(struct sock *sk, 657 struct sk_buff *skb, int *size, 658 unsigned int remaining, 659 bool has_ts, 660 struct mptcp_out_options *opts) 661 { 662 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 663 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 664 struct mptcp_addr_info addr; 665 bool drop_ts = has_ts; 666 bool echo; 667 668 /* add addr will strip the existing options, be sure to avoid breaking 669 * MPC/MPJ handshakes 670 */ 671 if (!mptcp_pm_should_add_signal(msk) || 672 (opts->suboptions & (OPTION_MPTCP_MPJ_ACK | OPTION_MPTCP_MPC_ACK)) || 673 !skb || !skb_is_tcp_pure_ack(skb) || 674 !mptcp_pm_add_addr_signal(msk, size, remaining, &addr, &echo, 675 &drop_ts)) 676 return false; 677 678 pr_debug("drop other suboptions\n"); 679 opts->suboptions = OPTION_MPTCP_ADD_ADDR; 680 opts->drop_ts = drop_ts; 681 opts->addr = addr; 682 if (!echo) { 683 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ADDADDRTX); 684 opts->ahmac = add_addr_generate_hmac(READ_ONCE(msk->local_key), 685 READ_ONCE(msk->remote_key), 686 &opts->addr); 687 } else { 688 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ECHOADDTX); 689 opts->ahmac = 0; 690 } 691 pr_debug("addr_id=%d, ahmac=%llu, echo=%d, port=%d\n", 692 opts->addr.id, opts->ahmac, echo, ntohs(opts->addr.port)); 693 694 return true; 695 } 696 697 static bool mptcp_established_options_rm_addr(struct sock *sk, int *size, 698 unsigned int remaining, 699 struct mptcp_out_options *opts) 700 { 701 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 702 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 703 struct mptcp_rm_list rm_list; 704 int i; 705 706 if (!mptcp_pm_should_rm_signal(msk) || 707 !(mptcp_pm_rm_addr_signal(msk, remaining, &rm_list, size))) 708 return false; 709 710 opts->suboptions |= OPTION_MPTCP_RM_ADDR; 711 opts->rm_list = rm_list; 712 713 for (i = 0; i < opts->rm_list.nr; i++) 714 pr_debug("rm_list_ids[%d]=%d\n", i, opts->rm_list.ids[i]); 715 MPTCP_ADD_STATS(sock_net(sk), MPTCP_MIB_RMADDRTX, opts->rm_list.nr); 716 return true; 717 } 718 719 static bool mptcp_established_options_mp_prio(struct sock *sk, int *size, 720 unsigned int remaining, 721 struct mptcp_out_options *opts) 722 { 723 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 724 725 /* can't send MP_PRIO with MPC, as they share the same option space: 726 * 'backup'. Also it makes no sense at all 727 */ 728 if (!subflow->send_mp_prio || (opts->suboptions & OPTIONS_MPTCP_MPC)) 729 return false; 730 731 /* account for the trailing 'nop' option */ 732 if (remaining < TCPOLEN_MPTCP_PRIO_ALIGN) 733 return false; 734 735 *size = TCPOLEN_MPTCP_PRIO_ALIGN; 736 opts->suboptions |= OPTION_MPTCP_PRIO; 737 opts->backup = subflow->request_bkup; 738 739 pr_debug("prio=%d\n", opts->backup); 740 741 return true; 742 } 743 744 static noinline bool mptcp_established_options_rst(struct sock *sk, 745 int *size, 746 unsigned int remaining, 747 struct mptcp_out_options *opts) 748 { 749 const struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 750 751 if (remaining < TCPOLEN_MPTCP_RST) 752 return false; 753 754 *size = TCPOLEN_MPTCP_RST; 755 opts->suboptions |= OPTION_MPTCP_RST; 756 opts->reset_transient = subflow->reset_transient; 757 opts->reset_reason = subflow->reset_reason; 758 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPRSTTX); 759 760 return true; 761 } 762 763 static bool mptcp_established_options_fastclose(struct sock *sk, int *size, 764 unsigned int remaining, 765 struct mptcp_out_options *opts) 766 { 767 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 768 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 769 770 if (likely(!subflow->send_fastclose)) 771 return false; 772 773 if (remaining < TCPOLEN_MPTCP_FASTCLOSE) 774 return false; 775 776 *size = TCPOLEN_MPTCP_FASTCLOSE; 777 opts->suboptions |= OPTION_MPTCP_FASTCLOSE; 778 opts->rcvr_key = READ_ONCE(msk->remote_key); 779 780 pr_debug("FASTCLOSE key=%llu\n", opts->rcvr_key); 781 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFASTCLOSETX); 782 return true; 783 } 784 785 static bool mptcp_established_options_mp_fail(struct sock *sk, int *size, 786 unsigned int remaining, 787 struct mptcp_out_options *opts) 788 { 789 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 790 791 if (likely(!subflow->send_mp_fail)) 792 return false; 793 794 if (remaining < TCPOLEN_MPTCP_FAIL) 795 return false; 796 797 *size = TCPOLEN_MPTCP_FAIL; 798 opts->suboptions |= OPTION_MPTCP_FAIL; 799 opts->fail_seq = subflow->map_seq; 800 801 pr_debug("MP_FAIL fail_seq=%llu\n", opts->fail_seq); 802 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFAILTX); 803 804 return true; 805 } 806 807 int mptcp_established_options(struct sock *sk, struct sk_buff *skb, 808 unsigned int remaining, bool has_ts, 809 struct mptcp_out_options *opts) 810 { 811 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 812 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 813 int total_size = 0; 814 bool snd_data_fin; 815 bool ret = false; 816 int opt_size = 0; 817 818 opts->suboptions = 0; 819 820 /* Force later mptcp_write_options(), but do not use any actual 821 * option space. 822 */ 823 if (unlikely(__mptcp_check_fallback(msk) && !mptcp_check_infinite_map(skb))) 824 return 0; 825 826 if (unlikely(skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_RST)) { 827 if (mptcp_established_options_fastclose(sk, &opt_size, remaining, opts) || 828 mptcp_established_options_mp_fail(sk, &opt_size, remaining, opts)) { 829 total_size += opt_size; 830 remaining -= opt_size; 831 } 832 /* MP_RST can be used with MP_FASTCLOSE and MP_FAIL if there is room */ 833 if (mptcp_established_options_rst(sk, &opt_size, remaining, opts)) { 834 total_size += opt_size; 835 remaining -= opt_size; 836 } 837 return total_size; 838 } 839 840 snd_data_fin = mptcp_data_fin_enabled(msk); 841 if (mptcp_established_options_mp(sk, skb, snd_data_fin, &opt_size, opts)) 842 ret = true; 843 else if (mptcp_established_options_dss(sk, skb, snd_data_fin, &opt_size, opts)) { 844 int mp_fail_size; 845 846 ret = true; 847 if (mptcp_established_options_mp_fail(sk, &mp_fail_size, 848 remaining - opt_size, opts)) { 849 total_size += opt_size + mp_fail_size; 850 remaining -= opt_size - mp_fail_size; 851 return total_size; 852 } 853 } 854 855 /* we reserved enough space for the above options, and exceeding the 856 * TCP option space would be fatal 857 */ 858 if (WARN_ON_ONCE(opt_size > remaining)) 859 return -1; 860 861 total_size += opt_size; 862 remaining -= opt_size; 863 if (mptcp_established_options_add_addr(sk, skb, &opt_size, remaining, 864 has_ts, opts)) { 865 total_size += opt_size; 866 remaining -= opt_size; 867 ret = true; 868 } else if (mptcp_established_options_rm_addr(sk, &opt_size, remaining, opts)) { 869 total_size += opt_size; 870 remaining -= opt_size; 871 ret = true; 872 } 873 874 if (mptcp_established_options_mp_prio(sk, &opt_size, remaining, opts)) { 875 total_size += opt_size; 876 remaining -= opt_size; 877 ret = true; 878 } 879 880 return ret ? total_size : -1; 881 } 882 883 bool mptcp_synack_options(const struct request_sock *req, unsigned int *size, 884 struct mptcp_out_options *opts) 885 { 886 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); 887 888 if (subflow_req->mp_capable) { 889 opts->suboptions = OPTION_MPTCP_MPC_SYNACK; 890 opts->sndr_key = subflow_req->local_key; 891 opts->csum_reqd = subflow_req->csum_reqd; 892 opts->allow_join_id0 = subflow_req->allow_join_id0; 893 *size = TCPOLEN_MPTCP_MPC_SYNACK; 894 pr_debug("subflow_req=%p, local_key=%llu\n", 895 subflow_req, subflow_req->local_key); 896 return true; 897 } else if (subflow_req->mp_join) { 898 opts->suboptions = OPTION_MPTCP_MPJ_SYNACK; 899 opts->backup = subflow_req->request_bkup; 900 opts->join_id = subflow_req->local_id; 901 opts->thmac = subflow_req->thmac; 902 opts->nonce = subflow_req->local_nonce; 903 pr_debug("req=%p, bkup=%u, id=%u, thmac=%llu, nonce=%u\n", 904 subflow_req, opts->backup, opts->join_id, 905 opts->thmac, opts->nonce); 906 *size = TCPOLEN_MPTCP_MPJ_SYNACK; 907 return true; 908 } 909 return false; 910 } 911 912 static bool check_fully_established(struct mptcp_sock *msk, struct sock *ssk, 913 struct mptcp_subflow_context *subflow, 914 struct sk_buff *skb, 915 struct mptcp_options_received *mp_opt) 916 { 917 /* here we can process OoO, in-window pkts, only in-sequence 4th ack 918 * will make the subflow fully established 919 */ 920 if (likely(READ_ONCE(subflow->fully_established))) { 921 /* on passive sockets, check for 3rd ack retransmission 922 * note that msk is always set by subflow_syn_recv_sock() 923 * for mp_join subflows 924 */ 925 if (TCP_SKB_CB(skb)->seq == subflow->ssn_offset + 1 && 926 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq && 927 subflow->mp_join && (mp_opt->suboptions & OPTIONS_MPTCP_MPJ) && 928 !subflow->request_join) 929 tcp_send_ack(ssk); 930 goto check_notify; 931 } 932 933 /* we must process OoO packets before the first subflow is fully 934 * established. OoO packets are instead a protocol violation 935 * for MP_JOIN subflows as the peer must not send any data 936 * before receiving the forth ack - cfr. RFC 8684 section 3.2. 937 */ 938 if (TCP_SKB_CB(skb)->seq != subflow->ssn_offset + 1) { 939 if (subflow->mp_join) 940 goto reset; 941 if (subflow->is_mptfo && mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) 942 goto set_fully_established; 943 return subflow->mp_capable; 944 } 945 946 if (subflow->remote_key_valid && 947 (((mp_opt->suboptions & OPTION_MPTCP_DSS) && mp_opt->use_ack) || 948 ((mp_opt->suboptions & OPTION_MPTCP_ADD_ADDR) && 949 (!mp_opt->echo || subflow->mp_join)))) { 950 /* subflows are fully established as soon as we get any 951 * additional ack, including ADD_ADDR. 952 */ 953 goto set_fully_established; 954 } 955 956 /* If the first established packet does not contain MP_CAPABLE + data 957 * then fallback to TCP. Fallback scenarios requires a reset for 958 * MP_JOIN subflows. 959 */ 960 if (!(mp_opt->suboptions & OPTIONS_MPTCP_MPC)) { 961 if (subflow->mp_join) 962 goto reset; 963 subflow->mp_capable = 0; 964 if (!mptcp_try_fallback(ssk, MPTCP_MIB_MPCAPABLEDATAFALLBACK)) { 965 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED); 966 goto reset; 967 } 968 return false; 969 } 970 971 if (unlikely(!READ_ONCE(msk->pm.server_side))) 972 pr_warn_once("bogus mpc option on established client sk"); 973 974 set_fully_established: 975 if (mp_opt->deny_join_id0) 976 WRITE_ONCE(msk->pm.remote_deny_join_id0, true); 977 978 mptcp_data_lock((struct sock *)msk); 979 __mptcp_subflow_fully_established(msk, subflow, mp_opt); 980 mptcp_data_unlock((struct sock *)msk); 981 982 check_notify: 983 /* if the subflow is not already linked into the conn_list, we can't 984 * notify the PM: this subflow is still on the listener queue 985 * and the PM possibly acquiring the subflow lock could race with 986 * the listener close 987 */ 988 if (likely(subflow->pm_notified) || list_empty(&subflow->node)) 989 return true; 990 991 subflow->pm_notified = 1; 992 if (subflow->mp_join) { 993 clear_3rdack_retransmission(ssk); 994 mptcp_pm_subflow_established(msk); 995 } else { 996 mptcp_pm_fully_established(msk, ssk); 997 } 998 return true; 999 1000 reset: 1001 mptcp_subflow_reset(ssk); 1002 return false; 1003 } 1004 1005 u64 __mptcp_expand_seq(u64 old_seq, u64 cur_seq) 1006 { 1007 u32 old_seq32, cur_seq32; 1008 1009 old_seq32 = (u32)old_seq; 1010 cur_seq32 = (u32)cur_seq; 1011 cur_seq = (old_seq & GENMASK_ULL(63, 32)) + cur_seq32; 1012 if (unlikely(cur_seq32 < old_seq32 && before(old_seq32, cur_seq32))) 1013 return cur_seq + (1LL << 32); 1014 1015 /* reverse wrap could happen, too */ 1016 if (unlikely(cur_seq32 > old_seq32 && after(old_seq32, cur_seq32))) 1017 return cur_seq - (1LL << 32); 1018 return cur_seq; 1019 } 1020 1021 static void __mptcp_snd_una_update(struct mptcp_sock *msk, u64 new_snd_una) 1022 { 1023 msk->bytes_acked += new_snd_una - msk->snd_una; 1024 WRITE_ONCE(msk->snd_una, new_snd_una); 1025 } 1026 1027 static void rwin_update(struct mptcp_sock *msk, struct sock *ssk, 1028 struct sk_buff *skb) 1029 { 1030 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk); 1031 struct tcp_sock *tp = tcp_sk(ssk); 1032 u64 mptcp_rcv_wnd; 1033 1034 /* Avoid touching extra cachelines if TCP is going to accept this 1035 * skb without filling the TCP-level window even with a possibly 1036 * outdated mptcp-level rwin. 1037 */ 1038 if (!skb->len || skb->len < tcp_receive_window(tp)) 1039 return; 1040 1041 mptcp_rcv_wnd = atomic64_read(&msk->rcv_wnd_sent); 1042 if (!after64(mptcp_rcv_wnd, subflow->rcv_wnd_sent)) 1043 return; 1044 1045 /* Some other subflow grew the mptcp-level rwin since rcv_wup, 1046 * resync. 1047 */ 1048 tp->rcv_wnd += mptcp_rcv_wnd - subflow->rcv_wnd_sent; 1049 tcp_update_max_rcv_wnd_seq(tp); 1050 subflow->rcv_wnd_sent = mptcp_rcv_wnd; 1051 } 1052 1053 static void ack_update_msk(struct mptcp_sock *msk, 1054 struct sock *ssk, 1055 struct mptcp_options_received *mp_opt) 1056 { 1057 u64 new_wnd_end, new_snd_una, snd_nxt = READ_ONCE(msk->snd_nxt); 1058 struct sock *sk = (struct sock *)msk; 1059 u64 old_snd_una; 1060 1061 mptcp_data_lock(sk); 1062 1063 /* avoid ack expansion on update conflict, to reduce the risk of 1064 * wrongly expanding to a future ack sequence number, which is way 1065 * more dangerous than missing an ack 1066 */ 1067 old_snd_una = msk->snd_una; 1068 new_snd_una = mptcp_expand_seq(old_snd_una, mp_opt->data_ack, mp_opt->ack64); 1069 1070 /* ACK for data not even sent yet? Ignore.*/ 1071 if (unlikely(after64(new_snd_una, snd_nxt))) 1072 new_snd_una = old_snd_una; 1073 1074 new_wnd_end = new_snd_una + tcp_sk(ssk)->snd_wnd; 1075 1076 if (after64(new_wnd_end, msk->wnd_end)) 1077 WRITE_ONCE(msk->wnd_end, new_wnd_end); 1078 1079 /* this assumes mptcp_incoming_options() is invoked after tcp_ack() */ 1080 if (after64(msk->wnd_end, snd_nxt)) 1081 __mptcp_check_push(sk, ssk); 1082 1083 if (after64(new_snd_una, old_snd_una)) { 1084 __mptcp_snd_una_update(msk, new_snd_una); 1085 __mptcp_data_acked(sk); 1086 } 1087 msk->last_ack_recv = tcp_jiffies32; 1088 mptcp_data_unlock(sk); 1089 1090 trace_ack_update_msk(mp_opt->data_ack, 1091 old_snd_una, new_snd_una, 1092 new_wnd_end, READ_ONCE(msk->wnd_end)); 1093 } 1094 1095 bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit) 1096 { 1097 /* Skip if DATA_FIN was already received. 1098 * If updating simultaneously with the recvmsg loop, values 1099 * should match. If they mismatch, the peer is misbehaving and 1100 * we will prefer the most recent information. 1101 */ 1102 if (READ_ONCE(msk->rcv_data_fin)) 1103 return false; 1104 1105 WRITE_ONCE(msk->rcv_data_fin_seq, 1106 mptcp_expand_seq(READ_ONCE(msk->ack_seq), data_fin_seq, use_64bit)); 1107 WRITE_ONCE(msk->rcv_data_fin, 1); 1108 1109 return true; 1110 } 1111 1112 static bool add_addr_hmac_valid(struct mptcp_sock *msk, 1113 struct mptcp_options_received *mp_opt) 1114 { 1115 u64 hmac = 0; 1116 1117 if (mp_opt->echo) 1118 return true; 1119 1120 hmac = add_addr_generate_hmac(READ_ONCE(msk->remote_key), 1121 READ_ONCE(msk->local_key), 1122 &mp_opt->addr); 1123 1124 pr_debug("msk=%p, ahmac=%llu, mp_opt->ahmac=%llu\n", 1125 msk, hmac, mp_opt->ahmac); 1126 1127 return hmac == mp_opt->ahmac; 1128 } 1129 1130 /* Return false in case of error (or subflow has been reset), 1131 * else return true. 1132 */ 1133 bool mptcp_incoming_options(struct sock *sk, struct sk_buff *skb) 1134 { 1135 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); 1136 struct mptcp_sock *msk = mptcp_sk(subflow->conn); 1137 struct mptcp_options_received mp_opt; 1138 struct mptcp_ext *mpext; 1139 1140 if (__mptcp_check_fallback(msk)) { 1141 /* Keep it simple and unconditionally trigger send data cleanup and 1142 * pending queue spooling. We will need to acquire the data lock 1143 * for more accurate checks, and once the lock is acquired, such 1144 * helpers are cheap. 1145 */ 1146 mptcp_data_lock(subflow->conn); 1147 if (sk_stream_memory_free(sk)) 1148 __mptcp_check_push(subflow->conn, sk); 1149 1150 /* on fallback we just need to ignore the msk-level snd_una, as 1151 * this is really plain TCP 1152 */ 1153 __mptcp_snd_una_update(msk, READ_ONCE(msk->snd_nxt)); 1154 1155 __mptcp_data_acked(subflow->conn); 1156 mptcp_data_unlock(subflow->conn); 1157 return true; 1158 } 1159 1160 mptcp_get_options(skb, &mp_opt); 1161 1162 /* The subflow can be in close state only if check_fully_established() 1163 * just sent a reset. If so, tell the caller to ignore the current packet. 1164 */ 1165 if (!check_fully_established(msk, sk, subflow, skb, &mp_opt)) 1166 return sk->sk_state != TCP_CLOSE; 1167 1168 if (unlikely(mp_opt.suboptions != OPTION_MPTCP_DSS)) { 1169 if ((mp_opt.suboptions & OPTION_MPTCP_FASTCLOSE) && 1170 READ_ONCE(msk->local_key) == mp_opt.rcvr_key) { 1171 WRITE_ONCE(msk->rcv_fastclose, true); 1172 mptcp_schedule_work((struct sock *)msk); 1173 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFASTCLOSERX); 1174 } 1175 1176 if ((mp_opt.suboptions & OPTION_MPTCP_ADD_ADDR) && 1177 add_addr_hmac_valid(msk, &mp_opt)) { 1178 if (!mp_opt.echo) { 1179 mptcp_pm_add_addr_received(sk, &mp_opt.addr); 1180 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ADDADDR); 1181 } else { 1182 mptcp_pm_add_addr_echoed(msk, &mp_opt.addr); 1183 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ECHOADD); 1184 } 1185 1186 if (mp_opt.addr.port) 1187 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_PORTADD); 1188 } 1189 1190 if (mp_opt.suboptions & OPTION_MPTCP_RM_ADDR) 1191 mptcp_pm_rm_addr_received(msk, &mp_opt.rm_list); 1192 1193 if (mp_opt.suboptions & OPTION_MPTCP_PRIO) { 1194 mptcp_pm_mp_prio_received(sk, mp_opt.backup); 1195 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPPRIORX); 1196 } 1197 1198 if (mp_opt.suboptions & OPTION_MPTCP_FAIL) { 1199 mptcp_pm_mp_fail_received(sk, mp_opt.fail_seq); 1200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFAILRX); 1201 } 1202 1203 if (mp_opt.suboptions & OPTION_MPTCP_RST) { 1204 subflow->reset_seen = 1; 1205 subflow->reset_reason = mp_opt.reset_reason; 1206 subflow->reset_transient = mp_opt.reset_transient; 1207 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPRSTRX); 1208 } 1209 1210 if (!(mp_opt.suboptions & OPTION_MPTCP_DSS)) 1211 return true; 1212 } 1213 1214 /* we can't wait for recvmsg() to update the ack_seq, otherwise 1215 * monodirectional flows will stuck 1216 */ 1217 if (mp_opt.use_ack) 1218 ack_update_msk(msk, sk, &mp_opt); 1219 rwin_update(msk, sk, skb); 1220 1221 /* Zero-data-length packets are dropped by the caller and not 1222 * propagated to the MPTCP layer, so the skb extension does not 1223 * need to be allocated or populated. DATA_FIN information, if 1224 * present, needs to be updated here before the skb is freed. 1225 */ 1226 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 1227 if (mp_opt.data_fin && mp_opt.data_len == 1 && 1228 mptcp_update_rcv_data_fin(msk, mp_opt.data_seq, mp_opt.dsn64)) 1229 mptcp_schedule_work((struct sock *)msk); 1230 1231 return true; 1232 } 1233 1234 mpext = skb_ext_add(skb, SKB_EXT_MPTCP); 1235 if (!mpext) 1236 return false; 1237 1238 memset(mpext, 0, sizeof(*mpext)); 1239 1240 if (likely(mp_opt.use_map)) { 1241 if (mp_opt.mpc_map) { 1242 /* this is an MP_CAPABLE carrying MPTCP data 1243 * we know this map the first chunk of data 1244 */ 1245 mptcp_crypto_key_sha(subflow->remote_key, NULL, 1246 &mpext->data_seq); 1247 mpext->data_seq++; 1248 mpext->subflow_seq = 1; 1249 mpext->dsn64 = 1; 1250 mpext->mpc_map = 1; 1251 mpext->data_fin = 0; 1252 } else { 1253 mpext->data_seq = mp_opt.data_seq; 1254 mpext->subflow_seq = mp_opt.subflow_seq; 1255 mpext->dsn64 = mp_opt.dsn64; 1256 mpext->data_fin = mp_opt.data_fin; 1257 } 1258 mpext->data_len = mp_opt.data_len; 1259 mpext->use_map = 1; 1260 mpext->csum_reqd = !!(mp_opt.suboptions & OPTION_MPTCP_CSUMREQD); 1261 1262 if (mpext->csum_reqd) 1263 mpext->csum = mp_opt.csum; 1264 } 1265 1266 return true; 1267 } 1268 1269 static u64 mptcp_set_rwin(struct mptcp_sock *msk, struct tcp_sock *tp, 1270 struct tcphdr *th, u64 ack_seq) 1271 { 1272 const struct sock *ssk = (const struct sock *)tp; 1273 u64 rcv_wnd_old, rcv_wnd_new; 1274 u32 new_win; 1275 u64 win; 1276 1277 rcv_wnd_new = ack_seq + tp->rcv_wnd; 1278 1279 rcv_wnd_old = atomic64_read(&msk->rcv_wnd_sent); 1280 if (after64(rcv_wnd_new, rcv_wnd_old)) { 1281 u64 rcv_wnd; 1282 1283 for (;;) { 1284 rcv_wnd = atomic64_cmpxchg(&msk->rcv_wnd_sent, rcv_wnd_old, rcv_wnd_new); 1285 1286 if (rcv_wnd == rcv_wnd_old) 1287 break; 1288 1289 rcv_wnd_old = rcv_wnd; 1290 if (before64(rcv_wnd_new, rcv_wnd_old)) { 1291 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDCONFLICTUPDATE); 1292 goto raise_win; 1293 } 1294 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDCONFLICT); 1295 } 1296 goto update_wspace; 1297 } 1298 1299 if (rcv_wnd_new != rcv_wnd_old) { 1300 raise_win: 1301 /* The msk-level rcv wnd is after the tcp level one, 1302 * sync the latter. 1303 */ 1304 rcv_wnd_new = rcv_wnd_old; 1305 win = rcv_wnd_old - ack_seq; 1306 new_win = min_t(u64, win, U32_MAX); 1307 tp->rcv_wnd = new_win; 1308 tcp_update_max_rcv_wnd_seq(tp); 1309 1310 /* Make sure we do not exceed the maximum possible 1311 * scaled window. 1312 */ 1313 if (unlikely(th->syn)) 1314 new_win = min(new_win, 65535U) << tp->rx_opt.rcv_wscale; 1315 if (!tp->rx_opt.rcv_wscale && 1316 READ_ONCE(sock_net(ssk)->ipv4.sysctl_tcp_workaround_signed_windows)) 1317 new_win = min(new_win, MAX_TCP_WINDOW); 1318 else 1319 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 1320 1321 /* RFC1323 scaling applied */ 1322 new_win >>= tp->rx_opt.rcv_wscale; 1323 th->window = htons(new_win); 1324 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDSHARED); 1325 } 1326 1327 update_wspace: 1328 WRITE_ONCE(msk->old_wspace, tp->rcv_wnd); 1329 return rcv_wnd_new; 1330 } 1331 1332 static void mptcp_track_rwin(struct tcp_sock *tp) 1333 { 1334 const struct sock *ssk = (const struct sock *)tp; 1335 struct mptcp_subflow_context *subflow; 1336 struct mptcp_sock *msk; 1337 1338 if (!ssk) 1339 return; 1340 1341 subflow = mptcp_subflow_ctx(ssk); 1342 msk = mptcp_sk(subflow->conn); 1343 WRITE_ONCE(msk->old_wspace, tp->rcv_wnd); 1344 } 1345 1346 __sum16 __mptcp_make_csum(u64 data_seq, u32 subflow_seq, u16 data_len, __wsum sum) 1347 { 1348 struct csum_pseudo_header header; 1349 __wsum csum; 1350 1351 /* cfr RFC 8684 3.3.1.: 1352 * the data sequence number used in the pseudo-header is 1353 * always the 64-bit value, irrespective of what length is used in the 1354 * DSS option itself. 1355 */ 1356 header.data_seq = cpu_to_be64(data_seq); 1357 header.subflow_seq = htonl(subflow_seq); 1358 header.data_len = htons(data_len); 1359 header.csum = 0; 1360 1361 csum = csum_partial(&header, sizeof(header), sum); 1362 return csum_fold(csum); 1363 } 1364 1365 static __sum16 mptcp_make_csum(const struct mptcp_ext *mpext) 1366 { 1367 return __mptcp_make_csum(mpext->data_seq, mpext->subflow_seq, mpext->data_len, 1368 ~csum_unfold(mpext->csum)); 1369 } 1370 1371 static void put_len_csum(u16 len, __sum16 csum, void *data) 1372 { 1373 __sum16 *sumptr = data + 2; 1374 __be16 *ptr = data; 1375 1376 put_unaligned_be16(len, ptr); 1377 1378 put_unaligned(csum, sumptr); 1379 } 1380 1381 void mptcp_write_options(struct tcphdr *th, __be32 *ptr, struct tcp_sock *tp, 1382 struct mptcp_out_options *opts) 1383 { 1384 const struct sock *ssk = (const struct sock *)tp; 1385 struct mptcp_subflow_context *subflow; 1386 1387 /* Which options can be used together? 1388 * 1389 * X: mutually exclusive 1390 * O: often used together 1391 * C: can be used together in some cases 1392 * P: could be used together but we prefer not to (optimisations) 1393 * 1394 * Opt: | MPC | MPJ | DSS | ADD | RM | PRIO | FAIL | FC | 1395 * ------|------|------|------|------|------|------|------|------| 1396 * MPC |------|------|------|------|------|------|------|------| 1397 * MPJ | X |------|------|------|------|------|------|------| 1398 * DSS | X | X |------|------|------|------|------|------| 1399 * ADD | X | X | P |------|------|------|------|------| 1400 * RM | C | C | C | P |------|------|------|------| 1401 * PRIO | X | C | C | C | C |------|------|------| 1402 * FAIL | X | X | C | X | X | X |------|------| 1403 * FC | X | X | X | X | X | X | X |------| 1404 * RST | X | X | X | X | X | X | O | O | 1405 * ------|------|------|------|------|------|------|------|------| 1406 * 1407 * The same applies in mptcp_established_options() function. 1408 */ 1409 if (likely(OPTION_MPTCP_DSS & opts->suboptions)) { 1410 struct mptcp_ext *mpext = &opts->ext_copy; 1411 u8 len = TCPOLEN_MPTCP_DSS_BASE; 1412 u8 flags = 0; 1413 1414 if (mpext->use_ack) { 1415 flags = MPTCP_DSS_HAS_ACK; 1416 if (mpext->ack64) { 1417 len += TCPOLEN_MPTCP_DSS_ACK64; 1418 flags |= MPTCP_DSS_ACK64; 1419 } else { 1420 len += TCPOLEN_MPTCP_DSS_ACK32; 1421 } 1422 } 1423 1424 if (mpext->use_map) { 1425 len += TCPOLEN_MPTCP_DSS_MAP64; 1426 1427 /* Use only 64-bit mapping flags for now, add 1428 * support for optional 32-bit mappings later. 1429 */ 1430 flags |= MPTCP_DSS_HAS_MAP | MPTCP_DSS_DSN64; 1431 if (mpext->data_fin) 1432 flags |= MPTCP_DSS_DATA_FIN; 1433 1434 if (opts->csum_reqd) 1435 len += TCPOLEN_MPTCP_DSS_CHECKSUM; 1436 } 1437 1438 *ptr++ = mptcp_option(MPTCPOPT_DSS, len, 0, flags); 1439 1440 if (mpext->use_ack) { 1441 struct mptcp_sock *msk; 1442 u64 ack_seq; 1443 1444 /* DSS option is set only by mptcp_established_options, 1445 * the caller is __tcp_transmit_skb() and ssk is always 1446 * not NULL. 1447 */ 1448 subflow = mptcp_subflow_ctx(ssk); 1449 msk = mptcp_sk(subflow->conn); 1450 ack_seq = READ_ONCE(msk->ack_seq); 1451 if (mpext->ack64) { 1452 put_unaligned_be64(ack_seq, ptr); 1453 ptr += 2; 1454 } else { 1455 put_unaligned_be32(ack_seq, ptr); 1456 ptr += 1; 1457 } 1458 subflow->rcv_wnd_sent = mptcp_set_rwin(msk, tp, th, 1459 ack_seq); 1460 } 1461 1462 if (mpext->use_map) { 1463 put_unaligned_be64(mpext->data_seq, ptr); 1464 ptr += 2; 1465 put_unaligned_be32(mpext->subflow_seq, ptr); 1466 ptr += 1; 1467 if (opts->csum_reqd) { 1468 /* data_len == 0 is reserved for the infinite mapping, 1469 * the checksum will also be set to 0. 1470 */ 1471 put_len_csum(mpext->data_len, 1472 (mpext->data_len ? mptcp_make_csum(mpext) : 0), 1473 ptr); 1474 } else { 1475 put_unaligned_be32(mpext->data_len << 16 | 1476 TCPOPT_NOP << 8 | TCPOPT_NOP, ptr); 1477 } 1478 ptr += 1; 1479 } 1480 1481 /* We might need to add MP_FAIL options in rare cases */ 1482 if (unlikely(OPTION_MPTCP_FAIL & opts->suboptions)) 1483 goto mp_fail; 1484 } else if (OPTIONS_MPTCP_MPC & opts->suboptions) { 1485 u8 len, flag = MPTCP_CAP_HMAC_SHA256; 1486 1487 if (OPTION_MPTCP_MPC_SYN & opts->suboptions) { 1488 len = TCPOLEN_MPTCP_MPC_SYN; 1489 } else if (OPTION_MPTCP_MPC_SYNACK & opts->suboptions) { 1490 len = TCPOLEN_MPTCP_MPC_SYNACK; 1491 } else if (opts->data_len) { 1492 len = TCPOLEN_MPTCP_MPC_ACK_DATA; 1493 if (opts->csum_reqd) 1494 len += TCPOLEN_MPTCP_DSS_CHECKSUM; 1495 } else { 1496 len = TCPOLEN_MPTCP_MPC_ACK; 1497 } 1498 1499 if (opts->csum_reqd) 1500 flag |= MPTCP_CAP_CHECKSUM_REQD; 1501 1502 if (!opts->allow_join_id0) 1503 flag |= MPTCP_CAP_DENY_JOIN_ID0; 1504 1505 *ptr++ = mptcp_option(MPTCPOPT_MP_CAPABLE, len, 1506 MPTCP_SUPPORTED_VERSION, 1507 flag); 1508 1509 if (!((OPTION_MPTCP_MPC_SYNACK | OPTION_MPTCP_MPC_ACK) & 1510 opts->suboptions)) 1511 goto mp_capable_done; 1512 1513 put_unaligned_be64(opts->sndr_key, ptr); 1514 ptr += 2; 1515 if (!((OPTION_MPTCP_MPC_ACK) & opts->suboptions)) 1516 goto mp_capable_done; 1517 1518 put_unaligned_be64(opts->rcvr_key, ptr); 1519 ptr += 2; 1520 if (!opts->data_len) 1521 goto mp_capable_done; 1522 1523 if (opts->csum_reqd) { 1524 put_len_csum(opts->data_len, 1525 __mptcp_make_csum(opts->data_seq, 1526 opts->subflow_seq, 1527 opts->data_len, 1528 ~csum_unfold(opts->csum)), 1529 ptr); 1530 } else { 1531 put_unaligned_be32(opts->data_len << 16 | 1532 TCPOPT_NOP << 8 | TCPOPT_NOP, ptr); 1533 } 1534 ptr += 1; 1535 1536 /* MPC is additionally mutually exclusive with MP_PRIO */ 1537 goto mp_capable_done; 1538 } else if (OPTIONS_MPTCP_MPJ & opts->suboptions) { 1539 if (OPTION_MPTCP_MPJ_SYN & opts->suboptions) { 1540 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN, 1541 TCPOLEN_MPTCP_MPJ_SYN, 1542 opts->backup, opts->join_id); 1543 put_unaligned_be32(opts->token, ptr); 1544 ptr += 1; 1545 put_unaligned_be32(opts->nonce, ptr); 1546 ptr += 1; 1547 } else if (OPTION_MPTCP_MPJ_SYNACK & opts->suboptions) { 1548 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN, 1549 TCPOLEN_MPTCP_MPJ_SYNACK, 1550 opts->backup, opts->join_id); 1551 put_unaligned_be64(opts->thmac, ptr); 1552 ptr += 2; 1553 put_unaligned_be32(opts->nonce, ptr); 1554 ptr += 1; 1555 } else { 1556 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN, 1557 TCPOLEN_MPTCP_MPJ_ACK, 0, 0); 1558 memcpy(ptr, opts->hmac, MPTCPOPT_HMAC_LEN); 1559 ptr += 5; 1560 } 1561 } else if (OPTION_MPTCP_ADD_ADDR & opts->suboptions) { 1562 u8 len = TCPOLEN_MPTCP_ADD_ADDR_BASE; 1563 u8 echo = MPTCP_ADDR_ECHO; 1564 1565 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 1566 if (opts->addr.family == AF_INET6) 1567 len = TCPOLEN_MPTCP_ADD_ADDR6_BASE; 1568 #endif 1569 1570 if (opts->addr.port) 1571 len += TCPOLEN_MPTCP_PORT_LEN; 1572 1573 if (opts->ahmac) { 1574 len += sizeof(opts->ahmac); 1575 echo = 0; 1576 } 1577 1578 *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR, 1579 len, echo, opts->addr.id); 1580 if (opts->addr.family == AF_INET) { 1581 memcpy((u8 *)ptr, (u8 *)&opts->addr.addr.s_addr, 4); 1582 ptr += 1; 1583 } 1584 #if IS_ENABLED(CONFIG_MPTCP_IPV6) 1585 else if (opts->addr.family == AF_INET6) { 1586 memcpy((u8 *)ptr, opts->addr.addr6.s6_addr, 16); 1587 ptr += 4; 1588 } 1589 #endif 1590 1591 if (!opts->addr.port) { 1592 if (opts->ahmac) { 1593 put_unaligned_be64(opts->ahmac, ptr); 1594 ptr += 2; 1595 } 1596 } else { 1597 u16 port = ntohs(opts->addr.port); 1598 1599 if (opts->ahmac) { 1600 u8 *bptr = (u8 *)ptr; 1601 1602 put_unaligned_be16(port, bptr); 1603 bptr += 2; 1604 put_unaligned_be64(opts->ahmac, bptr); 1605 bptr += 8; 1606 put_unaligned_be16(TCPOPT_NOP << 8 | 1607 TCPOPT_NOP, bptr); 1608 1609 ptr += 3; 1610 } else { 1611 put_unaligned_be32(port << 16 | 1612 TCPOPT_NOP << 8 | 1613 TCPOPT_NOP, ptr); 1614 ptr += 1; 1615 } 1616 } 1617 } else if (unlikely(OPTION_MPTCP_FASTCLOSE & opts->suboptions)) { 1618 /* FASTCLOSE is mutually exclusive with others except RST */ 1619 *ptr++ = mptcp_option(MPTCPOPT_MP_FASTCLOSE, 1620 TCPOLEN_MPTCP_FASTCLOSE, 1621 0, 0); 1622 put_unaligned_be64(opts->rcvr_key, ptr); 1623 ptr += 2; 1624 1625 if (OPTION_MPTCP_RST & opts->suboptions) 1626 goto mp_rst; 1627 return; 1628 } else if (unlikely(OPTION_MPTCP_FAIL & opts->suboptions)) { 1629 mp_fail: 1630 /* MP_FAIL is mutually exclusive with others except RST */ 1631 subflow = mptcp_subflow_ctx(ssk); 1632 subflow->send_mp_fail = 0; 1633 1634 *ptr++ = mptcp_option(MPTCPOPT_MP_FAIL, 1635 TCPOLEN_MPTCP_FAIL, 1636 0, 0); 1637 put_unaligned_be64(opts->fail_seq, ptr); 1638 ptr += 2; 1639 1640 if (OPTION_MPTCP_RST & opts->suboptions) 1641 goto mp_rst; 1642 return; 1643 } else if (unlikely(OPTION_MPTCP_RST & opts->suboptions)) { 1644 mp_rst: 1645 *ptr++ = mptcp_option(MPTCPOPT_RST, 1646 TCPOLEN_MPTCP_RST, 1647 opts->reset_transient, 1648 opts->reset_reason); 1649 return; 1650 } else if (unlikely(!opts->suboptions)) { 1651 /* Fallback to TCP */ 1652 mptcp_track_rwin(tp); 1653 return; 1654 } 1655 1656 if (OPTION_MPTCP_PRIO & opts->suboptions) { 1657 subflow = mptcp_subflow_ctx(ssk); 1658 subflow->send_mp_prio = 0; 1659 1660 *ptr++ = mptcp_option(MPTCPOPT_MP_PRIO, 1661 TCPOLEN_MPTCP_PRIO, 1662 opts->backup, TCPOPT_NOP); 1663 1664 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPPRIOTX); 1665 } 1666 1667 mp_capable_done: 1668 if (OPTION_MPTCP_RM_ADDR & opts->suboptions) { 1669 u8 i = 1; 1670 1671 *ptr++ = mptcp_option(MPTCPOPT_RM_ADDR, 1672 TCPOLEN_MPTCP_RM_ADDR_BASE + opts->rm_list.nr, 1673 0, opts->rm_list.ids[0]); 1674 1675 while (i < opts->rm_list.nr) { 1676 u8 id1, id2, id3, id4; 1677 1678 id1 = opts->rm_list.ids[i]; 1679 id2 = i + 1 < opts->rm_list.nr ? opts->rm_list.ids[i + 1] : TCPOPT_NOP; 1680 id3 = i + 2 < opts->rm_list.nr ? opts->rm_list.ids[i + 2] : TCPOPT_NOP; 1681 id4 = i + 3 < opts->rm_list.nr ? opts->rm_list.ids[i + 3] : TCPOPT_NOP; 1682 put_unaligned_be32(id1 << 24 | id2 << 16 | id3 << 8 | id4, ptr); 1683 ptr += 1; 1684 i += 4; 1685 } 1686 } 1687 } 1688 1689 __be32 mptcp_get_reset_option(const struct sk_buff *skb) 1690 { 1691 const struct mptcp_ext *ext = mptcp_get_ext(skb); 1692 u8 flags, reason; 1693 1694 if (ext) { 1695 flags = ext->reset_transient; 1696 reason = ext->reset_reason; 1697 1698 return mptcp_option(MPTCPOPT_RST, TCPOLEN_MPTCP_RST, 1699 flags, reason); 1700 } 1701 1702 return htonl(0u); 1703 } 1704 EXPORT_SYMBOL_GPL(mptcp_get_reset_option); 1705