net/mptcp/options.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Multipath TCP
   3  *
   4  * Copyright (c) 2017 - 2019, Intel Corporation.
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <net/tcp.h>
   9 #include <net/mptcp.h>
  10 #include "protocol.h"
  11
  12 static bool mptcp_cap_flag_sha256(u8 flags)
  13 {
  14         return (flags & MPTCP_CAP_FLAG_MASK) == MPTCP_CAP_HMAC_SHA256;
  15 }
  16
  17 void mptcp_parse_option(const struct sk_buff *skb, const unsigned char *ptr,
  18                         int opsize, struct tcp_options_received *opt_rx)
  19 {
  20         struct mptcp_options_received *mp_opt = &opt_rx->mptcp;
  21         u8 subtype = *ptr >> 4;
  22         int expected_opsize;
  23         u8 version;
  24         u8 flags;
  25
  26         switch (subtype) {
  27         case MPTCPOPT_MP_CAPABLE:
  28                 /* strict size checking */
  29                 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
  30                         if (skb->len > tcp_hdr(skb)->doff << 2)
  31                                 expected_opsize = TCPOLEN_MPTCP_MPC_ACK_DATA;
  32                         else
  33                                 expected_opsize = TCPOLEN_MPTCP_MPC_ACK;
  34                 } else {
  35                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)
  36                                 expected_opsize = TCPOLEN_MPTCP_MPC_SYNACK;
  37                         else
  38                                 expected_opsize = TCPOLEN_MPTCP_MPC_SYN;
  39                 }
  40                 if (opsize != expected_opsize)
  41                         break;
  42
  43                 /* try to be gentle vs future versions on the initial syn */
  44                 version = *ptr++ & MPTCP_VERSION_MASK;
  45                 if (opsize != TCPOLEN_MPTCP_MPC_SYN) {
  46                         if (version != MPTCP_SUPPORTED_VERSION)
  47                                 break;
  48                 } else if (version < MPTCP_SUPPORTED_VERSION) {
  49                         break;
  50                 }
  51
  52                 flags = *ptr++;
  53                 if (!mptcp_cap_flag_sha256(flags) ||
  54                     (flags & MPTCP_CAP_EXTENSIBILITY))
  55                         break;
  56
  57                 /* RFC 6824, Section 3.1:
  58                  * "For the Checksum Required bit (labeled "A"), if either
  59                  * host requires the use of checksums, checksums MUST be used.
  60                  * In other words, the only way for checksums not to be used
  61                  * is if both hosts in their SYNs set A=0."
  62                  *
  63                  * Section 3.3.0:
  64                  * "If a checksum is not present when its use has been
  65                  * negotiated, the receiver MUST close the subflow with a RST as
  66                  * it is considered broken."
  67                  *
  68                  * We don't implement DSS checksum - fall back to TCP.
  69                  */
  70                 if (flags & MPTCP_CAP_CHECKSUM_REQD)
  71                         break;
  72
  73                 mp_opt->mp_capable = 1;
  74                 if (opsize >= TCPOLEN_MPTCP_MPC_SYNACK) {
  75                         mp_opt->sndr_key = get_unaligned_be64(ptr);
  76                         ptr += 8;
  77                 }
  78                 if (opsize >= TCPOLEN_MPTCP_MPC_ACK) {
  79                         mp_opt->rcvr_key = get_unaligned_be64(ptr);
  80                         ptr += 8;
  81                 }
  82                 if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA) {
  83                         /* Section 3.1.:
  84                          * "the data parameters in a MP_CAPABLE are semantically
  85                          * equivalent to those in a DSS option and can be used
  86                          * interchangeably."
  87                          */
  88                         mp_opt->dss = 1;
  89                         mp_opt->use_map = 1;
  90                         mp_opt->mpc_map = 1;
  91                         mp_opt->data_len = get_unaligned_be16(ptr);
  92                         ptr += 2;
  93                 }
  94                 pr_debug("MP_CAPABLE version=%x, flags=%x, optlen=%d sndr=%llu, rcvr=%llu len=%d",
  95                          version, flags, opsize, mp_opt->sndr_key,
  96                          mp_opt->rcvr_key, mp_opt->data_len);
  97                 break;
  98
  99         case MPTCPOPT_MP_JOIN:
 100                 mp_opt->mp_join = 1;
 101                 if (opsize == TCPOLEN_MPTCP_MPJ_SYN) {
 102                         mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP;
 103                         mp_opt->join_id = *ptr++;
 104                         mp_opt->token = get_unaligned_be32(ptr);
 105                         ptr += 4;
 106                         mp_opt->nonce = get_unaligned_be32(ptr);
 107                         ptr += 4;
 108                         pr_debug("MP_JOIN bkup=%u, id=%u, token=%u, nonce=%u",
 109                                  mp_opt->backup, mp_opt->join_id,
 110                                  mp_opt->token, mp_opt->nonce);
 111                 } else if (opsize == TCPOLEN_MPTCP_MPJ_SYNACK) {
 112                         mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP;
 113                         mp_opt->join_id = *ptr++;
 114                         mp_opt->thmac = get_unaligned_be64(ptr);
 115                         ptr += 8;
 116                         mp_opt->nonce = get_unaligned_be32(ptr);
 117                         ptr += 4;
 118                         pr_debug("MP_JOIN bkup=%u, id=%u, thmac=%llu, nonce=%u",
 119                                  mp_opt->backup, mp_opt->join_id,
 120                                  mp_opt->thmac, mp_opt->nonce);
 121                 } else if (opsize == TCPOLEN_MPTCP_MPJ_ACK) {
 122                         ptr += 2;
 123                         memcpy(mp_opt->hmac, ptr, MPTCPOPT_HMAC_LEN);
 124                         pr_debug("MP_JOIN hmac");
 125                 } else {
 126                         pr_warn("MP_JOIN bad option size");
 127                         mp_opt->mp_join = 0;
 128                 }
 129                 break;
 130
 131         case MPTCPOPT_DSS:
 132                 pr_debug("DSS");
 133                 ptr++;
 134
 135                 /* we must clear 'mpc_map' be able to detect MP_CAPABLE
 136                  * map vs DSS map in mptcp_incoming_options(), and reconstruct
 137                  * map info accordingly
 138                  */
 139                 mp_opt->mpc_map = 0;
 140                 flags = (*ptr++) & MPTCP_DSS_FLAG_MASK;
 141                 mp_opt->data_fin = (flags & MPTCP_DSS_DATA_FIN) != 0;
 142                 mp_opt->dsn64 = (flags & MPTCP_DSS_DSN64) != 0;
 143                 mp_opt->use_map = (flags & MPTCP_DSS_HAS_MAP) != 0;
 144                 mp_opt->ack64 = (flags & MPTCP_DSS_ACK64) != 0;
 145                 mp_opt->use_ack = (flags & MPTCP_DSS_HAS_ACK);
 146
 147                 pr_debug("data_fin=%d dsn64=%d use_map=%d ack64=%d use_ack=%d",
 148                          mp_opt->data_fin, mp_opt->dsn64,
 149                          mp_opt->use_map, mp_opt->ack64,
 150                          mp_opt->use_ack);
 151
 152                 expected_opsize = TCPOLEN_MPTCP_DSS_BASE;
 153
 154                 if (mp_opt->use_ack) {
 155                         if (mp_opt->ack64)
 156                                 expected_opsize += TCPOLEN_MPTCP_DSS_ACK64;
 157                         else
 158                                 expected_opsize += TCPOLEN_MPTCP_DSS_ACK32;
 159                 }
 160
 161                 if (mp_opt->use_map) {
 162                         if (mp_opt->dsn64)
 163                                 expected_opsize += TCPOLEN_MPTCP_DSS_MAP64;
 164                         else
 165                                 expected_opsize += TCPOLEN_MPTCP_DSS_MAP32;
 166                 }
 167
 168                 /* RFC 6824, Section 3.3:
 169                  * If a checksum is present, but its use had
 170                  * not been negotiated in the MP_CAPABLE handshake,
 171                  * the checksum field MUST be ignored.
 172                  */
 173                 if (opsize != expected_opsize &&
 174                     opsize != expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM)
 175                         break;
 176
 177                 mp_opt->dss = 1;
 178
 179                 if (mp_opt->use_ack) {
 180                         if (mp_opt->ack64) {
 181                                 mp_opt->data_ack = get_unaligned_be64(ptr);
 182                                 ptr += 8;
 183                         } else {
 184                                 mp_opt->data_ack = get_unaligned_be32(ptr);
 185                                 ptr += 4;
 186                         }
 187
 188                         pr_debug("data_ack=%llu", mp_opt->data_ack);
 189                 }
 190
 191                 if (mp_opt->use_map) {
 192                         if (mp_opt->dsn64) {
 193                                 mp_opt->data_seq = get_unaligned_be64(ptr);
 194                                 ptr += 8;
 195                         } else {
 196                                 mp_opt->data_seq = get_unaligned_be32(ptr);
 197                                 ptr += 4;
 198                         }
 199
 200                         mp_opt->subflow_seq = get_unaligned_be32(ptr);
 201                         ptr += 4;
 202
 203                         mp_opt->data_len = get_unaligned_be16(ptr);
 204                         ptr += 2;
 205
 206                         pr_debug("data_seq=%llu subflow_seq=%u data_len=%u",
 207                                  mp_opt->data_seq, mp_opt->subflow_seq,
 208                                  mp_opt->data_len);
 209                 }
 210
 211                 break;
 212
 213         case MPTCPOPT_ADD_ADDR:
 214                 mp_opt->echo = (*ptr++) & MPTCP_ADDR_ECHO;
 215                 if (!mp_opt->echo) {
 216                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR ||
 217                             opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT)
 218                                 mp_opt->family = MPTCP_ADDR_IPVERSION_4;
 219 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 220                         else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6 ||
 221                                  opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT)
 222                                 mp_opt->family = MPTCP_ADDR_IPVERSION_6;
 223 #endif
 224                         else
 225                                 break;
 226                 } else {
 227                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE ||
 228                             opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT)
 229                                 mp_opt->family = MPTCP_ADDR_IPVERSION_4;
 230 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 231                         else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE ||
 232                                  opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT)
 233                                 mp_opt->family = MPTCP_ADDR_IPVERSION_6;
 234 #endif
 235                         else
 236                                 break;
 237                 }
 238
 239                 mp_opt->add_addr = 1;
 240                 mp_opt->port = 0;
 241                 mp_opt->addr_id = *ptr++;
 242                 pr_debug("ADD_ADDR: id=%d", mp_opt->addr_id);
 243                 if (mp_opt->family == MPTCP_ADDR_IPVERSION_4) {
 244                         memcpy((u8 *)&mp_opt->addr.s_addr, (u8 *)ptr, 4);
 245                         ptr += 4;
 246                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT ||
 247                             opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) {
 248                                 mp_opt->port = get_unaligned_be16(ptr);
 249                                 ptr += 2;
 250                         }
 251                 }
 252 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 253                 else {
 254                         memcpy(mp_opt->addr6.s6_addr, (u8 *)ptr, 16);
 255                         ptr += 16;
 256                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT ||
 257                             opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) {
 258                                 mp_opt->port = get_unaligned_be16(ptr);
 259                                 ptr += 2;
 260                         }
 261                 }
 262 #endif
 263                 if (!mp_opt->echo) {
 264                         mp_opt->ahmac = get_unaligned_be64(ptr);
 265                         ptr += 8;
 266                 }
 267                 break;
 268
 269         case MPTCPOPT_RM_ADDR:
 270                 if (opsize != TCPOLEN_MPTCP_RM_ADDR_BASE)
 271                         break;
 272
 273                 mp_opt->rm_addr = 1;
 274                 mp_opt->rm_id = *ptr++;
 275                 pr_debug("RM_ADDR: id=%d", mp_opt->rm_id);
 276                 break;
 277
 278         default:
 279                 break;
 280         }
 281 }
 282
 283 void mptcp_get_options(const struct sk_buff *skb,
 284                        struct tcp_options_received *opt_rx)
 285 {
 286         const unsigned char *ptr;
 287         const struct tcphdr *th = tcp_hdr(skb);
 288         int length = (th->doff * 4) - sizeof(struct tcphdr);
 289
 290         ptr = (const unsigned char *)(th + 1);
 291
 292         while (length > 0) {
 293                 int opcode = *ptr++;
 294                 int opsize;
 295
 296                 switch (opcode) {
 297                 case TCPOPT_EOL:
 298                         return;
 299                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
 300                         length--;
 301                         continue;
 302                 default:
 303                         opsize = *ptr++;
 304                         if (opsize < 2) /* "silly options" */
 305                                 return;
 306                         if (opsize > length)
 307                                 return; /* don't parse partial options */
 308                         if (opcode == TCPOPT_MPTCP)
 309                                 mptcp_parse_option(skb, ptr, opsize, opt_rx);
 310                         ptr += opsize - 2;
 311                         length -= opsize;
 312                 }
 313         }
 314 }
 315
 316 bool mptcp_syn_options(struct sock *sk, const struct sk_buff *skb,
 317                        unsigned int *size, struct mptcp_out_options *opts)
 318 {
 319         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 320
 321         /* we will use snd_isn to detect first pkt [re]transmission
 322          * in mptcp_established_options_mp()
 323          */
 324         subflow->snd_isn = TCP_SKB_CB(skb)->end_seq;
 325         if (subflow->request_mptcp) {
 326                 pr_debug("local_key=%llu", subflow->local_key);
 327                 opts->suboptions = OPTION_MPTCP_MPC_SYN;
 328                 opts->sndr_key = subflow->local_key;
 329                 *size = TCPOLEN_MPTCP_MPC_SYN;
 330                 return true;
 331         }
 332         return false;
 333 }
 334
 335 void mptcp_rcv_synsent(struct sock *sk)
 336 {
 337         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 338         struct tcp_sock *tp = tcp_sk(sk);
 339
 340         pr_debug("subflow=%p", subflow);
 341         if (subflow->request_mptcp && tp->rx_opt.mptcp.mp_capable) {
 342                 subflow->mp_capable = 1;
 343                 subflow->can_ack = 1;
 344                 subflow->remote_key = tp->rx_opt.mptcp.sndr_key;
 345         } else {
 346                 tcp_sk(sk)->is_mptcp = 0;
 347         }
 348 }
 349
 350 static bool mptcp_established_options_mp(struct sock *sk, struct sk_buff *skb,
 351                                          unsigned int *size,
 352                                          unsigned int remaining,
 353                                          struct mptcp_out_options *opts)
 354 {
 355         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 356         struct mptcp_ext *mpext;
 357         unsigned int data_len;
 358
 359         pr_debug("subflow=%p fully established=%d seq=%x:%x remaining=%d",
 360                  subflow, subflow->fully_established, subflow->snd_isn,
 361                  skb ? TCP_SKB_CB(skb)->seq : 0, remaining);
 362
 363         if (subflow->mp_capable && !subflow->fully_established && skb &&
 364             subflow->snd_isn == TCP_SKB_CB(skb)->seq) {
 365                 /* When skb is not available, we better over-estimate the
 366                  * emitted options len. A full DSS option is longer than
 367                  * TCPOLEN_MPTCP_MPC_ACK_DATA, so let's the caller try to fit
 368                  * that.
 369                  */
 370                 mpext = mptcp_get_ext(skb);
 371                 data_len = mpext ? mpext->data_len : 0;
 372
 373                 /* we will check ext_copy.data_len in mptcp_write_options() to
 374                  * discriminate between TCPOLEN_MPTCP_MPC_ACK_DATA and
 375                  * TCPOLEN_MPTCP_MPC_ACK
 376                  */
 377                 opts->ext_copy.data_len = data_len;
 378                 opts->suboptions = OPTION_MPTCP_MPC_ACK;
 379                 opts->sndr_key = subflow->local_key;
 380                 opts->rcvr_key = subflow->remote_key;
 381
 382                 /* Section 3.1.
 383                  * The MP_CAPABLE option is carried on the SYN, SYN/ACK, and ACK
 384                  * packets that start the first subflow of an MPTCP connection,
 385                  * as well as the first packet that carries data
 386                  */
 387                 if (data_len > 0)
 388                         *size = ALIGN(TCPOLEN_MPTCP_MPC_ACK_DATA, 4);
 389                 else
 390                         *size = TCPOLEN_MPTCP_MPC_ACK;
 391
 392                 pr_debug("subflow=%p, local_key=%llu, remote_key=%llu map_len=%d",
 393                          subflow, subflow->local_key, subflow->remote_key,
 394                          data_len);
 395
 396                 return true;
 397         }
 398         return false;
 399 }
 400
 401 static void mptcp_write_data_fin(struct mptcp_subflow_context *subflow,
 402                                  struct mptcp_ext *ext)
 403 {
 404         if (!ext->use_map) {
 405                 /* RFC6824 requires a DSS mapping with specific values
 406                  * if DATA_FIN is set but no data payload is mapped
 407                  */
 408                 ext->data_fin = 1;
 409                 ext->use_map = 1;
 410                 ext->dsn64 = 1;
 411                 ext->data_seq = subflow->data_fin_tx_seq;
 412                 ext->subflow_seq = 0;
 413                 ext->data_len = 1;
 414         } else if (ext->data_seq + ext->data_len == subflow->data_fin_tx_seq) {
 415                 /* If there's an existing DSS mapping and it is the
 416                  * final mapping, DATA_FIN consumes 1 additional byte of
 417                  * mapping space.
 418                  */
 419                 ext->data_fin = 1;
 420                 ext->data_len++;
 421         }
 422 }
 423
 424 static bool mptcp_established_options_dss(struct sock *sk, struct sk_buff *skb,
 425                                           unsigned int *size,
 426                                           unsigned int remaining,
 427                                           struct mptcp_out_options *opts)
 428 {
 429         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 430         unsigned int dss_size = 0;
 431         struct mptcp_ext *mpext;
 432         struct mptcp_sock *msk;
 433         unsigned int ack_size;
 434         bool ret = false;
 435         u8 tcp_fin;
 436
 437         if (skb) {
 438                 mpext = mptcp_get_ext(skb);
 439                 tcp_fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
 440         } else {
 441                 mpext = NULL;
 442                 tcp_fin = 0;
 443         }
 444
 445         if (!skb || (mpext && mpext->use_map) || tcp_fin) {
 446                 unsigned int map_size;
 447
 448                 map_size = TCPOLEN_MPTCP_DSS_BASE + TCPOLEN_MPTCP_DSS_MAP64;
 449
 450                 remaining -= map_size;
 451                 dss_size = map_size;
 452                 if (mpext)
 453                         opts->ext_copy = *mpext;
 454
 455                 if (skb && tcp_fin && subflow->data_fin_tx_enable)
 456                         mptcp_write_data_fin(subflow, &opts->ext_copy);
 457                 ret = true;
 458         }
 459
 460         /* passive sockets msk will set the 'can_ack' after accept(), even
 461          * if the first subflow may have the already the remote key handy
 462          */
 463         opts->ext_copy.use_ack = 0;
 464         msk = mptcp_sk(subflow->conn);
 465         if (!READ_ONCE(msk->can_ack)) {
 466                 *size = ALIGN(dss_size, 4);
 467                 return ret;
 468         }
 469
 470         ack_size = TCPOLEN_MPTCP_DSS_ACK64;
 471
 472         /* Add kind/length/subtype/flag overhead if mapping is not populated */
 473         if (dss_size == 0)
 474                 ack_size += TCPOLEN_MPTCP_DSS_BASE;
 475
 476         dss_size += ack_size;
 477
 478         opts->ext_copy.data_ack = msk->ack_seq;
 479         opts->ext_copy.ack64 = 1;
 480         opts->ext_copy.use_ack = 1;
 481
 482         *size = ALIGN(dss_size, 4);
 483         return true;
 484 }
 485
 486 static u64 add_addr_generate_hmac(u64 key1, u64 key2, u8 addr_id,
 487                                   struct in_addr *addr)
 488 {
 489         u8 hmac[MPTCP_ADDR_HMAC_LEN];
 490         u8 msg[7];
 491
 492         msg[0] = addr_id;
 493         memcpy(&msg[1], &addr->s_addr, 4);
 494         msg[5] = 0;
 495         msg[6] = 0;
 496
 497         mptcp_crypto_hmac_sha(key1, key2, msg, 7, hmac);
 498
 499         return get_unaligned_be64(hmac);
 500 }
 501
 502 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 503 static u64 add_addr6_generate_hmac(u64 key1, u64 key2, u8 addr_id,
 504                                    struct in6_addr *addr)
 505 {
 506         u8 hmac[MPTCP_ADDR_HMAC_LEN];
 507         u8 msg[19];
 508
 509         msg[0] = addr_id;
 510         memcpy(&msg[1], &addr->s6_addr, 16);
 511         msg[17] = 0;
 512         msg[18] = 0;
 513
 514         mptcp_crypto_hmac_sha(key1, key2, msg, 19, hmac);
 515
 516         return get_unaligned_be64(hmac);
 517 }
 518 #endif
 519
 520 static bool mptcp_established_options_addr(struct sock *sk,
 521                                            unsigned int *size,
 522                                            unsigned int remaining,
 523                                            struct mptcp_out_options *opts)
 524 {
 525         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 526         struct mptcp_sock *msk = mptcp_sk(subflow->conn);
 527         struct mptcp_addr_info saddr;
 528         int len;
 529
 530         if (!mptcp_pm_should_signal(msk) ||
 531             !(mptcp_pm_addr_signal(msk, remaining, &saddr)))
 532                 return false;
 533
 534         len = mptcp_add_addr_len(saddr.family);
 535         if (remaining < len)
 536                 return false;
 537
 538         *size = len;
 539         opts->addr_id = saddr.id;
 540         if (saddr.family == AF_INET) {
 541                 opts->suboptions |= OPTION_MPTCP_ADD_ADDR;
 542                 opts->addr = saddr.addr;
 543                 opts->ahmac = add_addr_generate_hmac(msk->local_key,
 544                                                      msk->remote_key,
 545                                                      opts->addr_id,
 546                                                      &opts->addr);
 547         }
 548 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 549         else if (saddr.family == AF_INET6) {
 550                 opts->suboptions |= OPTION_MPTCP_ADD_ADDR6;
 551                 opts->addr6 = saddr.addr6;
 552                 opts->ahmac = add_addr6_generate_hmac(msk->local_key,
 553                                                       msk->remote_key,
 554                                                       opts->addr_id,
 555                                                       &opts->addr6);
 556         }
 557 #endif
 558         pr_debug("addr_id=%d, ahmac=%llu", opts->addr_id, opts->ahmac);
 559
 560         return true;
 561 }
 562
 563 bool mptcp_established_options(struct sock *sk, struct sk_buff *skb,
 564                                unsigned int *size, unsigned int remaining,
 565                                struct mptcp_out_options *opts)
 566 {
 567         unsigned int opt_size = 0;
 568         bool ret = false;
 569
 570         opts->suboptions = 0;
 571
 572         if (mptcp_established_options_mp(sk, skb, &opt_size, remaining, opts))
 573                 ret = true;
 574         else if (mptcp_established_options_dss(sk, skb, &opt_size, remaining,
 575                                                opts))
 576                 ret = true;
 577
 578         /* we reserved enough space for the above options, and exceeding the
 579          * TCP option space would be fatal
 580          */
 581         if (WARN_ON_ONCE(opt_size > remaining))
 582                 return false;
 583
 584         *size += opt_size;
 585         remaining -= opt_size;
 586         if (mptcp_established_options_addr(sk, &opt_size, remaining, opts)) {
 587                 *size += opt_size;
 588                 remaining -= opt_size;
 589                 ret = true;
 590         }
 591
 592         return ret;
 593 }
 594
 595 bool mptcp_synack_options(const struct request_sock *req, unsigned int *size,
 596                           struct mptcp_out_options *opts)
 597 {
 598         struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
 599
 600         if (subflow_req->mp_capable) {
 601                 opts->suboptions = OPTION_MPTCP_MPC_SYNACK;
 602                 opts->sndr_key = subflow_req->local_key;
 603                 *size = TCPOLEN_MPTCP_MPC_SYNACK;
 604                 pr_debug("subflow_req=%p, local_key=%llu",
 605                          subflow_req, subflow_req->local_key);
 606                 return true;
 607         } else if (subflow_req->mp_join) {
 608                 opts->suboptions = OPTION_MPTCP_MPJ_SYNACK;
 609                 opts->backup = subflow_req->backup;
 610                 opts->join_id = subflow_req->local_id;
 611                 opts->thmac = subflow_req->thmac;
 612                 opts->nonce = subflow_req->local_nonce;
 613                 pr_debug("req=%p, bkup=%u, id=%u, thmac=%llu, nonce=%u",
 614                          subflow_req, opts->backup, opts->join_id,
 615                          opts->thmac, opts->nonce);
 616                 *size = TCPOLEN_MPTCP_MPJ_SYNACK;
 617                 return true;
 618         }
 619         return false;
 620 }
 621
 622 static bool check_fully_established(struct mptcp_sock *msk, struct sock *sk,
 623                                     struct mptcp_subflow_context *subflow,
 624                                     struct sk_buff *skb,
 625                                     struct mptcp_options_received *mp_opt)
 626 {
 627         /* here we can process OoO, in-window pkts, only in-sequence 4th ack
 628          * will make the subflow fully established
 629          */
 630         if (likely(subflow->fully_established)) {
 631                 /* on passive sockets, check for 3rd ack retransmission
 632                  * note that msk is always set by subflow_syn_recv_sock()
 633                  * for mp_join subflows
 634                  */
 635                 if (TCP_SKB_CB(skb)->seq == subflow->ssn_offset + 1 &&
 636                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq &&
 637                     subflow->mp_join && mp_opt->mp_join &&
 638                     READ_ONCE(msk->pm.server_side))
 639                         tcp_send_ack(sk);
 640                 goto fully_established;
 641         }
 642
 643         /* we should process OoO packets before the first subflow is fully
 644          * established, but not expected for MP_JOIN subflows
 645          */
 646         if (TCP_SKB_CB(skb)->seq != subflow->ssn_offset + 1)
 647                 return subflow->mp_capable;
 648
 649         if (mp_opt->use_ack) {
 650                 /* subflows are fully established as soon as we get any
 651                  * additional ack.
 652                  */
 653                 subflow->fully_established = 1;
 654                 goto fully_established;
 655         }
 656
 657         WARN_ON_ONCE(subflow->can_ack);
 658
 659         /* If the first established packet does not contain MP_CAPABLE + data
 660          * then fallback to TCP
 661          */
 662         if (!mp_opt->mp_capable) {
 663                 subflow->mp_capable = 0;
 664                 tcp_sk(sk)->is_mptcp = 0;
 665                 return false;
 666         }
 667
 668         subflow->fully_established = 1;
 669         subflow->remote_key = mp_opt->sndr_key;
 670         subflow->can_ack = 1;
 671
 672 fully_established:
 673         if (likely(subflow->pm_notified))
 674                 return true;
 675
 676         subflow->pm_notified = 1;
 677         if (subflow->mp_join)
 678                 mptcp_pm_subflow_established(msk, subflow);
 679         else
 680                 mptcp_pm_fully_established(msk);
 681         return true;
 682 }
 683
 684 static bool add_addr_hmac_valid(struct mptcp_sock *msk,
 685                                 struct mptcp_options_received *mp_opt)
 686 {
 687         u64 hmac = 0;
 688
 689         if (mp_opt->echo)
 690                 return true;
 691
 692         if (mp_opt->family == MPTCP_ADDR_IPVERSION_4)
 693                 hmac = add_addr_generate_hmac(msk->remote_key,
 694                                               msk->local_key,
 695                                               mp_opt->addr_id, &mp_opt->addr);
 696 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 697         else
 698                 hmac = add_addr6_generate_hmac(msk->remote_key,
 699                                                msk->local_key,
 700                                                mp_opt->addr_id, &mp_opt->addr6);
 701 #endif
 702
 703         pr_debug("msk=%p, ahmac=%llu, mp_opt->ahmac=%llu\n",
 704                  msk, (unsigned long long)hmac,
 705                  (unsigned long long)mp_opt->ahmac);
 706
 707         return hmac == mp_opt->ahmac;
 708 }
 709
 710 void mptcp_incoming_options(struct sock *sk, struct sk_buff *skb,
 711                             struct tcp_options_received *opt_rx)
 712 {
 713         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 714         struct mptcp_sock *msk = mptcp_sk(subflow->conn);
 715         struct mptcp_options_received *mp_opt;
 716         struct mptcp_ext *mpext;
 717
 718         mp_opt = &opt_rx->mptcp;
 719         if (!check_fully_established(msk, sk, subflow, skb, mp_opt))
 720                 return;
 721
 722         if (mp_opt->add_addr && add_addr_hmac_valid(msk, mp_opt)) {
 723                 struct mptcp_addr_info addr;
 724
 725                 addr.port = htons(mp_opt->port);
 726                 addr.id = mp_opt->addr_id;
 727                 if (mp_opt->family == MPTCP_ADDR_IPVERSION_4) {
 728                         addr.family = AF_INET;
 729                         addr.addr = mp_opt->addr;
 730                 }
 731 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 732                 else if (mp_opt->family == MPTCP_ADDR_IPVERSION_6) {
 733                         addr.family = AF_INET6;
 734                         addr.addr6 = mp_opt->addr6;
 735                 }
 736 #endif
 737                 if (!mp_opt->echo)
 738                         mptcp_pm_add_addr_received(msk, &addr);
 739                 mp_opt->add_addr = 0;
 740         }
 741
 742         if (!mp_opt->dss)
 743                 return;
 744
 745         mpext = skb_ext_add(skb, SKB_EXT_MPTCP);
 746         if (!mpext)
 747                 return;
 748
 749         memset(mpext, 0, sizeof(*mpext));
 750
 751         if (mp_opt->use_map) {
 752                 if (mp_opt->mpc_map) {
 753                         /* this is an MP_CAPABLE carrying MPTCP data
 754                          * we know this map the first chunk of data
 755                          */
 756                         mptcp_crypto_key_sha(subflow->remote_key, NULL,
 757                                              &mpext->data_seq);
 758                         mpext->data_seq++;
 759                         mpext->subflow_seq = 1;
 760                         mpext->dsn64 = 1;
 761                         mpext->mpc_map = 1;
 762                 } else {
 763                         mpext->data_seq = mp_opt->data_seq;
 764                         mpext->subflow_seq = mp_opt->subflow_seq;
 765                         mpext->dsn64 = mp_opt->dsn64;
 766                 }
 767                 mpext->data_len = mp_opt->data_len;
 768                 mpext->use_map = 1;
 769         }
 770
 771         if (mp_opt->use_ack) {
 772                 mpext->data_ack = mp_opt->data_ack;
 773                 mpext->use_ack = 1;
 774                 mpext->ack64 = mp_opt->ack64;
 775         }
 776
 777         mpext->data_fin = mp_opt->data_fin;
 778 }
 779
 780 void mptcp_write_options(__be32 *ptr, struct mptcp_out_options *opts)
 781 {
 782         if ((OPTION_MPTCP_MPC_SYN | OPTION_MPTCP_MPC_SYNACK |
 783              OPTION_MPTCP_MPC_ACK) & opts->suboptions) {
 784                 u8 len;
 785
 786                 if (OPTION_MPTCP_MPC_SYN & opts->suboptions)
 787                         len = TCPOLEN_MPTCP_MPC_SYN;
 788                 else if (OPTION_MPTCP_MPC_SYNACK & opts->suboptions)
 789                         len = TCPOLEN_MPTCP_MPC_SYNACK;
 790                 else if (opts->ext_copy.data_len)
 791                         len = TCPOLEN_MPTCP_MPC_ACK_DATA;
 792                 else
 793                         len = TCPOLEN_MPTCP_MPC_ACK;
 794
 795                 *ptr++ = mptcp_option(MPTCPOPT_MP_CAPABLE, len,
 796                                       MPTCP_SUPPORTED_VERSION,
 797                                       MPTCP_CAP_HMAC_SHA256);
 798
 799                 if (!((OPTION_MPTCP_MPC_SYNACK | OPTION_MPTCP_MPC_ACK) &
 800                     opts->suboptions))
 801                         goto mp_capable_done;
 802
 803                 put_unaligned_be64(opts->sndr_key, ptr);
 804                 ptr += 2;
 805                 if (!((OPTION_MPTCP_MPC_ACK) & opts->suboptions))
 806                         goto mp_capable_done;
 807
 808                 put_unaligned_be64(opts->rcvr_key, ptr);
 809                 ptr += 2;
 810                 if (!opts->ext_copy.data_len)
 811                         goto mp_capable_done;
 812
 813                 put_unaligned_be32(opts->ext_copy.data_len << 16 |
 814                                    TCPOPT_NOP << 8 | TCPOPT_NOP, ptr);
 815                 ptr += 1;
 816         }
 817
 818 mp_capable_done:
 819         if (OPTION_MPTCP_ADD_ADDR & opts->suboptions) {
 820                 if (opts->ahmac)
 821                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 822                                               TCPOLEN_MPTCP_ADD_ADDR, 0,
 823                                               opts->addr_id);
 824                 else
 825                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 826                                               TCPOLEN_MPTCP_ADD_ADDR_BASE,
 827                                               MPTCP_ADDR_ECHO,
 828                                               opts->addr_id);
 829                 memcpy((u8 *)ptr, (u8 *)&opts->addr.s_addr, 4);
 830                 ptr += 1;
 831                 if (opts->ahmac) {
 832                         put_unaligned_be64(opts->ahmac, ptr);
 833                         ptr += 2;
 834                 }
 835         }
 836
 837 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 838         if (OPTION_MPTCP_ADD_ADDR6 & opts->suboptions) {
 839                 if (opts->ahmac)
 840                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 841                                               TCPOLEN_MPTCP_ADD_ADDR6, 0,
 842                                               opts->addr_id);
 843                 else
 844                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 845                                               TCPOLEN_MPTCP_ADD_ADDR6_BASE,
 846                                               MPTCP_ADDR_ECHO,
 847                                               opts->addr_id);
 848                 memcpy((u8 *)ptr, opts->addr6.s6_addr, 16);
 849                 ptr += 4;
 850                 if (opts->ahmac) {
 851                         put_unaligned_be64(opts->ahmac, ptr);
 852                         ptr += 2;
 853                 }
 854         }
 855 #endif
 856
 857         if (OPTION_MPTCP_RM_ADDR & opts->suboptions) {
 858                 *ptr++ = mptcp_option(MPTCPOPT_RM_ADDR,
 859                                       TCPOLEN_MPTCP_RM_ADDR_BASE,
 860                                       0, opts->rm_id);
 861         }
 862
 863         if (OPTION_MPTCP_MPJ_SYNACK & opts->suboptions) {
 864                 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
 865                                       TCPOLEN_MPTCP_MPJ_SYNACK,
 866                                       opts->backup, opts->join_id);
 867                 put_unaligned_be64(opts->thmac, ptr);
 868                 ptr += 2;
 869                 put_unaligned_be32(opts->nonce, ptr);
 870                 ptr += 1;
 871         }
 872
 873         if (opts->ext_copy.use_ack || opts->ext_copy.use_map) {
 874                 struct mptcp_ext *mpext = &opts->ext_copy;
 875                 u8 len = TCPOLEN_MPTCP_DSS_BASE;
 876                 u8 flags = 0;
 877
 878                 if (mpext->use_ack) {
 879                         len += TCPOLEN_MPTCP_DSS_ACK64;
 880                         flags = MPTCP_DSS_HAS_ACK | MPTCP_DSS_ACK64;
 881                 }
 882
 883                 if (mpext->use_map) {
 884                         len += TCPOLEN_MPTCP_DSS_MAP64;
 885
 886                         /* Use only 64-bit mapping flags for now, add
 887                          * support for optional 32-bit mappings later.
 888                          */
 889                         flags |= MPTCP_DSS_HAS_MAP | MPTCP_DSS_DSN64;
 890                         if (mpext->data_fin)
 891                                 flags |= MPTCP_DSS_DATA_FIN;
 892                 }
 893
 894                 *ptr++ = mptcp_option(MPTCPOPT_DSS, len, 0, flags);
 895
 896                 if (mpext->use_ack) {
 897                         put_unaligned_be64(mpext->data_ack, ptr);
 898                         ptr += 2;
 899                 }
 900
 901                 if (mpext->use_map) {
 902                         put_unaligned_be64(mpext->data_seq, ptr);
 903                         ptr += 2;
 904                         put_unaligned_be32(mpext->subflow_seq, ptr);
 905                         ptr += 1;
 906                         put_unaligned_be32(mpext->data_len << 16 |
 907                                            TCPOPT_NOP << 8 | TCPOPT_NOP, ptr);
 908                 }
 909         }
 910 }