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 #define pr_fmt(fmt) "MPTCP: " fmt
   8
   9 #include <linux/kernel.h>
  10 #include <crypto/sha.h>
  11 #include <net/tcp.h>
  12 #include <net/mptcp.h>
  13 #include "protocol.h"
  14
  15 static bool mptcp_cap_flag_sha256(u8 flags)
  16 {
  17         return (flags & MPTCP_CAP_FLAG_MASK) == MPTCP_CAP_HMAC_SHA256;
  18 }
  19
  20 static void mptcp_parse_option(const struct sk_buff *skb,
  21                                const unsigned char *ptr, int opsize,
  22                                struct mptcp_options_received *mp_opt)
  23 {
  24         u8 subtype = *ptr >> 4;
  25         int expected_opsize;
  26         u8 version;
  27         u8 flags;
  28
  29         switch (subtype) {
  30         case MPTCPOPT_MP_CAPABLE:
  31                 /* strict size checking */
  32                 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
  33                         if (skb->len > tcp_hdr(skb)->doff << 2)
  34                                 expected_opsize = TCPOLEN_MPTCP_MPC_ACK_DATA;
  35                         else
  36                                 expected_opsize = TCPOLEN_MPTCP_MPC_ACK;
  37                 } else {
  38                         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)
  39                                 expected_opsize = TCPOLEN_MPTCP_MPC_SYNACK;
  40                         else
  41                                 expected_opsize = TCPOLEN_MPTCP_MPC_SYN;
  42                 }
  43                 if (opsize != expected_opsize)
  44                         break;
  45
  46                 /* try to be gentle vs future versions on the initial syn */
  47                 version = *ptr++ & MPTCP_VERSION_MASK;
  48                 if (opsize != TCPOLEN_MPTCP_MPC_SYN) {
  49                         if (version != MPTCP_SUPPORTED_VERSION)
  50                                 break;
  51                 } else if (version < MPTCP_SUPPORTED_VERSION) {
  52                         break;
  53                 }
  54
  55                 flags = *ptr++;
  56                 if (!mptcp_cap_flag_sha256(flags) ||
  57                     (flags & MPTCP_CAP_EXTENSIBILITY))
  58                         break;
  59
  60                 /* RFC 6824, Section 3.1:
  61                  * "For the Checksum Required bit (labeled "A"), if either
  62                  * host requires the use of checksums, checksums MUST be used.
  63                  * In other words, the only way for checksums not to be used
  64                  * is if both hosts in their SYNs set A=0."
  65                  *
  66                  * Section 3.3.0:
  67                  * "If a checksum is not present when its use has been
  68                  * negotiated, the receiver MUST close the subflow with a RST as
  69                  * it is considered broken."
  70                  *
  71                  * We don't implement DSS checksum - fall back to TCP.
  72                  */
  73                 if (flags & MPTCP_CAP_CHECKSUM_REQD)
  74                         break;
  75
  76                 mp_opt->mp_capable = 1;
  77                 if (opsize >= TCPOLEN_MPTCP_MPC_SYNACK) {
  78                         mp_opt->sndr_key = get_unaligned_be64(ptr);
  79                         ptr += 8;
  80                 }
  81                 if (opsize >= TCPOLEN_MPTCP_MPC_ACK) {
  82                         mp_opt->rcvr_key = get_unaligned_be64(ptr);
  83                         ptr += 8;
  84                 }
  85                 if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA) {
  86                         /* Section 3.1.:
  87                          * "the data parameters in a MP_CAPABLE are semantically
  88                          * equivalent to those in a DSS option and can be used
  89                          * interchangeably."
  90                          */
  91                         mp_opt->dss = 1;
  92                         mp_opt->use_map = 1;
  93                         mp_opt->mpc_map = 1;
  94                         mp_opt->data_len = get_unaligned_be16(ptr);
  95                         ptr += 2;
  96                 }
  97                 pr_debug("MP_CAPABLE version=%x, flags=%x, optlen=%d sndr=%llu, rcvr=%llu len=%d",
  98                          version, flags, opsize, mp_opt->sndr_key,
  99                          mp_opt->rcvr_key, mp_opt->data_len);
 100                 break;
 101
 102         case MPTCPOPT_MP_JOIN:
 103                 mp_opt->mp_join = 1;
 104                 if (opsize == TCPOLEN_MPTCP_MPJ_SYN) {
 105                         mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP;
 106                         mp_opt->join_id = *ptr++;
 107                         mp_opt->token = get_unaligned_be32(ptr);
 108                         ptr += 4;
 109                         mp_opt->nonce = get_unaligned_be32(ptr);
 110                         ptr += 4;
 111                         pr_debug("MP_JOIN bkup=%u, id=%u, token=%u, nonce=%u",
 112                                  mp_opt->backup, mp_opt->join_id,
 113                                  mp_opt->token, mp_opt->nonce);
 114                 } else if (opsize == TCPOLEN_MPTCP_MPJ_SYNACK) {
 115                         mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP;
 116                         mp_opt->join_id = *ptr++;
 117                         mp_opt->thmac = get_unaligned_be64(ptr);
 118                         ptr += 8;
 119                         mp_opt->nonce = get_unaligned_be32(ptr);
 120                         ptr += 4;
 121                         pr_debug("MP_JOIN bkup=%u, id=%u, thmac=%llu, nonce=%u",
 122                                  mp_opt->backup, mp_opt->join_id,
 123                                  mp_opt->thmac, mp_opt->nonce);
 124                 } else if (opsize == TCPOLEN_MPTCP_MPJ_ACK) {
 125                         ptr += 2;
 126                         memcpy(mp_opt->hmac, ptr, MPTCPOPT_HMAC_LEN);
 127                         pr_debug("MP_JOIN hmac");
 128                 } else {
 129                         pr_warn("MP_JOIN bad option size");
 130                         mp_opt->mp_join = 0;
 131                 }
 132                 break;
 133
 134         case MPTCPOPT_DSS:
 135                 pr_debug("DSS");
 136                 ptr++;
 137
 138                 /* we must clear 'mpc_map' be able to detect MP_CAPABLE
 139                  * map vs DSS map in mptcp_incoming_options(), and reconstruct
 140                  * map info accordingly
 141                  */
 142                 mp_opt->mpc_map = 0;
 143                 flags = (*ptr++) & MPTCP_DSS_FLAG_MASK;
 144                 mp_opt->data_fin = (flags & MPTCP_DSS_DATA_FIN) != 0;
 145                 mp_opt->dsn64 = (flags & MPTCP_DSS_DSN64) != 0;
 146                 mp_opt->use_map = (flags & MPTCP_DSS_HAS_MAP) != 0;
 147                 mp_opt->ack64 = (flags & MPTCP_DSS_ACK64) != 0;
 148                 mp_opt->use_ack = (flags & MPTCP_DSS_HAS_ACK);
 149
 150                 pr_debug("data_fin=%d dsn64=%d use_map=%d ack64=%d use_ack=%d",
 151                          mp_opt->data_fin, mp_opt->dsn64,
 152                          mp_opt->use_map, mp_opt->ack64,
 153                          mp_opt->use_ack);
 154
 155                 expected_opsize = TCPOLEN_MPTCP_DSS_BASE;
 156
 157                 if (mp_opt->use_ack) {
 158                         if (mp_opt->ack64)
 159                                 expected_opsize += TCPOLEN_MPTCP_DSS_ACK64;
 160                         else
 161                                 expected_opsize += TCPOLEN_MPTCP_DSS_ACK32;
 162                 }
 163
 164                 if (mp_opt->use_map) {
 165                         if (mp_opt->dsn64)
 166                                 expected_opsize += TCPOLEN_MPTCP_DSS_MAP64;
 167                         else
 168                                 expected_opsize += TCPOLEN_MPTCP_DSS_MAP32;
 169                 }
 170
 171                 /* RFC 6824, Section 3.3:
 172                  * If a checksum is present, but its use had
 173                  * not been negotiated in the MP_CAPABLE handshake,
 174                  * the checksum field MUST be ignored.
 175                  */
 176                 if (opsize != expected_opsize &&
 177                     opsize != expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM)
 178                         break;
 179
 180                 mp_opt->dss = 1;
 181
 182                 if (mp_opt->use_ack) {
 183                         if (mp_opt->ack64) {
 184                                 mp_opt->data_ack = get_unaligned_be64(ptr);
 185                                 ptr += 8;
 186                         } else {
 187                                 mp_opt->data_ack = get_unaligned_be32(ptr);
 188                                 ptr += 4;
 189                         }
 190
 191                         pr_debug("data_ack=%llu", mp_opt->data_ack);
 192                 }
 193
 194                 if (mp_opt->use_map) {
 195                         if (mp_opt->dsn64) {
 196                                 mp_opt->data_seq = get_unaligned_be64(ptr);
 197                                 ptr += 8;
 198                         } else {
 199                                 mp_opt->data_seq = get_unaligned_be32(ptr);
 200                                 ptr += 4;
 201                         }
 202
 203                         mp_opt->subflow_seq = get_unaligned_be32(ptr);
 204                         ptr += 4;
 205
 206                         mp_opt->data_len = get_unaligned_be16(ptr);
 207                         ptr += 2;
 208
 209                         pr_debug("data_seq=%llu subflow_seq=%u data_len=%u",
 210                                  mp_opt->data_seq, mp_opt->subflow_seq,
 211                                  mp_opt->data_len);
 212                 }
 213
 214                 break;
 215
 216         case MPTCPOPT_ADD_ADDR:
 217                 mp_opt->echo = (*ptr++) & MPTCP_ADDR_ECHO;
 218                 if (!mp_opt->echo) {
 219                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR ||
 220                             opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT)
 221                                 mp_opt->family = MPTCP_ADDR_IPVERSION_4;
 222 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 223                         else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6 ||
 224                                  opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT)
 225                                 mp_opt->family = MPTCP_ADDR_IPVERSION_6;
 226 #endif
 227                         else
 228                                 break;
 229                 } else {
 230                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE ||
 231                             opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT)
 232                                 mp_opt->family = MPTCP_ADDR_IPVERSION_4;
 233 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 234                         else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE ||
 235                                  opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT)
 236                                 mp_opt->family = MPTCP_ADDR_IPVERSION_6;
 237 #endif
 238                         else
 239                                 break;
 240                 }
 241
 242                 mp_opt->add_addr = 1;
 243                 mp_opt->port = 0;
 244                 mp_opt->addr_id = *ptr++;
 245                 pr_debug("ADD_ADDR: id=%d", mp_opt->addr_id);
 246                 if (mp_opt->family == MPTCP_ADDR_IPVERSION_4) {
 247                         memcpy((u8 *)&mp_opt->addr.s_addr, (u8 *)ptr, 4);
 248                         ptr += 4;
 249                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT ||
 250                             opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) {
 251                                 mp_opt->port = get_unaligned_be16(ptr);
 252                                 ptr += 2;
 253                         }
 254                 }
 255 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 256                 else {
 257                         memcpy(mp_opt->addr6.s6_addr, (u8 *)ptr, 16);
 258                         ptr += 16;
 259                         if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT ||
 260                             opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) {
 261                                 mp_opt->port = get_unaligned_be16(ptr);
 262                                 ptr += 2;
 263                         }
 264                 }
 265 #endif
 266                 if (!mp_opt->echo) {
 267                         mp_opt->ahmac = get_unaligned_be64(ptr);
 268                         ptr += 8;
 269                 }
 270                 break;
 271
 272         case MPTCPOPT_RM_ADDR:
 273                 if (opsize != TCPOLEN_MPTCP_RM_ADDR_BASE)
 274                         break;
 275
 276                 ptr++;
 277
 278                 mp_opt->rm_addr = 1;
 279                 mp_opt->rm_id = *ptr++;
 280                 pr_debug("RM_ADDR: id=%d", mp_opt->rm_id);
 281                 break;
 282
 283         default:
 284                 break;
 285         }
 286 }
 287
 288 void mptcp_get_options(const struct sk_buff *skb,
 289                        struct mptcp_options_received *mp_opt)
 290 {
 291         const struct tcphdr *th = tcp_hdr(skb);
 292         const unsigned char *ptr;
 293         int length;
 294
 295         /* initialize option status */
 296         mp_opt->mp_capable = 0;
 297         mp_opt->mp_join = 0;
 298         mp_opt->add_addr = 0;
 299         mp_opt->rm_addr = 0;
 300         mp_opt->dss = 0;
 301
 302         length = (th->doff * 4) - sizeof(struct tcphdr);
 303         ptr = (const unsigned char *)(th + 1);
 304
 305         while (length > 0) {
 306                 int opcode = *ptr++;
 307                 int opsize;
 308
 309                 switch (opcode) {
 310                 case TCPOPT_EOL:
 311                         return;
 312                 case TCPOPT_NOP:        /* Ref: RFC 793 section 3.1 */
 313                         length--;
 314                         continue;
 315                 default:
 316                         opsize = *ptr++;
 317                         if (opsize < 2) /* "silly options" */
 318                                 return;
 319                         if (opsize > length)
 320                                 return; /* don't parse partial options */
 321                         if (opcode == TCPOPT_MPTCP)
 322                                 mptcp_parse_option(skb, ptr, opsize, mp_opt);
 323                         ptr += opsize - 2;
 324                         length -= opsize;
 325                 }
 326         }
 327 }
 328
 329 bool mptcp_syn_options(struct sock *sk, const struct sk_buff *skb,
 330                        unsigned int *size, struct mptcp_out_options *opts)
 331 {
 332         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 333
 334         /* we will use snd_isn to detect first pkt [re]transmission
 335          * in mptcp_established_options_mp()
 336          */
 337         subflow->snd_isn = TCP_SKB_CB(skb)->end_seq;
 338         if (subflow->request_mptcp) {
 339                 opts->suboptions = OPTION_MPTCP_MPC_SYN;
 340                 *size = TCPOLEN_MPTCP_MPC_SYN;
 341                 return true;
 342         } else if (subflow->request_join) {
 343                 pr_debug("remote_token=%u, nonce=%u", subflow->remote_token,
 344                          subflow->local_nonce);
 345                 opts->suboptions = OPTION_MPTCP_MPJ_SYN;
 346                 opts->join_id = subflow->local_id;
 347                 opts->token = subflow->remote_token;
 348                 opts->nonce = subflow->local_nonce;
 349                 opts->backup = subflow->request_bkup;
 350                 *size = TCPOLEN_MPTCP_MPJ_SYN;
 351                 return true;
 352         }
 353         return false;
 354 }
 355
 356 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
 357  * TCP can't schedule delack timer before the subflow is fully established.
 358  * MPTCP uses the delack timer to do 3rd ack retransmissions
 359  */
 360 static void schedule_3rdack_retransmission(struct sock *sk)
 361 {
 362         struct inet_connection_sock *icsk = inet_csk(sk);
 363         struct tcp_sock *tp = tcp_sk(sk);
 364         unsigned long timeout;
 365
 366         /* reschedule with a timeout above RTT, as we must look only for drop */
 367         if (tp->srtt_us)
 368                 timeout = tp->srtt_us << 1;
 369         else
 370                 timeout = TCP_TIMEOUT_INIT;
 371
 372         WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
 373         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
 374         icsk->icsk_ack.timeout = timeout;
 375         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
 376 }
 377
 378 static void clear_3rdack_retransmission(struct sock *sk)
 379 {
 380         struct inet_connection_sock *icsk = inet_csk(sk);
 381
 382         sk_stop_timer(sk, &icsk->icsk_delack_timer);
 383         icsk->icsk_ack.timeout = 0;
 384         icsk->icsk_ack.ato = 0;
 385         icsk->icsk_ack.pending &= ~(ICSK_ACK_SCHED | ICSK_ACK_TIMER);
 386 }
 387
 388 static bool mptcp_established_options_mp(struct sock *sk, struct sk_buff *skb,
 389                                          unsigned int *size,
 390                                          unsigned int remaining,
 391                                          struct mptcp_out_options *opts)
 392 {
 393         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 394         struct mptcp_ext *mpext;
 395         unsigned int data_len;
 396
 397         /* When skb is not available, we better over-estimate the emitted
 398          * options len. A full DSS option (28 bytes) is longer than
 399          * TCPOLEN_MPTCP_MPC_ACK_DATA(22) or TCPOLEN_MPTCP_MPJ_ACK(24), so
 400          * tell the caller to defer the estimate to
 401          * mptcp_established_options_dss(), which will reserve enough space.
 402          */
 403         if (!skb)
 404                 return false;
 405
 406         /* MPC/MPJ needed only on 3rd ack packet */
 407         if (subflow->fully_established ||
 408             subflow->snd_isn != TCP_SKB_CB(skb)->seq)
 409                 return false;
 410
 411         if (subflow->mp_capable) {
 412                 mpext = mptcp_get_ext(skb);
 413                 data_len = mpext ? mpext->data_len : 0;
 414
 415                 /* we will check ext_copy.data_len in mptcp_write_options() to
 416                  * discriminate between TCPOLEN_MPTCP_MPC_ACK_DATA and
 417                  * TCPOLEN_MPTCP_MPC_ACK
 418                  */
 419                 opts->ext_copy.data_len = data_len;
 420                 opts->suboptions = OPTION_MPTCP_MPC_ACK;
 421                 opts->sndr_key = subflow->local_key;
 422                 opts->rcvr_key = subflow->remote_key;
 423
 424                 /* Section 3.1.
 425                  * The MP_CAPABLE option is carried on the SYN, SYN/ACK, and ACK
 426                  * packets that start the first subflow of an MPTCP connection,
 427                  * as well as the first packet that carries data
 428                  */
 429                 if (data_len > 0)
 430                         *size = ALIGN(TCPOLEN_MPTCP_MPC_ACK_DATA, 4);
 431                 else
 432                         *size = TCPOLEN_MPTCP_MPC_ACK;
 433
 434                 pr_debug("subflow=%p, local_key=%llu, remote_key=%llu map_len=%d",
 435                          subflow, subflow->local_key, subflow->remote_key,
 436                          data_len);
 437
 438                 return true;
 439         } else if (subflow->mp_join) {
 440                 opts->suboptions = OPTION_MPTCP_MPJ_ACK;
 441                 memcpy(opts->hmac, subflow->hmac, MPTCPOPT_HMAC_LEN);
 442                 *size = TCPOLEN_MPTCP_MPJ_ACK;
 443                 pr_debug("subflow=%p", subflow);
 444
 445                 schedule_3rdack_retransmission(sk);
 446                 return true;
 447         }
 448         return false;
 449 }
 450
 451 static void mptcp_write_data_fin(struct mptcp_subflow_context *subflow,
 452                                  struct mptcp_ext *ext)
 453 {
 454         if (!ext->use_map) {
 455                 /* RFC6824 requires a DSS mapping with specific values
 456                  * if DATA_FIN is set but no data payload is mapped
 457                  */
 458                 ext->data_fin = 1;
 459                 ext->use_map = 1;
 460                 ext->dsn64 = 1;
 461                 ext->data_seq = subflow->data_fin_tx_seq;
 462                 ext->subflow_seq = 0;
 463                 ext->data_len = 1;
 464         } else if (ext->data_seq + ext->data_len == subflow->data_fin_tx_seq) {
 465                 /* If there's an existing DSS mapping and it is the
 466                  * final mapping, DATA_FIN consumes 1 additional byte of
 467                  * mapping space.
 468                  */
 469                 ext->data_fin = 1;
 470                 ext->data_len++;
 471         }
 472 }
 473
 474 static bool mptcp_established_options_dss(struct sock *sk, struct sk_buff *skb,
 475                                           unsigned int *size,
 476                                           unsigned int remaining,
 477                                           struct mptcp_out_options *opts)
 478 {
 479         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 480         unsigned int dss_size = 0;
 481         struct mptcp_ext *mpext;
 482         struct mptcp_sock *msk;
 483         unsigned int ack_size;
 484         bool ret = false;
 485         u8 tcp_fin;
 486
 487         if (skb) {
 488                 mpext = mptcp_get_ext(skb);
 489                 tcp_fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
 490         } else {
 491                 mpext = NULL;
 492                 tcp_fin = 0;
 493         }
 494
 495         if (!skb || (mpext && mpext->use_map) || tcp_fin) {
 496                 unsigned int map_size;
 497
 498                 map_size = TCPOLEN_MPTCP_DSS_BASE + TCPOLEN_MPTCP_DSS_MAP64;
 499
 500                 remaining -= map_size;
 501                 dss_size = map_size;
 502                 if (mpext)
 503                         opts->ext_copy = *mpext;
 504
 505                 if (skb && tcp_fin && subflow->data_fin_tx_enable)
 506                         mptcp_write_data_fin(subflow, &opts->ext_copy);
 507                 ret = true;
 508         }
 509
 510         /* passive sockets msk will set the 'can_ack' after accept(), even
 511          * if the first subflow may have the already the remote key handy
 512          */
 513         opts->ext_copy.use_ack = 0;
 514         msk = mptcp_sk(subflow->conn);
 515         if (!READ_ONCE(msk->can_ack)) {
 516                 *size = ALIGN(dss_size, 4);
 517                 return ret;
 518         }
 519
 520         if (subflow->use_64bit_ack) {
 521                 ack_size = TCPOLEN_MPTCP_DSS_ACK64;
 522                 opts->ext_copy.data_ack = msk->ack_seq;
 523                 opts->ext_copy.ack64 = 1;
 524         } else {
 525                 ack_size = TCPOLEN_MPTCP_DSS_ACK32;
 526                 opts->ext_copy.data_ack32 = (uint32_t)(msk->ack_seq);
 527                 opts->ext_copy.ack64 = 0;
 528         }
 529         opts->ext_copy.use_ack = 1;
 530
 531         /* Add kind/length/subtype/flag overhead if mapping is not populated */
 532         if (dss_size == 0)
 533                 ack_size += TCPOLEN_MPTCP_DSS_BASE;
 534
 535         dss_size += ack_size;
 536
 537         *size = ALIGN(dss_size, 4);
 538         return true;
 539 }
 540
 541 static u64 add_addr_generate_hmac(u64 key1, u64 key2, u8 addr_id,
 542                                   struct in_addr *addr)
 543 {
 544         u8 hmac[SHA256_DIGEST_SIZE];
 545         u8 msg[7];
 546
 547         msg[0] = addr_id;
 548         memcpy(&msg[1], &addr->s_addr, 4);
 549         msg[5] = 0;
 550         msg[6] = 0;
 551
 552         mptcp_crypto_hmac_sha(key1, key2, msg, 7, hmac);
 553
 554         return get_unaligned_be64(&hmac[SHA256_DIGEST_SIZE - sizeof(u64)]);
 555 }
 556
 557 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 558 static u64 add_addr6_generate_hmac(u64 key1, u64 key2, u8 addr_id,
 559                                    struct in6_addr *addr)
 560 {
 561         u8 hmac[SHA256_DIGEST_SIZE];
 562         u8 msg[19];
 563
 564         msg[0] = addr_id;
 565         memcpy(&msg[1], &addr->s6_addr, 16);
 566         msg[17] = 0;
 567         msg[18] = 0;
 568
 569         mptcp_crypto_hmac_sha(key1, key2, msg, 19, hmac);
 570
 571         return get_unaligned_be64(&hmac[SHA256_DIGEST_SIZE - sizeof(u64)]);
 572 }
 573 #endif
 574
 575 static bool mptcp_established_options_addr(struct sock *sk,
 576                                            unsigned int *size,
 577                                            unsigned int remaining,
 578                                            struct mptcp_out_options *opts)
 579 {
 580         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 581         struct mptcp_sock *msk = mptcp_sk(subflow->conn);
 582         struct mptcp_addr_info saddr;
 583         int len;
 584
 585         if (!mptcp_pm_should_signal(msk) ||
 586             !(mptcp_pm_addr_signal(msk, remaining, &saddr)))
 587                 return false;
 588
 589         len = mptcp_add_addr_len(saddr.family);
 590         if (remaining < len)
 591                 return false;
 592
 593         *size = len;
 594         opts->addr_id = saddr.id;
 595         if (saddr.family == AF_INET) {
 596                 opts->suboptions |= OPTION_MPTCP_ADD_ADDR;
 597                 opts->addr = saddr.addr;
 598                 opts->ahmac = add_addr_generate_hmac(msk->local_key,
 599                                                      msk->remote_key,
 600                                                      opts->addr_id,
 601                                                      &opts->addr);
 602         }
 603 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 604         else if (saddr.family == AF_INET6) {
 605                 opts->suboptions |= OPTION_MPTCP_ADD_ADDR6;
 606                 opts->addr6 = saddr.addr6;
 607                 opts->ahmac = add_addr6_generate_hmac(msk->local_key,
 608                                                       msk->remote_key,
 609                                                       opts->addr_id,
 610                                                       &opts->addr6);
 611         }
 612 #endif
 613         pr_debug("addr_id=%d, ahmac=%llu", opts->addr_id, opts->ahmac);
 614
 615         return true;
 616 }
 617
 618 bool mptcp_established_options(struct sock *sk, struct sk_buff *skb,
 619                                unsigned int *size, unsigned int remaining,
 620                                struct mptcp_out_options *opts)
 621 {
 622         unsigned int opt_size = 0;
 623         bool ret = false;
 624
 625         opts->suboptions = 0;
 626
 627         if (mptcp_established_options_mp(sk, skb, &opt_size, remaining, opts))
 628                 ret = true;
 629         else if (mptcp_established_options_dss(sk, skb, &opt_size, remaining,
 630                                                opts))
 631                 ret = true;
 632
 633         /* we reserved enough space for the above options, and exceeding the
 634          * TCP option space would be fatal
 635          */
 636         if (WARN_ON_ONCE(opt_size > remaining))
 637                 return false;
 638
 639         *size += opt_size;
 640         remaining -= opt_size;
 641         if (mptcp_established_options_addr(sk, &opt_size, remaining, opts)) {
 642                 *size += opt_size;
 643                 remaining -= opt_size;
 644                 ret = true;
 645         }
 646
 647         return ret;
 648 }
 649
 650 bool mptcp_synack_options(const struct request_sock *req, unsigned int *size,
 651                           struct mptcp_out_options *opts)
 652 {
 653         struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
 654
 655         if (subflow_req->mp_capable) {
 656                 opts->suboptions = OPTION_MPTCP_MPC_SYNACK;
 657                 opts->sndr_key = subflow_req->local_key;
 658                 *size = TCPOLEN_MPTCP_MPC_SYNACK;
 659                 pr_debug("subflow_req=%p, local_key=%llu",
 660                          subflow_req, subflow_req->local_key);
 661                 return true;
 662         } else if (subflow_req->mp_join) {
 663                 opts->suboptions = OPTION_MPTCP_MPJ_SYNACK;
 664                 opts->backup = subflow_req->backup;
 665                 opts->join_id = subflow_req->local_id;
 666                 opts->thmac = subflow_req->thmac;
 667                 opts->nonce = subflow_req->local_nonce;
 668                 pr_debug("req=%p, bkup=%u, id=%u, thmac=%llu, nonce=%u",
 669                          subflow_req, opts->backup, opts->join_id,
 670                          opts->thmac, opts->nonce);
 671                 *size = TCPOLEN_MPTCP_MPJ_SYNACK;
 672                 return true;
 673         }
 674         return false;
 675 }
 676
 677 static bool check_fully_established(struct mptcp_sock *msk, struct sock *sk,
 678                                     struct mptcp_subflow_context *subflow,
 679                                     struct sk_buff *skb,
 680                                     struct mptcp_options_received *mp_opt)
 681 {
 682         /* here we can process OoO, in-window pkts, only in-sequence 4th ack
 683          * will make the subflow fully established
 684          */
 685         if (likely(subflow->fully_established)) {
 686                 /* on passive sockets, check for 3rd ack retransmission
 687                  * note that msk is always set by subflow_syn_recv_sock()
 688                  * for mp_join subflows
 689                  */
 690                 if (TCP_SKB_CB(skb)->seq == subflow->ssn_offset + 1 &&
 691                     TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq &&
 692                     subflow->mp_join && mp_opt->mp_join &&
 693                     READ_ONCE(msk->pm.server_side))
 694                         tcp_send_ack(sk);
 695                 goto fully_established;
 696         }
 697
 698         /* we should process OoO packets before the first subflow is fully
 699          * established, but not expected for MP_JOIN subflows
 700          */
 701         if (TCP_SKB_CB(skb)->seq != subflow->ssn_offset + 1)
 702                 return subflow->mp_capable;
 703
 704         if (mp_opt->dss && mp_opt->use_ack) {
 705                 /* subflows are fully established as soon as we get any
 706                  * additional ack.
 707                  */
 708                 subflow->fully_established = 1;
 709                 goto fully_established;
 710         }
 711
 712         /* If the first established packet does not contain MP_CAPABLE + data
 713          * then fallback to TCP
 714          */
 715         if (!mp_opt->mp_capable) {
 716                 subflow->mp_capable = 0;
 717                 tcp_sk(sk)->is_mptcp = 0;
 718                 return false;
 719         }
 720
 721         if (unlikely(!READ_ONCE(msk->pm.server_side)))
 722                 pr_warn_once("bogus mpc option on established client sk");
 723         subflow->fully_established = 1;
 724         subflow->remote_key = mp_opt->sndr_key;
 725         subflow->can_ack = 1;
 726
 727 fully_established:
 728         if (likely(subflow->pm_notified))
 729                 return true;
 730
 731         subflow->pm_notified = 1;
 732         if (subflow->mp_join) {
 733                 clear_3rdack_retransmission(sk);
 734                 mptcp_pm_subflow_established(msk, subflow);
 735         } else {
 736                 mptcp_pm_fully_established(msk);
 737         }
 738         return true;
 739 }
 740
 741 static u64 expand_ack(u64 old_ack, u64 cur_ack, bool use_64bit)
 742 {
 743         u32 old_ack32, cur_ack32;
 744
 745         if (use_64bit)
 746                 return cur_ack;
 747
 748         old_ack32 = (u32)old_ack;
 749         cur_ack32 = (u32)cur_ack;
 750         cur_ack = (old_ack & GENMASK_ULL(63, 32)) + cur_ack32;
 751         if (unlikely(before(cur_ack32, old_ack32)))
 752                 return cur_ack + (1LL << 32);
 753         return cur_ack;
 754 }
 755
 756 static void update_una(struct mptcp_sock *msk,
 757                        struct mptcp_options_received *mp_opt)
 758 {
 759         u64 new_snd_una, snd_una, old_snd_una = atomic64_read(&msk->snd_una);
 760         u64 write_seq = READ_ONCE(msk->write_seq);
 761
 762         /* avoid ack expansion on update conflict, to reduce the risk of
 763          * wrongly expanding to a future ack sequence number, which is way
 764          * more dangerous than missing an ack
 765          */
 766         new_snd_una = expand_ack(old_snd_una, mp_opt->data_ack, mp_opt->ack64);
 767
 768         /* ACK for data not even sent yet? Ignore. */
 769         if (after64(new_snd_una, write_seq))
 770                 new_snd_una = old_snd_una;
 771
 772         while (after64(new_snd_una, old_snd_una)) {
 773                 snd_una = old_snd_una;
 774                 old_snd_una = atomic64_cmpxchg(&msk->snd_una, snd_una,
 775                                                new_snd_una);
 776                 if (old_snd_una == snd_una) {
 777                         mptcp_data_acked((struct sock *)msk);
 778                         break;
 779                 }
 780         }
 781 }
 782
 783 static bool add_addr_hmac_valid(struct mptcp_sock *msk,
 784                                 struct mptcp_options_received *mp_opt)
 785 {
 786         u64 hmac = 0;
 787
 788         if (mp_opt->echo)
 789                 return true;
 790
 791         if (mp_opt->family == MPTCP_ADDR_IPVERSION_4)
 792                 hmac = add_addr_generate_hmac(msk->remote_key,
 793                                               msk->local_key,
 794                                               mp_opt->addr_id, &mp_opt->addr);
 795 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 796         else
 797                 hmac = add_addr6_generate_hmac(msk->remote_key,
 798                                                msk->local_key,
 799                                                mp_opt->addr_id, &mp_opt->addr6);
 800 #endif
 801
 802         pr_debug("msk=%p, ahmac=%llu, mp_opt->ahmac=%llu\n",
 803                  msk, (unsigned long long)hmac,
 804                  (unsigned long long)mp_opt->ahmac);
 805
 806         return hmac == mp_opt->ahmac;
 807 }
 808
 809 void mptcp_incoming_options(struct sock *sk, struct sk_buff *skb,
 810                             struct tcp_options_received *opt_rx)
 811 {
 812         struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
 813         struct mptcp_sock *msk = mptcp_sk(subflow->conn);
 814         struct mptcp_options_received mp_opt;
 815         struct mptcp_ext *mpext;
 816
 817         mptcp_get_options(skb, &mp_opt);
 818         if (!check_fully_established(msk, sk, subflow, skb, &mp_opt))
 819                 return;
 820
 821         if (mp_opt.add_addr && add_addr_hmac_valid(msk, &mp_opt)) {
 822                 struct mptcp_addr_info addr;
 823
 824                 addr.port = htons(mp_opt.port);
 825                 addr.id = mp_opt.addr_id;
 826                 if (mp_opt.family == MPTCP_ADDR_IPVERSION_4) {
 827                         addr.family = AF_INET;
 828                         addr.addr = mp_opt.addr;
 829                 }
 830 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 831                 else if (mp_opt.family == MPTCP_ADDR_IPVERSION_6) {
 832                         addr.family = AF_INET6;
 833                         addr.addr6 = mp_opt.addr6;
 834                 }
 835 #endif
 836                 if (!mp_opt.echo)
 837                         mptcp_pm_add_addr_received(msk, &addr);
 838                 mp_opt.add_addr = 0;
 839         }
 840
 841         if (!mp_opt.dss)
 842                 return;
 843
 844         /* we can't wait for recvmsg() to update the ack_seq, otherwise
 845          * monodirectional flows will stuck
 846          */
 847         if (mp_opt.use_ack)
 848                 update_una(msk, &mp_opt);
 849
 850         mpext = skb_ext_add(skb, SKB_EXT_MPTCP);
 851         if (!mpext)
 852                 return;
 853
 854         memset(mpext, 0, sizeof(*mpext));
 855
 856         if (mp_opt.use_map) {
 857                 if (mp_opt.mpc_map) {
 858                         /* this is an MP_CAPABLE carrying MPTCP data
 859                          * we know this map the first chunk of data
 860                          */
 861                         mptcp_crypto_key_sha(subflow->remote_key, NULL,
 862                                              &mpext->data_seq);
 863                         mpext->data_seq++;
 864                         mpext->subflow_seq = 1;
 865                         mpext->dsn64 = 1;
 866                         mpext->mpc_map = 1;
 867                         mpext->data_fin = 0;
 868                 } else {
 869                         mpext->data_seq = mp_opt.data_seq;
 870                         mpext->subflow_seq = mp_opt.subflow_seq;
 871                         mpext->dsn64 = mp_opt.dsn64;
 872                         mpext->data_fin = mp_opt.data_fin;
 873                 }
 874                 mpext->data_len = mp_opt.data_len;
 875                 mpext->use_map = 1;
 876         }
 877 }
 878
 879 void mptcp_write_options(__be32 *ptr, struct mptcp_out_options *opts)
 880 {
 881         if ((OPTION_MPTCP_MPC_SYN | OPTION_MPTCP_MPC_SYNACK |
 882              OPTION_MPTCP_MPC_ACK) & opts->suboptions) {
 883                 u8 len;
 884
 885                 if (OPTION_MPTCP_MPC_SYN & opts->suboptions)
 886                         len = TCPOLEN_MPTCP_MPC_SYN;
 887                 else if (OPTION_MPTCP_MPC_SYNACK & opts->suboptions)
 888                         len = TCPOLEN_MPTCP_MPC_SYNACK;
 889                 else if (opts->ext_copy.data_len)
 890                         len = TCPOLEN_MPTCP_MPC_ACK_DATA;
 891                 else
 892                         len = TCPOLEN_MPTCP_MPC_ACK;
 893
 894                 *ptr++ = mptcp_option(MPTCPOPT_MP_CAPABLE, len,
 895                                       MPTCP_SUPPORTED_VERSION,
 896                                       MPTCP_CAP_HMAC_SHA256);
 897
 898                 if (!((OPTION_MPTCP_MPC_SYNACK | OPTION_MPTCP_MPC_ACK) &
 899                     opts->suboptions))
 900                         goto mp_capable_done;
 901
 902                 put_unaligned_be64(opts->sndr_key, ptr);
 903                 ptr += 2;
 904                 if (!((OPTION_MPTCP_MPC_ACK) & opts->suboptions))
 905                         goto mp_capable_done;
 906
 907                 put_unaligned_be64(opts->rcvr_key, ptr);
 908                 ptr += 2;
 909                 if (!opts->ext_copy.data_len)
 910                         goto mp_capable_done;
 911
 912                 put_unaligned_be32(opts->ext_copy.data_len << 16 |
 913                                    TCPOPT_NOP << 8 | TCPOPT_NOP, ptr);
 914                 ptr += 1;
 915         }
 916
 917 mp_capable_done:
 918         if (OPTION_MPTCP_ADD_ADDR & opts->suboptions) {
 919                 if (opts->ahmac)
 920                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 921                                               TCPOLEN_MPTCP_ADD_ADDR, 0,
 922                                               opts->addr_id);
 923                 else
 924                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 925                                               TCPOLEN_MPTCP_ADD_ADDR_BASE,
 926                                               MPTCP_ADDR_ECHO,
 927                                               opts->addr_id);
 928                 memcpy((u8 *)ptr, (u8 *)&opts->addr.s_addr, 4);
 929                 ptr += 1;
 930                 if (opts->ahmac) {
 931                         put_unaligned_be64(opts->ahmac, ptr);
 932                         ptr += 2;
 933                 }
 934         }
 935
 936 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 937         if (OPTION_MPTCP_ADD_ADDR6 & opts->suboptions) {
 938                 if (opts->ahmac)
 939                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 940                                               TCPOLEN_MPTCP_ADD_ADDR6, 0,
 941                                               opts->addr_id);
 942                 else
 943                         *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
 944                                               TCPOLEN_MPTCP_ADD_ADDR6_BASE,
 945                                               MPTCP_ADDR_ECHO,
 946                                               opts->addr_id);
 947                 memcpy((u8 *)ptr, opts->addr6.s6_addr, 16);
 948                 ptr += 4;
 949                 if (opts->ahmac) {
 950                         put_unaligned_be64(opts->ahmac, ptr);
 951                         ptr += 2;
 952                 }
 953         }
 954 #endif
 955
 956         if (OPTION_MPTCP_RM_ADDR & opts->suboptions) {
 957                 *ptr++ = mptcp_option(MPTCPOPT_RM_ADDR,
 958                                       TCPOLEN_MPTCP_RM_ADDR_BASE,
 959                                       0, opts->rm_id);
 960         }
 961
 962         if (OPTION_MPTCP_MPJ_SYN & opts->suboptions) {
 963                 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
 964                                       TCPOLEN_MPTCP_MPJ_SYN,
 965                                       opts->backup, opts->join_id);
 966                 put_unaligned_be32(opts->token, ptr);
 967                 ptr += 1;
 968                 put_unaligned_be32(opts->nonce, ptr);
 969                 ptr += 1;
 970         }
 971
 972         if (OPTION_MPTCP_MPJ_SYNACK & opts->suboptions) {
 973                 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
 974                                       TCPOLEN_MPTCP_MPJ_SYNACK,
 975                                       opts->backup, opts->join_id);
 976                 put_unaligned_be64(opts->thmac, ptr);
 977                 ptr += 2;
 978                 put_unaligned_be32(opts->nonce, ptr);
 979                 ptr += 1;
 980         }
 981
 982         if (OPTION_MPTCP_MPJ_ACK & opts->suboptions) {
 983                 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
 984                                       TCPOLEN_MPTCP_MPJ_ACK, 0, 0);
 985                 memcpy(ptr, opts->hmac, MPTCPOPT_HMAC_LEN);
 986                 ptr += 5;
 987         }
 988
 989         if (opts->ext_copy.use_ack || opts->ext_copy.use_map) {
 990                 struct mptcp_ext *mpext = &opts->ext_copy;
 991                 u8 len = TCPOLEN_MPTCP_DSS_BASE;
 992                 u8 flags = 0;
 993
 994                 if (mpext->use_ack) {
 995                         flags = MPTCP_DSS_HAS_ACK;
 996                         if (mpext->ack64) {
 997                                 len += TCPOLEN_MPTCP_DSS_ACK64;
 998                                 flags |= MPTCP_DSS_ACK64;
 999                         } else {
1000                                 len += TCPOLEN_MPTCP_DSS_ACK32;
1001                         }
1002                 }
1003
1004                 if (mpext->use_map) {
1005                         len += TCPOLEN_MPTCP_DSS_MAP64;
1006
1007                         /* Use only 64-bit mapping flags for now, add
1008                          * support for optional 32-bit mappings later.
1009                          */
1010                         flags |= MPTCP_DSS_HAS_MAP | MPTCP_DSS_DSN64;
1011                         if (mpext->data_fin)
1012                                 flags |= MPTCP_DSS_DATA_FIN;
1013                 }
1014
1015                 *ptr++ = mptcp_option(MPTCPOPT_DSS, len, 0, flags);
1016
1017                 if (mpext->use_ack) {
1018                         if (mpext->ack64) {
1019                                 put_unaligned_be64(mpext->data_ack, ptr);
1020                                 ptr += 2;
1021                         } else {
1022                                 put_unaligned_be32(mpext->data_ack32, ptr);
1023                                 ptr += 1;
1024                         }
1025                 }
1026
1027                 if (mpext->use_map) {
1028                         put_unaligned_be64(mpext->data_seq, ptr);
1029                         ptr += 2;
1030                         put_unaligned_be32(mpext->subflow_seq, ptr);
1031                         ptr += 1;
1032                         put_unaligned_be32(mpext->data_len << 16 |
1033                                            TCPOPT_NOP << 8 | TCPOPT_NOP, ptr);
1034                 }
1035         }
1036 }