net/netfilter/nf_conntrack_proto_dccp.c

   1 /*
   2  * DCCP connection tracking protocol helper
   3  *
   4  * Copyright (c) 2005, 2006, 2008 Patrick McHardy <kaber@trash.net>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  *
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/module.h>
  13 #include <linux/init.h>
  14 #include <linux/sysctl.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/dccp.h>
  18 #include <linux/slab.h>
  19
  20 #include <net/net_namespace.h>
  21 #include <net/netns/generic.h>
  22
  23 #include <linux/netfilter/nfnetlink_conntrack.h>
  24 #include <net/netfilter/nf_conntrack.h>
  25 #include <net/netfilter/nf_conntrack_l4proto.h>
  26 #include <net/netfilter/nf_conntrack_ecache.h>
  27 #include <net/netfilter/nf_log.h>
  28
  29 /* Timeouts are based on values from RFC4340:
  30  *
  31  * - REQUEST:
  32  *
  33  *   8.1.2. Client Request
  34  *
  35  *   A client MAY give up on its DCCP-Requests after some time
  36  *   (3 minutes, for example).
  37  *
  38  * - RESPOND:
  39  *
  40  *   8.1.3. Server Response
  41  *
  42  *   It MAY also leave the RESPOND state for CLOSED after a timeout of
  43  *   not less than 4MSL (8 minutes);
  44  *
  45  * - PARTOPEN:
  46  *
  47  *   8.1.5. Handshake Completion
  48  *
  49  *   If the client remains in PARTOPEN for more than 4MSL (8 minutes),
  50  *   it SHOULD reset the connection with Reset Code 2, "Aborted".
  51  *
  52  * - OPEN:
  53  *
  54  *   The DCCP timestamp overflows after 11.9 hours. If the connection
  55  *   stays idle this long the sequence number won't be recognized
  56  *   as valid anymore.
  57  *
  58  * - CLOSEREQ/CLOSING:
  59  *
  60  *   8.3. Termination
  61  *
  62  *   The retransmission timer should initially be set to go off in two
  63  *   round-trip times and should back off to not less than once every
  64  *   64 seconds ...
  65  *
  66  * - TIMEWAIT:
  67  *
  68  *   4.3. States
  69  *
  70  *   A server or client socket remains in this state for 2MSL (4 minutes)
  71  *   after the connection has been town down, ...
  72  */
  73
  74 #define DCCP_MSL (2 * 60 * HZ)
  75
  76 static const char * const dccp_state_names[] = {
  77         [CT_DCCP_NONE]          = "NONE",
  78         [CT_DCCP_REQUEST]       = "REQUEST",
  79         [CT_DCCP_RESPOND]       = "RESPOND",
  80         [CT_DCCP_PARTOPEN]      = "PARTOPEN",
  81         [CT_DCCP_OPEN]          = "OPEN",
  82         [CT_DCCP_CLOSEREQ]      = "CLOSEREQ",
  83         [CT_DCCP_CLOSING]       = "CLOSING",
  84         [CT_DCCP_TIMEWAIT]      = "TIMEWAIT",
  85         [CT_DCCP_IGNORE]        = "IGNORE",
  86         [CT_DCCP_INVALID]       = "INVALID",
  87 };
  88
  89 #define sNO     CT_DCCP_NONE
  90 #define sRQ     CT_DCCP_REQUEST
  91 #define sRS     CT_DCCP_RESPOND
  92 #define sPO     CT_DCCP_PARTOPEN
  93 #define sOP     CT_DCCP_OPEN
  94 #define sCR     CT_DCCP_CLOSEREQ
  95 #define sCG     CT_DCCP_CLOSING
  96 #define sTW     CT_DCCP_TIMEWAIT
  97 #define sIG     CT_DCCP_IGNORE
  98 #define sIV     CT_DCCP_INVALID
  99
 100 /*
 101  * DCCP state transition table
 102  *
 103  * The assumption is the same as for TCP tracking:
 104  *
 105  * We are the man in the middle. All the packets go through us but might
 106  * get lost in transit to the destination. It is assumed that the destination
 107  * can't receive segments we haven't seen.
 108  *
 109  * The following states exist:
 110  *
 111  * NONE:        Initial state, expecting Request
 112  * REQUEST:     Request seen, waiting for Response from server
 113  * RESPOND:     Response from server seen, waiting for Ack from client
 114  * PARTOPEN:    Ack after Response seen, waiting for packet other than Response,
 115  *              Reset or Sync from server
 116  * OPEN:        Packet other than Response, Reset or Sync seen
 117  * CLOSEREQ:    CloseReq from server seen, expecting Close from client
 118  * CLOSING:     Close seen, expecting Reset
 119  * TIMEWAIT:    Reset seen
 120  * IGNORE:      Not determinable whether packet is valid
 121  *
 122  * Some states exist only on one side of the connection: REQUEST, RESPOND,
 123  * PARTOPEN, CLOSEREQ. For the other side these states are equivalent to
 124  * the one it was in before.
 125  *
 126  * Packets are marked as ignored (sIG) if we don't know if they're valid
 127  * (for example a reincarnation of a connection we didn't notice is dead
 128  * already) and the server may send back a connection closing Reset or a
 129  * Response. They're also used for Sync/SyncAck packets, which we don't
 130  * care about.
 131  */
 132 static const u_int8_t
 133 dccp_state_table[CT_DCCP_ROLE_MAX + 1][DCCP_PKT_SYNCACK + 1][CT_DCCP_MAX + 1] = {
 134         [CT_DCCP_ROLE_CLIENT] = {
 135                 [DCCP_PKT_REQUEST] = {
 136                 /*
 137                  * sNO -> sRQ           Regular Request
 138                  * sRQ -> sRQ           Retransmitted Request or reincarnation
 139                  * sRS -> sRS           Retransmitted Request (apparently Response
 140                  *                      got lost after we saw it) or reincarnation
 141                  * sPO -> sIG           Ignore, conntrack might be out of sync
 142                  * sOP -> sIG           Ignore, conntrack might be out of sync
 143                  * sCR -> sIG           Ignore, conntrack might be out of sync
 144                  * sCG -> sIG           Ignore, conntrack might be out of sync
 145                  * sTW -> sRQ           Reincarnation
 146                  *
 147                  *      sNO, sRQ, sRS, sPO. sOP, sCR, sCG, sTW, */
 148                         sRQ, sRQ, sRS, sIG, sIG, sIG, sIG, sRQ,
 149                 },
 150                 [DCCP_PKT_RESPONSE] = {
 151                 /*
 152                  * sNO -> sIV           Invalid
 153                  * sRQ -> sIG           Ignore, might be response to ignored Request
 154                  * sRS -> sIG           Ignore, might be response to ignored Request
 155                  * sPO -> sIG           Ignore, might be response to ignored Request
 156                  * sOP -> sIG           Ignore, might be response to ignored Request
 157                  * sCR -> sIG           Ignore, might be response to ignored Request
 158                  * sCG -> sIG           Ignore, might be response to ignored Request
 159                  * sTW -> sIV           Invalid, reincarnation in reverse direction
 160                  *                      goes through sRQ
 161                  *
 162                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 163                         sIV, sIG, sIG, sIG, sIG, sIG, sIG, sIV,
 164                 },
 165                 [DCCP_PKT_ACK] = {
 166                 /*
 167                  * sNO -> sIV           No connection
 168                  * sRQ -> sIV           No connection
 169                  * sRS -> sPO           Ack for Response, move to PARTOPEN (8.1.5.)
 170                  * sPO -> sPO           Retransmitted Ack for Response, remain in PARTOPEN
 171                  * sOP -> sOP           Regular ACK, remain in OPEN
 172                  * sCR -> sCR           Ack in CLOSEREQ MAY be processed (8.3.)
 173                  * sCG -> sCG           Ack in CLOSING MAY be processed (8.3.)
 174                  * sTW -> sIV
 175                  *
 176                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 177                         sIV, sIV, sPO, sPO, sOP, sCR, sCG, sIV
 178                 },
 179                 [DCCP_PKT_DATA] = {
 180                 /*
 181                  * sNO -> sIV           No connection
 182                  * sRQ -> sIV           No connection
 183                  * sRS -> sIV           No connection
 184                  * sPO -> sIV           MUST use DataAck in PARTOPEN state (8.1.5.)
 185                  * sOP -> sOP           Regular Data packet
 186                  * sCR -> sCR           Data in CLOSEREQ MAY be processed (8.3.)
 187                  * sCG -> sCG           Data in CLOSING MAY be processed (8.3.)
 188                  * sTW -> sIV
 189                  *
 190                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 191                         sIV, sIV, sIV, sIV, sOP, sCR, sCG, sIV,
 192                 },
 193                 [DCCP_PKT_DATAACK] = {
 194                 /*
 195                  * sNO -> sIV           No connection
 196                  * sRQ -> sIV           No connection
 197                  * sRS -> sPO           Ack for Response, move to PARTOPEN (8.1.5.)
 198                  * sPO -> sPO           Remain in PARTOPEN state
 199                  * sOP -> sOP           Regular DataAck packet in OPEN state
 200                  * sCR -> sCR           DataAck in CLOSEREQ MAY be processed (8.3.)
 201                  * sCG -> sCG           DataAck in CLOSING MAY be processed (8.3.)
 202                  * sTW -> sIV
 203                  *
 204                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 205                         sIV, sIV, sPO, sPO, sOP, sCR, sCG, sIV
 206                 },
 207                 [DCCP_PKT_CLOSEREQ] = {
 208                 /*
 209                  * CLOSEREQ may only be sent by the server.
 210                  *
 211                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 212                         sIV, sIV, sIV, sIV, sIV, sIV, sIV, sIV
 213                 },
 214                 [DCCP_PKT_CLOSE] = {
 215                 /*
 216                  * sNO -> sIV           No connection
 217                  * sRQ -> sIV           No connection
 218                  * sRS -> sIV           No connection
 219                  * sPO -> sCG           Client-initiated close
 220                  * sOP -> sCG           Client-initiated close
 221                  * sCR -> sCG           Close in response to CloseReq (8.3.)
 222                  * sCG -> sCG           Retransmit
 223                  * sTW -> sIV           Late retransmit, already in TIME_WAIT
 224                  *
 225                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 226                         sIV, sIV, sIV, sCG, sCG, sCG, sIV, sIV
 227                 },
 228                 [DCCP_PKT_RESET] = {
 229                 /*
 230                  * sNO -> sIV           No connection
 231                  * sRQ -> sTW           Sync received or timeout, SHOULD send Reset (8.1.1.)
 232                  * sRS -> sTW           Response received without Request
 233                  * sPO -> sTW           Timeout, SHOULD send Reset (8.1.5.)
 234                  * sOP -> sTW           Connection reset
 235                  * sCR -> sTW           Connection reset
 236                  * sCG -> sTW           Connection reset
 237                  * sTW -> sIG           Ignore (don't refresh timer)
 238                  *
 239                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 240                         sIV, sTW, sTW, sTW, sTW, sTW, sTW, sIG
 241                 },
 242                 [DCCP_PKT_SYNC] = {
 243                 /*
 244                  * We currently ignore Sync packets
 245                  *
 246                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 247                         sIG, sIG, sIG, sIG, sIG, sIG, sIG, sIG,
 248                 },
 249                 [DCCP_PKT_SYNCACK] = {
 250                 /*
 251                  * We currently ignore SyncAck packets
 252                  *
 253                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 254                         sIG, sIG, sIG, sIG, sIG, sIG, sIG, sIG,
 255                 },
 256         },
 257         [CT_DCCP_ROLE_SERVER] = {
 258                 [DCCP_PKT_REQUEST] = {
 259                 /*
 260                  * sNO -> sIV           Invalid
 261                  * sRQ -> sIG           Ignore, conntrack might be out of sync
 262                  * sRS -> sIG           Ignore, conntrack might be out of sync
 263                  * sPO -> sIG           Ignore, conntrack might be out of sync
 264                  * sOP -> sIG           Ignore, conntrack might be out of sync
 265                  * sCR -> sIG           Ignore, conntrack might be out of sync
 266                  * sCG -> sIG           Ignore, conntrack might be out of sync
 267                  * sTW -> sRQ           Reincarnation, must reverse roles
 268                  *
 269                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 270                         sIV, sIG, sIG, sIG, sIG, sIG, sIG, sRQ
 271                 },
 272                 [DCCP_PKT_RESPONSE] = {
 273                 /*
 274                  * sNO -> sIV           Response without Request
 275                  * sRQ -> sRS           Response to clients Request
 276                  * sRS -> sRS           Retransmitted Response (8.1.3. SHOULD NOT)
 277                  * sPO -> sIG           Response to an ignored Request or late retransmit
 278                  * sOP -> sIG           Ignore, might be response to ignored Request
 279                  * sCR -> sIG           Ignore, might be response to ignored Request
 280                  * sCG -> sIG           Ignore, might be response to ignored Request
 281                  * sTW -> sIV           Invalid, Request from client in sTW moves to sRQ
 282                  *
 283                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 284                         sIV, sRS, sRS, sIG, sIG, sIG, sIG, sIV
 285                 },
 286                 [DCCP_PKT_ACK] = {
 287                 /*
 288                  * sNO -> sIV           No connection
 289                  * sRQ -> sIV           No connection
 290                  * sRS -> sIV           No connection
 291                  * sPO -> sOP           Enter OPEN state (8.1.5.)
 292                  * sOP -> sOP           Regular Ack in OPEN state
 293                  * sCR -> sIV           Waiting for Close from client
 294                  * sCG -> sCG           Ack in CLOSING MAY be processed (8.3.)
 295                  * sTW -> sIV
 296                  *
 297                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 298                         sIV, sIV, sIV, sOP, sOP, sIV, sCG, sIV
 299                 },
 300                 [DCCP_PKT_DATA] = {
 301                 /*
 302                  * sNO -> sIV           No connection
 303                  * sRQ -> sIV           No connection
 304                  * sRS -> sIV           No connection
 305                  * sPO -> sOP           Enter OPEN state (8.1.5.)
 306                  * sOP -> sOP           Regular Data packet in OPEN state
 307                  * sCR -> sIV           Waiting for Close from client
 308                  * sCG -> sCG           Data in CLOSING MAY be processed (8.3.)
 309                  * sTW -> sIV
 310                  *
 311                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 312                         sIV, sIV, sIV, sOP, sOP, sIV, sCG, sIV
 313                 },
 314                 [DCCP_PKT_DATAACK] = {
 315                 /*
 316                  * sNO -> sIV           No connection
 317                  * sRQ -> sIV           No connection
 318                  * sRS -> sIV           No connection
 319                  * sPO -> sOP           Enter OPEN state (8.1.5.)
 320                  * sOP -> sOP           Regular DataAck in OPEN state
 321                  * sCR -> sIV           Waiting for Close from client
 322                  * sCG -> sCG           Data in CLOSING MAY be processed (8.3.)
 323                  * sTW -> sIV
 324                  *
 325                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 326                         sIV, sIV, sIV, sOP, sOP, sIV, sCG, sIV
 327                 },
 328                 [DCCP_PKT_CLOSEREQ] = {
 329                 /*
 330                  * sNO -> sIV           No connection
 331                  * sRQ -> sIV           No connection
 332                  * sRS -> sIV           No connection
 333                  * sPO -> sOP -> sCR    Move directly to CLOSEREQ (8.1.5.)
 334                  * sOP -> sCR           CloseReq in OPEN state
 335                  * sCR -> sCR           Retransmit
 336                  * sCG -> sCR           Simultaneous close, client sends another Close
 337                  * sTW -> sIV           Already closed
 338                  *
 339                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 340                         sIV, sIV, sIV, sCR, sCR, sCR, sCR, sIV
 341                 },
 342                 [DCCP_PKT_CLOSE] = {
 343                 /*
 344                  * sNO -> sIV           No connection
 345                  * sRQ -> sIV           No connection
 346                  * sRS -> sIV           No connection
 347                  * sPO -> sOP -> sCG    Move direcly to CLOSING
 348                  * sOP -> sCG           Move to CLOSING
 349                  * sCR -> sIV           Close after CloseReq is invalid
 350                  * sCG -> sCG           Retransmit
 351                  * sTW -> sIV           Already closed
 352                  *
 353                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 354                         sIV, sIV, sIV, sCG, sCG, sIV, sCG, sIV
 355                 },
 356                 [DCCP_PKT_RESET] = {
 357                 /*
 358                  * sNO -> sIV           No connection
 359                  * sRQ -> sTW           Reset in response to Request
 360                  * sRS -> sTW           Timeout, SHOULD send Reset (8.1.3.)
 361                  * sPO -> sTW           Timeout, SHOULD send Reset (8.1.3.)
 362                  * sOP -> sTW
 363                  * sCR -> sTW
 364                  * sCG -> sTW
 365                  * sTW -> sIG           Ignore (don't refresh timer)
 366                  *
 367                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW, sTW */
 368                         sIV, sTW, sTW, sTW, sTW, sTW, sTW, sTW, sIG
 369                 },
 370                 [DCCP_PKT_SYNC] = {
 371                 /*
 372                  * We currently ignore Sync packets
 373                  *
 374                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 375                         sIG, sIG, sIG, sIG, sIG, sIG, sIG, sIG,
 376                 },
 377                 [DCCP_PKT_SYNCACK] = {
 378                 /*
 379                  * We currently ignore SyncAck packets
 380                  *
 381                  *      sNO, sRQ, sRS, sPO, sOP, sCR, sCG, sTW */
 382                         sIG, sIG, sIG, sIG, sIG, sIG, sIG, sIG,
 383                 },
 384         },
 385 };
 386
 387 /* this module per-net specifics */
 388 static int dccp_net_id __read_mostly;
 389 struct dccp_net {
 390         struct nf_proto_net pn;
 391         int dccp_loose;
 392         unsigned int dccp_timeout[CT_DCCP_MAX + 1];
 393 };
 394
 395 static inline struct dccp_net *dccp_pernet(struct net *net)
 396 {
 397         return net_generic(net, dccp_net_id);
 398 }
 399
 400 static bool dccp_pkt_to_tuple(const struct sk_buff *skb, unsigned int dataoff,
 401                               struct net *net, struct nf_conntrack_tuple *tuple)
 402 {
 403         struct dccp_hdr _hdr, *dh;
 404
 405         /* Actually only need first 4 bytes to get ports. */
 406         dh = skb_header_pointer(skb, dataoff, 4, &_hdr);
 407         if (dh == NULL)
 408                 return false;
 409
 410         tuple->src.u.dccp.port = dh->dccph_sport;
 411         tuple->dst.u.dccp.port = dh->dccph_dport;
 412         return true;
 413 }
 414
 415 static bool dccp_invert_tuple(struct nf_conntrack_tuple *inv,
 416                               const struct nf_conntrack_tuple *tuple)
 417 {
 418         inv->src.u.dccp.port = tuple->dst.u.dccp.port;
 419         inv->dst.u.dccp.port = tuple->src.u.dccp.port;
 420         return true;
 421 }
 422
 423 static bool dccp_new(struct nf_conn *ct, const struct sk_buff *skb,
 424                      unsigned int dataoff, unsigned int *timeouts)
 425 {
 426         struct net *net = nf_ct_net(ct);
 427         struct dccp_net *dn;
 428         struct dccp_hdr _dh, *dh;
 429         const char *msg;
 430         u_int8_t state;
 431
 432         dh = skb_header_pointer(skb, dataoff, sizeof(_dh), &_dh);
 433         BUG_ON(dh == NULL);
 434
 435         state = dccp_state_table[CT_DCCP_ROLE_CLIENT][dh->dccph_type][CT_DCCP_NONE];
 436         switch (state) {
 437         default:
 438                 dn = dccp_pernet(net);
 439                 if (dn->dccp_loose == 0) {
 440                         msg = "nf_ct_dccp: not picking up existing connection ";
 441                         goto out_invalid;
 442                 }
 443         case CT_DCCP_REQUEST:
 444                 break;
 445         case CT_DCCP_INVALID:
 446                 msg = "nf_ct_dccp: invalid state transition ";
 447                 goto out_invalid;
 448         }
 449
 450         ct->proto.dccp.role[IP_CT_DIR_ORIGINAL] = CT_DCCP_ROLE_CLIENT;
 451         ct->proto.dccp.role[IP_CT_DIR_REPLY] = CT_DCCP_ROLE_SERVER;
 452         ct->proto.dccp.state = CT_DCCP_NONE;
 453         ct->proto.dccp.last_pkt = DCCP_PKT_REQUEST;
 454         ct->proto.dccp.last_dir = IP_CT_DIR_ORIGINAL;
 455         ct->proto.dccp.handshake_seq = 0;
 456         return true;
 457
 458 out_invalid:
 459         if (LOG_INVALID(net, IPPROTO_DCCP))
 460                 nf_log_packet(net, nf_ct_l3num(ct), 0, skb, NULL, NULL,
 461                               NULL, "%s", msg);
 462         return false;
 463 }
 464
 465 static u64 dccp_ack_seq(const struct dccp_hdr *dh)
 466 {
 467         const struct dccp_hdr_ack_bits *dhack;
 468
 469         dhack = (void *)dh + __dccp_basic_hdr_len(dh);
 470         return ((u64)ntohs(dhack->dccph_ack_nr_high) << 32) +
 471                      ntohl(dhack->dccph_ack_nr_low);
 472 }
 473
 474 static unsigned int *dccp_get_timeouts(struct net *net)
 475 {
 476         return dccp_pernet(net)->dccp_timeout;
 477 }
 478
 479 static int dccp_packet(struct nf_conn *ct, const struct sk_buff *skb,
 480                        unsigned int dataoff, enum ip_conntrack_info ctinfo,
 481                        u_int8_t pf, unsigned int hooknum,
 482                        unsigned int *timeouts)
 483 {
 484         struct net *net = nf_ct_net(ct);
 485         enum ip_conntrack_dir dir = CTINFO2DIR(ctinfo);
 486         struct dccp_hdr _dh, *dh;
 487         u_int8_t type, old_state, new_state;
 488         enum ct_dccp_roles role;
 489
 490         dh = skb_header_pointer(skb, dataoff, sizeof(_dh), &_dh);
 491         BUG_ON(dh == NULL);
 492         type = dh->dccph_type;
 493
 494         if (type == DCCP_PKT_RESET &&
 495             !test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
 496                 /* Tear down connection immediately if only reply is a RESET */
 497                 nf_ct_kill_acct(ct, ctinfo, skb);
 498                 return NF_ACCEPT;
 499         }
 500
 501         spin_lock_bh(&ct->lock);
 502
 503         role = ct->proto.dccp.role[dir];
 504         old_state = ct->proto.dccp.state;
 505         new_state = dccp_state_table[role][type][old_state];
 506
 507         switch (new_state) {
 508         case CT_DCCP_REQUEST:
 509                 if (old_state == CT_DCCP_TIMEWAIT &&
 510                     role == CT_DCCP_ROLE_SERVER) {
 511                         /* Reincarnation in the reverse direction: reopen and
 512                          * reverse client/server roles. */
 513                         ct->proto.dccp.role[dir] = CT_DCCP_ROLE_CLIENT;
 514                         ct->proto.dccp.role[!dir] = CT_DCCP_ROLE_SERVER;
 515                 }
 516                 break;
 517         case CT_DCCP_RESPOND:
 518                 if (old_state == CT_DCCP_REQUEST)
 519                         ct->proto.dccp.handshake_seq = dccp_hdr_seq(dh);
 520                 break;
 521         case CT_DCCP_PARTOPEN:
 522                 if (old_state == CT_DCCP_RESPOND &&
 523                     type == DCCP_PKT_ACK &&
 524                     dccp_ack_seq(dh) == ct->proto.dccp.handshake_seq)
 525                         set_bit(IPS_ASSURED_BIT, &ct->status);
 526                 break;
 527         case CT_DCCP_IGNORE:
 528                 /*
 529                  * Connection tracking might be out of sync, so we ignore
 530                  * packets that might establish a new connection and resync
 531                  * if the server responds with a valid Response.
 532                  */
 533                 if (ct->proto.dccp.last_dir == !dir &&
 534                     ct->proto.dccp.last_pkt == DCCP_PKT_REQUEST &&
 535                     type == DCCP_PKT_RESPONSE) {
 536                         ct->proto.dccp.role[!dir] = CT_DCCP_ROLE_CLIENT;
 537                         ct->proto.dccp.role[dir] = CT_DCCP_ROLE_SERVER;
 538                         ct->proto.dccp.handshake_seq = dccp_hdr_seq(dh);
 539                         new_state = CT_DCCP_RESPOND;
 540                         break;
 541                 }
 542                 ct->proto.dccp.last_dir = dir;
 543                 ct->proto.dccp.last_pkt = type;
 544
 545                 spin_unlock_bh(&ct->lock);
 546                 if (LOG_INVALID(net, IPPROTO_DCCP))
 547                         nf_log_packet(net, pf, 0, skb, NULL, NULL, NULL,
 548                                       "nf_ct_dccp: invalid packet ignored ");
 549                 return NF_ACCEPT;
 550         case CT_DCCP_INVALID:
 551                 spin_unlock_bh(&ct->lock);
 552                 if (LOG_INVALID(net, IPPROTO_DCCP))
 553                         nf_log_packet(net, pf, 0, skb, NULL, NULL, NULL,
 554                                       "nf_ct_dccp: invalid state transition ");
 555                 return -NF_ACCEPT;
 556         }
 557
 558         ct->proto.dccp.last_dir = dir;
 559         ct->proto.dccp.last_pkt = type;
 560         ct->proto.dccp.state = new_state;
 561         spin_unlock_bh(&ct->lock);
 562
 563         if (new_state != old_state)
 564                 nf_conntrack_event_cache(IPCT_PROTOINFO, ct);
 565
 566         nf_ct_refresh_acct(ct, ctinfo, skb, timeouts[new_state]);
 567
 568         return NF_ACCEPT;
 569 }
 570
 571 static int dccp_error(struct net *net, struct nf_conn *tmpl,
 572                       struct sk_buff *skb, unsigned int dataoff,
 573                       enum ip_conntrack_info *ctinfo,
 574                       u_int8_t pf, unsigned int hooknum)
 575 {
 576         struct dccp_hdr _dh, *dh;
 577         unsigned int dccp_len = skb->len - dataoff;
 578         unsigned int cscov;
 579         const char *msg;
 580
 581         dh = skb_header_pointer(skb, dataoff, sizeof(_dh), &_dh);
 582         if (dh == NULL) {
 583                 msg = "nf_ct_dccp: short packet ";
 584                 goto out_invalid;
 585         }
 586
 587         if (dh->dccph_doff * 4 < sizeof(struct dccp_hdr) ||
 588             dh->dccph_doff * 4 > dccp_len) {
 589                 msg = "nf_ct_dccp: truncated/malformed packet ";
 590                 goto out_invalid;
 591         }
 592
 593         cscov = dccp_len;
 594         if (dh->dccph_cscov) {
 595                 cscov = (dh->dccph_cscov - 1) * 4;
 596                 if (cscov > dccp_len) {
 597                         msg = "nf_ct_dccp: bad checksum coverage ";
 598                         goto out_invalid;
 599                 }
 600         }
 601
 602         if (net->ct.sysctl_checksum && hooknum == NF_INET_PRE_ROUTING &&
 603             nf_checksum_partial(skb, hooknum, dataoff, cscov, IPPROTO_DCCP,
 604                                 pf)) {
 605                 msg = "nf_ct_dccp: bad checksum ";
 606                 goto out_invalid;
 607         }
 608
 609         if (dh->dccph_type >= DCCP_PKT_INVALID) {
 610                 msg = "nf_ct_dccp: reserved packet type ";
 611                 goto out_invalid;
 612         }
 613
 614         return NF_ACCEPT;
 615
 616 out_invalid:
 617         if (LOG_INVALID(net, IPPROTO_DCCP))
 618                 nf_log_packet(net, pf, 0, skb, NULL, NULL, NULL, "%s", msg);
 619         return -NF_ACCEPT;
 620 }
 621
 622 static void dccp_print_tuple(struct seq_file *s,
 623                              const struct nf_conntrack_tuple *tuple)
 624 {
 625         seq_printf(s, "sport=%hu dport=%hu ",
 626                    ntohs(tuple->src.u.dccp.port),
 627                    ntohs(tuple->dst.u.dccp.port));
 628 }
 629
 630 static void dccp_print_conntrack(struct seq_file *s, struct nf_conn *ct)
 631 {
 632         seq_printf(s, "%s ", dccp_state_names[ct->proto.dccp.state]);
 633 }
 634
 635 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
 636 static int dccp_to_nlattr(struct sk_buff *skb, struct nlattr *nla,
 637                           struct nf_conn *ct)
 638 {
 639         struct nlattr *nest_parms;
 640
 641         spin_lock_bh(&ct->lock);
 642         nest_parms = nla_nest_start(skb, CTA_PROTOINFO_DCCP | NLA_F_NESTED);
 643         if (!nest_parms)
 644                 goto nla_put_failure;
 645         if (nla_put_u8(skb, CTA_PROTOINFO_DCCP_STATE, ct->proto.dccp.state) ||
 646             nla_put_u8(skb, CTA_PROTOINFO_DCCP_ROLE,
 647                        ct->proto.dccp.role[IP_CT_DIR_ORIGINAL]) ||
 648             nla_put_be64(skb, CTA_PROTOINFO_DCCP_HANDSHAKE_SEQ,
 649                          cpu_to_be64(ct->proto.dccp.handshake_seq),
 650                          CTA_PROTOINFO_DCCP_PAD))
 651                 goto nla_put_failure;
 652         nla_nest_end(skb, nest_parms);
 653         spin_unlock_bh(&ct->lock);
 654         return 0;
 655
 656 nla_put_failure:
 657         spin_unlock_bh(&ct->lock);
 658         return -1;
 659 }
 660
 661 static const struct nla_policy dccp_nla_policy[CTA_PROTOINFO_DCCP_MAX + 1] = {
 662         [CTA_PROTOINFO_DCCP_STATE]      = { .type = NLA_U8 },
 663         [CTA_PROTOINFO_DCCP_ROLE]       = { .type = NLA_U8 },
 664         [CTA_PROTOINFO_DCCP_HANDSHAKE_SEQ] = { .type = NLA_U64 },
 665         [CTA_PROTOINFO_DCCP_PAD]        = { .type = NLA_UNSPEC },
 666 };
 667
 668 static int nlattr_to_dccp(struct nlattr *cda[], struct nf_conn *ct)
 669 {
 670         struct nlattr *attr = cda[CTA_PROTOINFO_DCCP];
 671         struct nlattr *tb[CTA_PROTOINFO_DCCP_MAX + 1];
 672         int err;
 673
 674         if (!attr)
 675                 return 0;
 676
 677         err = nla_parse_nested(tb, CTA_PROTOINFO_DCCP_MAX, attr,
 678                                dccp_nla_policy);
 679         if (err < 0)
 680                 return err;
 681
 682         if (!tb[CTA_PROTOINFO_DCCP_STATE] ||
 683             !tb[CTA_PROTOINFO_DCCP_ROLE] ||
 684             nla_get_u8(tb[CTA_PROTOINFO_DCCP_ROLE]) > CT_DCCP_ROLE_MAX ||
 685             nla_get_u8(tb[CTA_PROTOINFO_DCCP_STATE]) >= CT_DCCP_IGNORE) {
 686                 return -EINVAL;
 687         }
 688
 689         spin_lock_bh(&ct->lock);
 690         ct->proto.dccp.state = nla_get_u8(tb[CTA_PROTOINFO_DCCP_STATE]);
 691         if (nla_get_u8(tb[CTA_PROTOINFO_DCCP_ROLE]) == CT_DCCP_ROLE_CLIENT) {
 692                 ct->proto.dccp.role[IP_CT_DIR_ORIGINAL] = CT_DCCP_ROLE_CLIENT;
 693                 ct->proto.dccp.role[IP_CT_DIR_REPLY] = CT_DCCP_ROLE_SERVER;
 694         } else {
 695                 ct->proto.dccp.role[IP_CT_DIR_ORIGINAL] = CT_DCCP_ROLE_SERVER;
 696                 ct->proto.dccp.role[IP_CT_DIR_REPLY] = CT_DCCP_ROLE_CLIENT;
 697         }
 698         if (tb[CTA_PROTOINFO_DCCP_HANDSHAKE_SEQ]) {
 699                 ct->proto.dccp.handshake_seq =
 700                 be64_to_cpu(nla_get_be64(tb[CTA_PROTOINFO_DCCP_HANDSHAKE_SEQ]));
 701         }
 702         spin_unlock_bh(&ct->lock);
 703         return 0;
 704 }
 705
 706 static int dccp_nlattr_size(void)
 707 {
 708         return nla_total_size(0)        /* CTA_PROTOINFO_DCCP */
 709                 + nla_policy_len(dccp_nla_policy, CTA_PROTOINFO_DCCP_MAX + 1);
 710 }
 711
 712 #endif
 713
 714 #if IS_ENABLED(CONFIG_NF_CT_NETLINK_TIMEOUT)
 715
 716 #include <linux/netfilter/nfnetlink.h>
 717 #include <linux/netfilter/nfnetlink_cttimeout.h>
 718
 719 static int dccp_timeout_nlattr_to_obj(struct nlattr *tb[],
 720                                       struct net *net, void *data)
 721 {
 722         struct dccp_net *dn = dccp_pernet(net);
 723         unsigned int *timeouts = data;
 724         int i;
 725
 726         /* set default DCCP timeouts. */
 727         for (i=0; i<CT_DCCP_MAX; i++)
 728                 timeouts[i] = dn->dccp_timeout[i];
 729
 730         /* there's a 1:1 mapping between attributes and protocol states. */
 731         for (i=CTA_TIMEOUT_DCCP_UNSPEC+1; i<CTA_TIMEOUT_DCCP_MAX+1; i++) {
 732                 if (tb[i]) {
 733                         timeouts[i] = ntohl(nla_get_be32(tb[i])) * HZ;
 734                 }
 735         }
 736         return 0;
 737 }
 738
 739 static int
 740 dccp_timeout_obj_to_nlattr(struct sk_buff *skb, const void *data)
 741 {
 742         const unsigned int *timeouts = data;
 743         int i;
 744
 745         for (i=CTA_TIMEOUT_DCCP_UNSPEC+1; i<CTA_TIMEOUT_DCCP_MAX+1; i++) {
 746                 if (nla_put_be32(skb, i, htonl(timeouts[i] / HZ)))
 747                         goto nla_put_failure;
 748         }
 749         return 0;
 750
 751 nla_put_failure:
 752         return -ENOSPC;
 753 }
 754
 755 static const struct nla_policy
 756 dccp_timeout_nla_policy[CTA_TIMEOUT_DCCP_MAX+1] = {
 757         [CTA_TIMEOUT_DCCP_REQUEST]      = { .type = NLA_U32 },
 758         [CTA_TIMEOUT_DCCP_RESPOND]      = { .type = NLA_U32 },
 759         [CTA_TIMEOUT_DCCP_PARTOPEN]     = { .type = NLA_U32 },
 760         [CTA_TIMEOUT_DCCP_OPEN]         = { .type = NLA_U32 },
 761         [CTA_TIMEOUT_DCCP_CLOSEREQ]     = { .type = NLA_U32 },
 762         [CTA_TIMEOUT_DCCP_CLOSING]      = { .type = NLA_U32 },
 763         [CTA_TIMEOUT_DCCP_TIMEWAIT]     = { .type = NLA_U32 },
 764 };
 765 #endif /* CONFIG_NF_CT_NETLINK_TIMEOUT */
 766
 767 #ifdef CONFIG_SYSCTL
 768 /* template, data assigned later */
 769 static struct ctl_table dccp_sysctl_table[] = {
 770         {
 771                 .procname       = "nf_conntrack_dccp_timeout_request",
 772                 .maxlen         = sizeof(unsigned int),
 773                 .mode           = 0644,
 774                 .proc_handler   = proc_dointvec_jiffies,
 775         },
 776         {
 777                 .procname       = "nf_conntrack_dccp_timeout_respond",
 778                 .maxlen         = sizeof(unsigned int),
 779                 .mode           = 0644,
 780                 .proc_handler   = proc_dointvec_jiffies,
 781         },
 782         {
 783                 .procname       = "nf_conntrack_dccp_timeout_partopen",
 784                 .maxlen         = sizeof(unsigned int),
 785                 .mode           = 0644,
 786                 .proc_handler   = proc_dointvec_jiffies,
 787         },
 788         {
 789                 .procname       = "nf_conntrack_dccp_timeout_open",
 790                 .maxlen         = sizeof(unsigned int),
 791                 .mode           = 0644,
 792                 .proc_handler   = proc_dointvec_jiffies,
 793         },
 794         {
 795                 .procname       = "nf_conntrack_dccp_timeout_closereq",
 796                 .maxlen         = sizeof(unsigned int),
 797                 .mode           = 0644,
 798                 .proc_handler   = proc_dointvec_jiffies,
 799         },
 800         {
 801                 .procname       = "nf_conntrack_dccp_timeout_closing",
 802                 .maxlen         = sizeof(unsigned int),
 803                 .mode           = 0644,
 804                 .proc_handler   = proc_dointvec_jiffies,
 805         },
 806         {
 807                 .procname       = "nf_conntrack_dccp_timeout_timewait",
 808                 .maxlen         = sizeof(unsigned int),
 809                 .mode           = 0644,
 810                 .proc_handler   = proc_dointvec_jiffies,
 811         },
 812         {
 813                 .procname       = "nf_conntrack_dccp_loose",
 814                 .maxlen         = sizeof(int),
 815                 .mode           = 0644,
 816                 .proc_handler   = proc_dointvec,
 817         },
 818         { }
 819 };
 820 #endif /* CONFIG_SYSCTL */
 821
 822 static int dccp_kmemdup_sysctl_table(struct net *net, struct nf_proto_net *pn,
 823                                      struct dccp_net *dn)
 824 {
 825 #ifdef CONFIG_SYSCTL
 826         if (pn->ctl_table)
 827                 return 0;
 828
 829         pn->ctl_table = kmemdup(dccp_sysctl_table,
 830                                 sizeof(dccp_sysctl_table),
 831                                 GFP_KERNEL);
 832         if (!pn->ctl_table)
 833                 return -ENOMEM;
 834
 835         pn->ctl_table[0].data = &dn->dccp_timeout[CT_DCCP_REQUEST];
 836         pn->ctl_table[1].data = &dn->dccp_timeout[CT_DCCP_RESPOND];
 837         pn->ctl_table[2].data = &dn->dccp_timeout[CT_DCCP_PARTOPEN];
 838         pn->ctl_table[3].data = &dn->dccp_timeout[CT_DCCP_OPEN];
 839         pn->ctl_table[4].data = &dn->dccp_timeout[CT_DCCP_CLOSEREQ];
 840         pn->ctl_table[5].data = &dn->dccp_timeout[CT_DCCP_CLOSING];
 841         pn->ctl_table[6].data = &dn->dccp_timeout[CT_DCCP_TIMEWAIT];
 842         pn->ctl_table[7].data = &dn->dccp_loose;
 843
 844         /* Don't export sysctls to unprivileged users */
 845         if (net->user_ns != &init_user_ns)
 846                 pn->ctl_table[0].procname = NULL;
 847 #endif
 848         return 0;
 849 }
 850
 851 static int dccp_init_net(struct net *net, u_int16_t proto)
 852 {
 853         struct dccp_net *dn = dccp_pernet(net);
 854         struct nf_proto_net *pn = &dn->pn;
 855
 856         if (!pn->users) {
 857                 /* default values */
 858                 dn->dccp_loose = 1;
 859                 dn->dccp_timeout[CT_DCCP_REQUEST]       = 2 * DCCP_MSL;
 860                 dn->dccp_timeout[CT_DCCP_RESPOND]       = 4 * DCCP_MSL;
 861                 dn->dccp_timeout[CT_DCCP_PARTOPEN]      = 4 * DCCP_MSL;
 862                 dn->dccp_timeout[CT_DCCP_OPEN]          = 12 * 3600 * HZ;
 863                 dn->dccp_timeout[CT_DCCP_CLOSEREQ]      = 64 * HZ;
 864                 dn->dccp_timeout[CT_DCCP_CLOSING]       = 64 * HZ;
 865                 dn->dccp_timeout[CT_DCCP_TIMEWAIT]      = 2 * DCCP_MSL;
 866         }
 867
 868         return dccp_kmemdup_sysctl_table(net, pn, dn);
 869 }
 870
 871 static struct nf_conntrack_l4proto dccp_proto4 __read_mostly = {
 872         .l3proto                = AF_INET,
 873         .l4proto                = IPPROTO_DCCP,
 874         .name                   = "dccp",
 875         .pkt_to_tuple           = dccp_pkt_to_tuple,
 876         .invert_tuple           = dccp_invert_tuple,
 877         .new                    = dccp_new,
 878         .packet                 = dccp_packet,
 879         .get_timeouts           = dccp_get_timeouts,
 880         .error                  = dccp_error,
 881         .print_tuple            = dccp_print_tuple,
 882         .print_conntrack        = dccp_print_conntrack,
 883 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
 884         .to_nlattr              = dccp_to_nlattr,
 885         .nlattr_size            = dccp_nlattr_size,
 886         .from_nlattr            = nlattr_to_dccp,
 887         .tuple_to_nlattr        = nf_ct_port_tuple_to_nlattr,
 888         .nlattr_tuple_size      = nf_ct_port_nlattr_tuple_size,
 889         .nlattr_to_tuple        = nf_ct_port_nlattr_to_tuple,
 890         .nla_policy             = nf_ct_port_nla_policy,
 891 #endif
 892 #if IS_ENABLED(CONFIG_NF_CT_NETLINK_TIMEOUT)
 893         .ctnl_timeout           = {
 894                 .nlattr_to_obj  = dccp_timeout_nlattr_to_obj,
 895                 .obj_to_nlattr  = dccp_timeout_obj_to_nlattr,
 896                 .nlattr_max     = CTA_TIMEOUT_DCCP_MAX,
 897                 .obj_size       = sizeof(unsigned int) * CT_DCCP_MAX,
 898                 .nla_policy     = dccp_timeout_nla_policy,
 899         },
 900 #endif /* CONFIG_NF_CT_NETLINK_TIMEOUT */
 901         .net_id                 = &dccp_net_id,
 902         .init_net               = dccp_init_net,
 903 };
 904
 905 static struct nf_conntrack_l4proto dccp_proto6 __read_mostly = {
 906         .l3proto                = AF_INET6,
 907         .l4proto                = IPPROTO_DCCP,
 908         .name                   = "dccp",
 909         .pkt_to_tuple           = dccp_pkt_to_tuple,
 910         .invert_tuple           = dccp_invert_tuple,
 911         .new                    = dccp_new,
 912         .packet                 = dccp_packet,
 913         .get_timeouts           = dccp_get_timeouts,
 914         .error                  = dccp_error,
 915         .print_tuple            = dccp_print_tuple,
 916         .print_conntrack        = dccp_print_conntrack,
 917 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
 918         .to_nlattr              = dccp_to_nlattr,
 919         .nlattr_size            = dccp_nlattr_size,
 920         .from_nlattr            = nlattr_to_dccp,
 921         .tuple_to_nlattr        = nf_ct_port_tuple_to_nlattr,
 922         .nlattr_tuple_size      = nf_ct_port_nlattr_tuple_size,
 923         .nlattr_to_tuple        = nf_ct_port_nlattr_to_tuple,
 924         .nla_policy             = nf_ct_port_nla_policy,
 925 #endif
 926 #if IS_ENABLED(CONFIG_NF_CT_NETLINK_TIMEOUT)
 927         .ctnl_timeout           = {
 928                 .nlattr_to_obj  = dccp_timeout_nlattr_to_obj,
 929                 .obj_to_nlattr  = dccp_timeout_obj_to_nlattr,
 930                 .nlattr_max     = CTA_TIMEOUT_DCCP_MAX,
 931                 .obj_size       = sizeof(unsigned int) * CT_DCCP_MAX,
 932                 .nla_policy     = dccp_timeout_nla_policy,
 933         },
 934 #endif /* CONFIG_NF_CT_NETLINK_TIMEOUT */
 935         .net_id                 = &dccp_net_id,
 936         .init_net               = dccp_init_net,
 937 };
 938
 939 static __net_init int dccp_net_init(struct net *net)
 940 {
 941         int ret = 0;
 942         ret = nf_ct_l4proto_pernet_register(net, &dccp_proto4);
 943         if (ret < 0) {
 944                 pr_err("nf_conntrack_dccp4: pernet registration failed.\n");
 945                 goto out;
 946         }
 947         ret = nf_ct_l4proto_pernet_register(net, &dccp_proto6);
 948         if (ret < 0) {
 949                 pr_err("nf_conntrack_dccp6: pernet registration failed.\n");
 950                 goto cleanup_dccp4;
 951         }
 952         return 0;
 953 cleanup_dccp4:
 954         nf_ct_l4proto_pernet_unregister(net, &dccp_proto4);
 955 out:
 956         return ret;
 957 }
 958
 959 static __net_exit void dccp_net_exit(struct net *net)
 960 {
 961         nf_ct_l4proto_pernet_unregister(net, &dccp_proto6);
 962         nf_ct_l4proto_pernet_unregister(net, &dccp_proto4);
 963 }
 964
 965 static struct pernet_operations dccp_net_ops = {
 966         .init = dccp_net_init,
 967         .exit = dccp_net_exit,
 968         .id   = &dccp_net_id,
 969         .size = sizeof(struct dccp_net),
 970 };
 971
 972 static int __init nf_conntrack_proto_dccp_init(void)
 973 {
 974         int ret;
 975
 976         ret = register_pernet_subsys(&dccp_net_ops);
 977         if (ret < 0)
 978                 goto out_pernet;
 979
 980         ret = nf_ct_l4proto_register(&dccp_proto4);
 981         if (ret < 0)
 982                 goto out_dccp4;
 983
 984         ret = nf_ct_l4proto_register(&dccp_proto6);
 985         if (ret < 0)
 986                 goto out_dccp6;
 987
 988         return 0;
 989 out_dccp6:
 990         nf_ct_l4proto_unregister(&dccp_proto4);
 991 out_dccp4:
 992         unregister_pernet_subsys(&dccp_net_ops);
 993 out_pernet:
 994         return ret;
 995 }
 996
 997 static void __exit nf_conntrack_proto_dccp_fini(void)
 998 {
 999         nf_ct_l4proto_unregister(&dccp_proto6);
1000         nf_ct_l4proto_unregister(&dccp_proto4);
1001         unregister_pernet_subsys(&dccp_net_ops);
1002 }
1003
1004 module_init(nf_conntrack_proto_dccp_init);
1005 module_exit(nf_conntrack_proto_dccp_fini);
1006
1007 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
1008 MODULE_DESCRIPTION("DCCP connection tracking protocol helper");
1009 MODULE_LICENSE("GPL");