KernelLand/Modules/IPStack/tcp.c

   1 /*
   2  * Acess2 IP Stack
   3  * - TCP Handling
   4  */
   5 #define DEBUG   0
   6 #include "ipstack.h"
   7 #include "ipv4.h"
   8 #include "ipv6.h"
   9 #include "tcp.h"
  10
  11 #define HEXDUMP_INCOMING        0
  12 #define HEXDUMP_OUTGOING        0
  13
  14 #define TCP_MIN_DYNPORT 0xC000
  15 #define TCP_MAX_HALFOPEN        1024    // Should be enough
  16
  17 #define TCP_MAX_PACKET_SIZE     1024
  18 #define TCP_WINDOW_SIZE 0x2000
  19 #define TCP_RECIEVE_BUFFER_SIZE 0x8000
  20 #define TCP_DACK_THRESHOLD      4096
  21 #define TCP_DACK_TIMEOUT        500
  22
  23 #define TCP_DEBUG       0       // Set to non-0 to enable TCP packet logging
  24
  25 // === PROTOTYPES ===
  26 void    TCP_Initialise(void);
  27 void    TCP_StartConnection(tTCPConnection *Conn);
  28 void    TCP_SendPacket(tTCPConnection *Conn, tTCPHeader *Header, size_t DataLen, const void *Data);
  29 void    TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer);
  30 void    TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length);
  31 int     TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length);
  32 void    TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection);
  33 void    TCP_int_SendDelayedACK(void *ConnPtr);
  34 void    TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason);
  35 Uint16  TCP_GetUnusedPort();
  36  int    TCP_AllocatePort(Uint16 Port);
  37  int    TCP_DeallocatePort(Uint16 Port);
  38 tTCPConnection  *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State);
  39 // --- Server
  40 tVFS_Node       *TCP_Server_Init(tInterface *Interface);
  41  int    TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Name[FILENAME_MAX]);
  42 tVFS_Node       *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags);
  43  int    TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data);
  44 void    TCP_Server_Close(tVFS_Node *Node);
  45 // --- Client
  46 tVFS_Node       *TCP_Client_Init(tInterface *Interface);
  47 size_t  TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags);
  48 size_t  TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags);
  49  int    TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data);
  50 void    TCP_Client_Close(tVFS_Node *Node);
  51 // --- Helpers
  52  int    WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue);
  53
  54 // === TEMPLATES ===
  55 tSocketFile     gTCP_ServerFile = {NULL, "tcps", TCP_Server_Init};
  56 tSocketFile     gTCP_ClientFile = {NULL, "tcpc", TCP_Client_Init};
  57 tVFS_NodeType   gTCP_ServerNodeType = {
  58         .TypeName = "TCP Server",
  59         .ReadDir = TCP_Server_ReadDir,
  60         .FindDir = TCP_Server_FindDir,
  61         .IOCtl   = TCP_Server_IOCtl,
  62         .Close   = TCP_Server_Close
  63         };
  64 tVFS_NodeType   gTCP_ClientNodeType = {
  65         .TypeName = "TCP Client/Connection",
  66         .Read  = TCP_Client_Read,
  67         .Write = TCP_Client_Write,
  68         .IOCtl = TCP_Client_IOCtl,
  69         .Close = TCP_Client_Close
  70         };
  71
  72 // === GLOBALS ===
  73  int    giTCP_NumHalfopen = 0;
  74 tShortSpinlock  glTCP_Listeners;
  75 tTCPListener    *gTCP_Listeners;
  76 tShortSpinlock  glTCP_OutbountCons;
  77 tTCPConnection  *gTCP_OutbountCons;
  78 Uint32  gaTCP_PortBitmap[0x800];
  79  int    giTCP_NextOutPort = TCP_MIN_DYNPORT;
  80
  81 // === CODE ===
  82 /**
  83  * \brief Initialise the TCP Layer
  84  *
  85  * Registers the client and server files and the GetPacket callback
  86  */
  87 void TCP_Initialise(void)
  88 {
  89         giTCP_NextOutPort += rand()%128;
  90         IPStack_AddFile(&gTCP_ServerFile);
  91         IPStack_AddFile(&gTCP_ClientFile);
  92         IPv4_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  93         IPv6_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  94 }
  95
  96 /**
  97  * \brief Sends a packet from the specified connection, calculating the checksums
  98  * \param Conn  Connection
  99  * \param Length        Length of data
 100  * \param Data  Packet data (cast as a TCP Header)
 101  */
 102 void TCP_SendPacket( tTCPConnection *Conn, tTCPHeader *Header, size_t Length, const void *Data )
 103 {
 104         tIPStackBuffer  *buffer;
 105         Uint16  checksum[3];
 106          int    packlen = sizeof(*Header) + Length;
 107
 108         buffer = IPStack_Buffer_CreateBuffer(2 + IPV4_BUFFERS);
 109         if( Data && Length )
 110                 IPStack_Buffer_AppendSubBuffer(buffer, Length, 0, Data, NULL, NULL);
 111         IPStack_Buffer_AppendSubBuffer(buffer, sizeof(*Header), 0, Header, NULL, NULL);
 112
 113         LOG("Sending %i+%i to %s:%i", sizeof(*Header), Length,
 114                 IPStack_PrintAddress(Conn->Interface->Type, &Conn->RemoteIP),
 115                 Conn->RemotePort
 116                 );
 117
 118         Header->Checksum = 0;
 119         checksum[1] = htons( ~IPv4_Checksum(Header, sizeof(tTCPHeader)) );
 120         checksum[2] = htons( ~IPv4_Checksum(Data, Length) );
 121
 122         // TODO: Fragment packet
 123
 124         switch( Conn->Interface->Type )
 125         {
 126         case 4:
 127                 // Get IPv4 pseudo-header checksum
 128                 {
 129                         Uint32  buf[3];
 130                         buf[0] = ((tIPv4*)Conn->Interface->Address)->L;
 131                         buf[1] = Conn->RemoteIP.v4.L;
 132                         buf[2] = (htons(packlen)<<16) | (6<<8) | 0;
 133                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 134                 }
 135                 // - Combine checksums
 136                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );
 137                 IPv4_SendPacket(Conn->Interface, Conn->RemoteIP.v4, IP4PROT_TCP, 0, buffer);
 138                 break;
 139
 140         case 6:
 141                 // Append IPv6 Pseudo Header
 142                 {
 143                         Uint32  buf[4+4+1+1];
 144                         memcpy(buf, Conn->Interface->Address, 16);
 145                         memcpy(&buf[4], &Conn->RemoteIP, 16);
 146                         buf[8] = htonl(packlen);
 147                         buf[9] = htonl(6);
 148                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 149                 }
 150                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );  // Combine the two
 151                 IPv6_SendPacket(Conn->Interface, Conn->RemoteIP.v6, IP4PROT_TCP, buffer);
 152                 break;
 153         }
 154 }
 155
 156 /**
 157  * \brief Handles a packet from the IP Layer
 158  * \param Interface     Interface the packet arrived from
 159  * \param Address       Pointer to the addres structure
 160  * \param Length        Size of packet in bytes
 161  * \param Buffer        Packet data
 162  */
 163 void TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer)
 164 {
 165         tTCPHeader      *hdr = Buffer;
 166         tTCPListener    *srv;
 167         tTCPConnection  *conn;
 168
 169         #if TCP_DEBUG
 170         Log_Log("TCP", "TCP_GetPacket: <Local>:%i from [%s]:%i, Flags = %s%s%s%s%s%s%s%s",
 171                 ntohs(hdr->DestPort),
 172                 IPStack_PrintAddress(Interface->Type, Address),
 173                 ntohs(hdr->SourcePort),
 174                 (hdr->Flags & TCP_FLAG_CWR) ? "CWR " : "",
 175                 (hdr->Flags & TCP_FLAG_ECE) ? "ECE " : "",
 176                 (hdr->Flags & TCP_FLAG_URG) ? "URG " : "",
 177                 (hdr->Flags & TCP_FLAG_ACK) ? "ACK " : "",
 178                 (hdr->Flags & TCP_FLAG_PSH) ? "PSH " : "",
 179                 (hdr->Flags & TCP_FLAG_RST) ? "RST " : "",
 180                 (hdr->Flags & TCP_FLAG_SYN) ? "SYN " : "",
 181                 (hdr->Flags & TCP_FLAG_FIN) ? "FIN " : ""
 182                 );
 183         #endif
 184
 185         if( Length > (hdr->DataOffset >> 4)*4 )
 186         {
 187                 LOG("SequenceNumber = 0x%x", ntohl(hdr->SequenceNumber));
 188 #if HEXDUMP_INCOMING
 189                 Debug_HexDump(
 190                         "TCP_GetPacket: Packet Data = ",
 191                         (Uint8*)hdr + (hdr->DataOffset >> 4)*4,
 192                         Length - (hdr->DataOffset >> 4)*4
 193                         );
 194 #endif
 195         }
 196
 197         // Check Servers
 198         for( srv = gTCP_Listeners; srv; srv = srv->Next )
 199         {
 200                 // Check if the server is active
 201                 if(srv->Port == 0)      continue;
 202                 // Check the interface
 203                 if(srv->Interface && srv->Interface != Interface)       continue;
 204                 // Check the destination port
 205                 if(srv->Port != htons(hdr->DestPort))   continue;
 206
 207                 Log_Log("TCP", "TCP_GetPacket: Matches server %p", srv);
 208                 // Is this in an established connection?
 209                 for( conn = srv->Connections; conn; conn = conn->Next )
 210                 {
 211                         // Check that it is coming in on the same interface
 212                         if(conn->Interface != Interface)        continue;
 213
 214                         // Check Source Port
 215                         Log_Log("TCP", "TCP_GetPacket: conn->RemotePort(%i) == hdr->SourcePort(%i)",
 216                                 conn->RemotePort, ntohs(hdr->SourcePort));
 217                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 218
 219                         // Check Source IP
 220                         Log_Debug("TCP", "TCP_GetPacket: conn->RemoteIP(%s)",
 221                                 IPStack_PrintAddress(conn->Interface->Type, &conn->RemoteIP));
 222                         Log_Debug("TCP", "                == Address(%s)",
 223                                 IPStack_PrintAddress(conn->Interface->Type, Address));
 224                         if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 225                                 continue ;
 226
 227                         Log_Log("TCP", "TCP_GetPacket: Matches connection %p", conn);
 228                         // We have a response!
 229                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 230
 231                         return;
 232                 }
 233
 234                 Log_Log("TCP", "TCP_GetPacket: Opening Connection");
 235                 // Open a new connection (well, check that it's a SYN)
 236                 if(hdr->Flags != TCP_FLAG_SYN) {
 237                         Log_Log("TCP", "TCP_GetPacket: Packet is not a SYN");
 238                         return ;
 239                 }
 240
 241                 // TODO: Check for halfopen max
 242
 243                 conn = TCP_int_CreateConnection(Interface, TCP_ST_SYN_RCVD);
 244                 conn->LocalPort = srv->Port;
 245                 conn->RemotePort = ntohs(hdr->SourcePort);
 246
 247                 switch(Interface->Type)
 248                 {
 249                 case 4: conn->RemoteIP.v4 = *(tIPv4*)Address;   break;
 250                 case 6: conn->RemoteIP.v6 = *(tIPv6*)Address;   break;
 251                 }
 252
 253                 conn->NextSequenceRcv = ntohl( hdr->SequenceNumber ) + 1;
 254                 conn->HighestSequenceRcvd = conn->NextSequenceRcv;
 255                 conn->NextSequenceSend = rand();
 256
 257                 conn->Node.ImplInt = srv->NextID ++;
 258
 259                 // Hmm... Theoretically, this lock will never have to wait,
 260                 // as the interface is locked to the watching thread, and this
 261                 // runs in the watching thread. But, it's a good idea to have
 262                 // it, just in case
 263                 // Oh, wait, there is a case where a wildcard can be used
 264                 // (srv->Interface == NULL) so having the lock is a good idea
 265                 SHORTLOCK(&srv->lConnections);
 266                 if( !srv->Connections )
 267                         srv->Connections = conn;
 268                 else
 269                         srv->ConnectionsTail->Next = conn;
 270                 srv->ConnectionsTail = conn;
 271                 if(!srv->NewConnections)
 272                         srv->NewConnections = conn;
 273                 VFS_MarkAvaliable( &srv->Node, 1 );
 274                 SHORTREL(&srv->lConnections);
 275                 Semaphore_Signal(&srv->WaitingConnections, 1);
 276
 277                 // Send the SYN ACK
 278                 hdr->Flags |= TCP_FLAG_ACK;
 279                 hdr->AcknowlegementNumber = htonl(conn->NextSequenceRcv);
 280                 hdr->SequenceNumber = htonl(conn->NextSequenceSend);
 281                 hdr->DestPort = hdr->SourcePort;
 282                 hdr->SourcePort = htons(srv->Port);
 283                 hdr->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 284                 TCP_SendPacket( conn, hdr, 0, NULL );
 285                 conn->NextSequenceSend ++;
 286                 return ;
 287         }
 288
 289         // Check Open Connections
 290         {
 291                 for( conn = gTCP_OutbountCons; conn; conn = conn->Next )
 292                 {
 293                         // Check that it is coming in on the same interface
 294                         if(conn->Interface != Interface)        continue;
 295
 296                         // Check Source Port
 297                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 298
 299                         // Check Source IP
 300                         if(conn->Interface->Type == 6 && !IP6_EQU(conn->RemoteIP.v6, *(tIPv6*)Address))
 301                                 continue;
 302                         if(conn->Interface->Type == 4 && !IP4_EQU(conn->RemoteIP.v4, *(tIPv4*)Address))
 303                                 continue;
 304
 305                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 306                         return ;
 307                 }
 308         }
 309
 310         Log_Log("TCP", "TCP_GetPacket: No Match");
 311 }
 312
 313 /**
 314  * \brief Handles a packet sent to a specific connection
 315  * \param Connection    TCP Connection pointer
 316  * \param Header        TCP Packet pointer
 317  * \param Length        Length of the packet
 318  */
 319 void TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length)
 320 {
 321          int    dataLen;
 322         Uint32  sequence_num;
 323
 324         // Silently drop once finished
 325         // TODO: Check if this needs to be here
 326         if( Connection->State == TCP_ST_FINISHED ) {
 327                 Log_Log("TCP", "Packet ignored - connection finnished");
 328                 return ;
 329         }
 330
 331         // Syncronise sequence values
 332         if(Header->Flags & TCP_FLAG_SYN) {
 333                 // TODO: What if the packet also has data?
 334                 if( Connection->LastACKSequence != Connection->NextSequenceRcv )
 335                         TCP_INT_SendACK(Connection, "SYN");
 336                 Connection->NextSequenceRcv = ntohl(Header->SequenceNumber);
 337                 // TODO: Process HighestSequenceRcvd
 338                 // HACK!
 339                 if( Connection->HighestSequenceRcvd == 0 )
 340                         Connection->HighestSequenceRcvd = Connection->NextSequenceRcv;
 341                 Connection->LastACKSequence = Connection->NextSequenceRcv;
 342         }
 343
 344         // Ackowledge a sent packet
 345         if(Header->Flags & TCP_FLAG_ACK) {
 346                 // TODO: Process an ACKed Packet
 347                 LOG("Conn %p, Sent packet 0x%x ACKed", Connection, Header->AcknowlegementNumber);
 348         }
 349
 350         // Get length of data
 351         dataLen = Length - (Header->DataOffset>>4)*4;
 352         LOG("dataLen = %i", dataLen);
 353         #if TCP_DEBUG
 354         Log_Debug("TCP", "State %i, dataLen = %x", Connection->State, dataLen);
 355         #endif
 356
 357         //
 358         // State Machine
 359         //
 360         switch( Connection->State )
 361         {
 362         // Pre-init connection?
 363         case TCP_ST_CLOSED:
 364                 Log_Log("TCP", "Packets to a closed connection?!");
 365                 break;
 366
 367         // --- Init States ---
 368         // SYN sent, expecting SYN-ACK Connection Opening
 369         case TCP_ST_SYN_SENT:
 370                 if( Header->Flags & TCP_FLAG_SYN )
 371                 {
 372                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 373                                 Connection->HighestSequenceRcvd ++;
 374                         Connection->NextSequenceRcv ++;
 375
 376                         if( Header->Flags & TCP_FLAG_ACK )
 377                         {
 378                                 Log_Log("TCP", "ACKing SYN-ACK");
 379                                 Connection->State = TCP_ST_OPEN;
 380                                 VFS_MarkFull(&Connection->Node, 0);
 381                         }
 382                         else
 383                         {
 384                                 Log_Log("TCP", "ACKing SYN");
 385                                 Connection->State = TCP_ST_SYN_RCVD;
 386                         }
 387                         Header->DestPort = Header->SourcePort;
 388                         Header->SourcePort = htons(Connection->LocalPort);
 389                         Header->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 390                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 391                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 392                         Header->Flags = TCP_FLAG_ACK;
 393                         Header->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 394                         TCP_SendPacket( Connection, Header, 0, NULL );
 395                 }
 396                 break;
 397
 398         // SYN-ACK sent, expecting ACK
 399         case TCP_ST_SYN_RCVD:
 400                 if( Header->Flags & TCP_FLAG_ACK )
 401                 {
 402                         // TODO: Handle max half-open limit
 403                         Log_Log("TCP", "Connection fully opened");
 404                         Connection->State = TCP_ST_OPEN;
 405                         VFS_MarkFull(&Connection->Node, 0);
 406                 }
 407                 break;
 408
 409         // --- Established State ---
 410         case TCP_ST_OPEN:
 411                 // - Handle State changes
 412                 //
 413                 if( Header->Flags & TCP_FLAG_FIN ) {
 414                         Log_Log("TCP", "Conn %p closed, recieved FIN", Connection);
 415                         VFS_MarkError(&Connection->Node, 1);
 416                         Connection->State = TCP_ST_CLOSE_WAIT;
 417 //                      Header->Flags &= ~TCP_FLAG_FIN;
 418                         // CLOSE WAIT requires the client to close (or does it?)
 419                         #if 0
 420
 421                         #endif
 422                 }
 423
 424                 // Check for an empty packet
 425                 if(dataLen == 0) {
 426                         if( Header->Flags == TCP_FLAG_ACK )
 427                         {
 428                                 Log_Log("TCP", "ACK only packet");
 429                                 return ;
 430                         }
 431                         // TODO: Is this right? (empty packet counts as one byte)
 432                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 433                                 Connection->HighestSequenceRcvd ++;
 434                         Connection->NextSequenceRcv ++;
 435                         Log_Log("TCP", "Empty Packet, inc and ACK the current sequence number");
 436                         TCP_INT_SendACK(Connection, "Empty");
 437                         return ;
 438                 }
 439
 440                 // NOTES:
 441                 // Flags
 442                 //    PSH - Has Data?
 443                 // /NOTES
 444
 445                 sequence_num = ntohl(Header->SequenceNumber);
 446
 447                 LOG("0x%08x <= 0x%08x < 0x%08x",
 448                         Connection->NextSequenceRcv,
 449                         ntohl(Header->SequenceNumber),
 450                         Connection->NextSequenceRcv + TCP_WINDOW_SIZE
 451                         );
 452
 453                 // Is this packet the next expected packet?
 454                 if( sequence_num == Connection->NextSequenceRcv )
 455                 {
 456                          int    rv;
 457                         // Ooh, Goodie! Add it to the recieved list
 458                         rv = TCP_INT_AppendRecieved(Connection,
 459                                 (Uint8*)Header + (Header->DataOffset>>4)*4,
 460                                 dataLen
 461                                 );
 462                         if(rv != 0) {
 463                                 Log_Notice("TCP", "TCP_INT_AppendRecieved rv %i", rv);
 464                                 break;
 465                         }
 466                         LOG("0x%08x += %i", Connection->NextSequenceRcv, dataLen);
 467                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 468                                 Connection->HighestSequenceRcvd += dataLen;
 469                         Connection->NextSequenceRcv += dataLen;
 470
 471                         // TODO: This should be moved out of the watcher thread,
 472                         // so that a single lost packet on one connection doesn't cause
 473                         // all connections on the interface to lag.
 474                         // - Meh, no real issue, as the cache shouldn't be that large
 475                         TCP_INT_UpdateRecievedFromFuture(Connection);
 476
 477                         #if 1
 478                         // - Only send an ACK if we've had a burst
 479                         if( Connection->NextSequenceRcv > (Uint32)(TCP_DACK_THRESHOLD + Connection->LastACKSequence) )
 480                         {
 481                                 TCP_INT_SendACK(Connection, "DACK Burst");
 482                                 // - Extend TCP deferred ACK timer
 483                                 Time_RemoveTimer(Connection->DeferredACKTimer);
 484                         }
 485                         // - Schedule the deferred ACK timer (if already scheduled, this is a NOP)
 486                         Time_ScheduleTimer(Connection->DeferredACKTimer, TCP_DACK_TIMEOUT);
 487                         #else
 488                         TCP_INT_SendACK(Connection, "RX");
 489                         #endif
 490                 }
 491                 // Check if the packet is in window
 492                 else if( WrapBetween(Connection->NextSequenceRcv, sequence_num,
 493                                 Connection->NextSequenceRcv+TCP_WINDOW_SIZE, 0xFFFFFFFF) )
 494                 {
 495                         Uint8   *dataptr = (Uint8*)Header + (Header->DataOffset>>4)*4;
 496                         #if CACHE_FUTURE_PACKETS_IN_BYTES
 497                         Uint32  index;
 498
 499                         index = sequence_num % TCP_WINDOW_SIZE;
 500                         for( int i = 0; i < dataLen; i ++ )
 501                         {
 502                                 Connection->FuturePacketValidBytes[index/8] |= 1 << (index%8);
 503                                 Connection->FuturePacketData[index] = dataptr[i];
 504                                 // Do a wrap increment
 505                                 index ++;
 506                                 if(index == TCP_WINDOW_SIZE)    index = 0;
 507                         }
 508                         #else
 509                         tTCPStoredPacket        *pkt, *tmp, *prev = NULL;
 510
 511                         // Allocate and fill cached packet
 512                         pkt = malloc( sizeof(tTCPStoredPacket) + dataLen );
 513                         pkt->Next = NULL;
 514                         pkt->Sequence = ntohl(Header->SequenceNumber);
 515                         pkt->Length = dataLen;
 516                         memcpy(pkt->Data, dataptr, dataLen);
 517
 518                         Log_Log("TCP", "We missed a packet, caching",
 519                                 pkt->Sequence, Connection->NextSequenceRcv);
 520
 521                         // No? Well, let's cache it and look at it later
 522                         SHORTLOCK( &Connection->lFuturePackets );
 523                         for(tmp = Connection->FuturePackets;
 524                                 tmp;
 525                                 prev = tmp, tmp = tmp->Next)
 526                         {
 527                                 if(tmp->Sequence >= pkt->Sequence)      break;
 528                         }
 529
 530                         // Add if before first, or sequences don't match
 531                         if( !tmp || tmp->Sequence != pkt->Sequence )
 532                         {
 533                                 if(prev)
 534                                         prev->Next = pkt;
 535                                 else
 536                                         Connection->FuturePackets = pkt;
 537                                 pkt->Next = tmp;
 538                         }
 539                         // Replace if larger
 540                         else if(pkt->Length > tmp->Length)
 541                         {
 542                                 if(prev)
 543                                         prev->Next = pkt;
 544                                 pkt->Next = tmp->Next;
 545                                 free(tmp);
 546                         }
 547                         else
 548                         {
 549                                 free(pkt);      // TODO: Find some way to remove this
 550                         }
 551                         SHORTREL( &Connection->lFuturePackets );
 552                         #endif
 553                 }
 554                 // Badly out of sequence packet
 555                 else
 556                 {
 557                         Log_Log("TCP", "Fully out of sequence packet (0x%08x not between 0x%08x and 0x%08x), dropped",
 558                                 sequence_num, Connection->NextSequenceRcv, Connection->NextSequenceRcv+TCP_WINDOW_SIZE);
 559                         // Spec says we should send an empty ACK with the current state
 560                         TCP_INT_SendACK(Connection, "Bad Seq");
 561                 }
 562                 break;
 563
 564         // --- Remote close states
 565         case TCP_ST_CLOSE_WAIT:
 566
 567                 // Ignore everything, CLOSE_WAIT is terminated by the client
 568                 Log_Debug("TCP", "CLOSE WAIT - Ignoring packets");
 569
 570                 break;
 571
 572         // LAST-ACK - Waiting for the ACK of FIN (from CLOSE WAIT)
 573         case TCP_ST_LAST_ACK:
 574                 if( Header->Flags & TCP_FLAG_ACK )
 575                 {
 576                         Connection->State = TCP_ST_FINISHED;    // Connection completed
 577                         Log_Log("TCP", "LAST-ACK to CLOSED - Connection remote closed");
 578                         // TODO: Destrory the TCB
 579                 }
 580                 break;
 581
 582         // --- Local close States
 583         case TCP_ST_FIN_WAIT1:
 584                 if( Header->Flags & TCP_FLAG_FIN )
 585                 {
 586                         Connection->State = TCP_ST_CLOSING;
 587                         Log_Debug("TCP", "Conn %p closed, sent FIN and recieved FIN", Connection);
 588                         VFS_MarkError(&Connection->Node, 1);
 589
 590                         // ACK Packet
 591                         Header->DestPort = Header->SourcePort;
 592                         Header->SourcePort = htons(Connection->LocalPort);
 593                         Header->AcknowlegementNumber = Header->SequenceNumber;
 594                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 595                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 596                         Header->Flags = TCP_FLAG_ACK;
 597                         TCP_SendPacket( Connection, Header, 0, NULL );
 598                         break ;
 599                 }
 600
 601                 // TODO: Make sure that the packet is actually ACKing the FIN
 602                 if( Header->Flags & TCP_FLAG_ACK )
 603                 {
 604                         Connection->State = TCP_ST_FIN_WAIT2;
 605                         Log_Debug("TCP", "Conn %p closed, sent FIN ACKed", Connection);
 606                         VFS_MarkError(&Connection->Node, 1);
 607                         return ;
 608                 }
 609                 break;
 610
 611         case TCP_ST_FIN_WAIT2:
 612                 if( Header->Flags & TCP_FLAG_FIN )
 613                 {
 614                         Connection->State = TCP_ST_TIME_WAIT;
 615                         Log_Debug("TCP", "FIN sent and recieved, ACKing and going into TIME WAIT %p FINWAIT-2 -> TIME WAIT", Connection);
 616                         // Send ACK
 617                         Header->DestPort = Header->SourcePort;
 618                         Header->SourcePort = htons(Connection->LocalPort);
 619                         Header->AcknowlegementNumber = Header->SequenceNumber;
 620                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 621                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 622                         Header->Flags = TCP_FLAG_ACK;
 623                         TCP_SendPacket( Connection, Header, 0, NULL );
 624                 }
 625                 break;
 626
 627         case TCP_ST_CLOSING:
 628                 // TODO: Make sure that the packet is actually ACKing the FIN
 629                 if( Header->Flags & TCP_FLAG_ACK )
 630                 {
 631                         Connection->State = TCP_ST_TIME_WAIT;
 632                         Log_Debug("TCP", "Conn %p CLOSING -> TIME WAIT", Connection);
 633                         VFS_MarkError(&Connection->Node, 1);
 634                         return ;
 635                 }
 636                 break;
 637
 638         // --- Closed (or near closed) states) ---
 639         case TCP_ST_TIME_WAIT:
 640                 Log_Log("TCP", "Packets on Time-Wait, ignored");
 641                 break;
 642
 643         case TCP_ST_FINISHED:
 644                 Log_Log("TCP", "Packets when CLOSED, ignoring");
 645                 break;
 646
 647         //default:
 648         //      Log_Warning("TCP", "Unhandled TCP state %i", Connection->State);
 649         //      break;
 650         }
 651
 652 }
 653
 654 /**
 655  * \brief Appends a packet to the recieved list
 656  * \param Connection    Connection structure
 657  * \param Data  Packet contents
 658  * \param Length        Length of \a Data
 659  */
 660 int TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length)
 661 {
 662         Mutex_Acquire( &Connection->lRecievedPackets );
 663
 664         if(Connection->RecievedBuffer->Length + Length > Connection->RecievedBuffer->Space )
 665         {
 666                 VFS_MarkAvaliable(&Connection->Node, 1);
 667                 Log_Error("TCP", "Buffer filled, packet dropped (:%i) - %i + %i > %i",
 668                         Connection->LocalPort, Connection->RecievedBuffer->Length, Length,
 669                         Connection->RecievedBuffer->Space
 670                         );
 671                 Mutex_Release( &Connection->lRecievedPackets );
 672                 return 1;
 673         }
 674
 675         RingBuffer_Write( Connection->RecievedBuffer, Data, Length );
 676
 677         VFS_MarkAvaliable(&Connection->Node, 1);
 678
 679         Mutex_Release( &Connection->lRecievedPackets );
 680         return 0;
 681 }
 682
 683 /**
 684  * \brief Updates the connections recieved list from the future list
 685  * \param Connection    Connection structure
 686  *
 687  * Updates the recieved packets list with packets from the future (out
 688  * of order) packets list that are now able to be added in direct
 689  * sequence.
 690  */
 691 void TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection)
 692 {
 693         #if CACHE_FUTURE_PACKETS_IN_BYTES
 694         // Calculate length of contiguous bytes
 695          int    length = Connection->HighestSequenceRcvd - Connection->NextSequenceRcv;
 696         Uint32  index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 697         LOG("length=%i, index=%i", length, index);
 698         for( int i = 0; i < length; i ++ )
 699         {
 700                  int    bit = index % 8;
 701                 Uint8   bitfield_byte = Connection->FuturePacketValidBytes[index / 8];
 702                 if( (bitfield_byte & (1 << bit)) == 0 ) {
 703                         length = i;
 704                         break;
 705                 }
 706
 707                 if( bitfield_byte == 0xFF ) {
 708                          int    inc = 8 - bit;
 709                         i += inc - 1;
 710                         index += inc;
 711                 }
 712                 else {
 713                         index ++;
 714                 }
 715                 if(index > TCP_WINDOW_SIZE)
 716                         index -= TCP_WINDOW_SIZE;
 717         }
 718
 719         index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 720
 721         // Write data to to the ring buffer
 722         if( TCP_WINDOW_SIZE - index > length )
 723         {
 724                 // Simple case
 725                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, length );
 726         }
 727         else
 728         {
 729                  int    endLen = TCP_WINDOW_SIZE - index;
 730                 // 2-part case
 731                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, endLen );
 732                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData, endLen - length );
 733         }
 734
 735         // Mark (now saved) bytes as invalid
 736         // - Align index
 737         while(index % 8 && length > 0)
 738         {
 739                 Connection->FuturePacketData[index] = 0;
 740                 Connection->FuturePacketValidBytes[index/8] &= ~(1 << (index%8));
 741                 index ++;
 742                 if(index > TCP_WINDOW_SIZE)
 743                         index -= TCP_WINDOW_SIZE;
 744                 length --;
 745         }
 746         while( length > 7 )
 747         {
 748                 Connection->FuturePacketData[index] = 0;
 749                 Connection->FuturePacketValidBytes[index/8] = 0;
 750                 length -= 8;
 751                 index += 8;
 752                 if(index > TCP_WINDOW_SIZE)
 753                         index -= TCP_WINDOW_SIZE;
 754         }
 755         while(length)
 756         {
 757                 Connection->FuturePacketData[index] = 0;
 758                 Connection->FuturePacketData[index/8] &= ~(1 << (index%8));
 759                 index ++;
 760                 if(index > TCP_WINDOW_SIZE)
 761                         index -= TCP_WINDOW_SIZE;
 762                 length --;
 763         }
 764
 765         #else
 766         tTCPStoredPacket        *pkt;
 767         for(;;)
 768         {
 769                 SHORTLOCK( &Connection->lFuturePackets );
 770
 771                 // Clear out duplicates from cache
 772                 // - If a packet has just been recieved, and it is expected, then
 773                 //   (since NextSequenceRcv = rcvd->Sequence + rcvd->Length) all
 774                 //   packets in cache that are smaller than the next expected
 775                 //   are now defunct.
 776                 pkt = Connection->FuturePackets;
 777                 while(pkt && pkt->Sequence < Connection->NextSequenceRcv)
 778                 {
 779                         tTCPStoredPacket        *next = pkt->Next;
 780                         free(pkt);
 781                         pkt = next;
 782                 }
 783
 784                 // If there's no packets left in cache, stop looking
 785                 if(!pkt || pkt->Sequence > Connection->NextSequenceRcv) {
 786                         SHORTREL( &Connection->lFuturePackets );
 787                         return;
 788                 }
 789
 790                 // Delete packet from future list
 791                 Connection->FuturePackets = pkt->Next;
 792
 793                 // Release list
 794                 SHORTREL( &Connection->lFuturePackets );
 795
 796                 // Looks like we found one
 797                 TCP_INT_AppendRecieved(Connection, pkt->Data, pkt->Length);
 798                 if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 799                         Connection->HighestSequenceRcvd += pkt->Length;
 800                 Connection->NextSequenceRcv += pkt->Length;
 801                 free(pkt);
 802         }
 803         #endif
 804 }
 805
 806 void TCP_int_SendDelayedACK(void *ConnPtr)
 807 {
 808         TCP_INT_SendACK(ConnPtr, "DACK Timeout");
 809 }
 810
 811 void TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason)
 812 {
 813         tTCPHeader      hdr;
 814         // ACK Packet
 815         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
 816         hdr.DestPort = htons(Connection->RemotePort);
 817         hdr.SourcePort = htons(Connection->LocalPort);
 818         hdr.AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 819         hdr.SequenceNumber = htonl(Connection->NextSequenceSend);
 820         hdr.WindowSize = htons(TCP_WINDOW_SIZE);
 821         hdr.Flags = TCP_FLAG_ACK;       // TODO: Determine if SYN is wanted too
 822         hdr.Checksum = 0;       // TODO: Checksum
 823         hdr.UrgentPointer = 0;
 824         Log_Debug("TCP", "Sending ACK for 0x%08x (%s)", Connection->NextSequenceRcv, Reason);
 825         TCP_SendPacket( Connection, &hdr, 0, NULL );
 826         //Connection->NextSequenceSend ++;
 827         Connection->LastACKSequence = Connection->NextSequenceRcv;
 828 }
 829
 830 /**
 831  * \fn Uint16 TCP_GetUnusedPort()
 832  * \brief Gets an unused port and allocates it
 833  */
 834 Uint16 TCP_GetUnusedPort()
 835 {
 836         Uint16  ret;
 837
 838         // Get Next outbound port
 839         ret = giTCP_NextOutPort++;
 840         while( gaTCP_PortBitmap[ret/32] & (1UL << (ret%32)) )
 841         {
 842                 ret ++;
 843                 giTCP_NextOutPort++;
 844                 if(giTCP_NextOutPort == 0x10000) {
 845                         ret = giTCP_NextOutPort = TCP_MIN_DYNPORT;
 846                 }
 847         }
 848
 849         // Mark the new port as used
 850         gaTCP_PortBitmap[ret/32] |= 1 << (ret%32);
 851
 852         return ret;
 853 }
 854
 855 /**
 856  * \fn int TCP_AllocatePort(Uint16 Port)
 857  * \brief Marks a port as used
 858  */
 859 int TCP_AllocatePort(Uint16 Port)
 860 {
 861         // Check if the port has already been allocated
 862         if( gaTCP_PortBitmap[Port/32] & (1 << (Port%32)) )
 863                 return 0;
 864
 865         // Allocate
 866         gaTCP_PortBitmap[Port/32] |= 1 << (Port%32);
 867
 868         return 1;
 869 }
 870
 871 /**
 872  * \fn int TCP_DeallocatePort(Uint16 Port)
 873  * \brief Marks a port as unused
 874  */
 875 int TCP_DeallocatePort(Uint16 Port)
 876 {
 877         // Check if the port has already been allocated
 878         if( !(gaTCP_PortBitmap[Port/32] & (1 << (Port%32))) )
 879                 return 0;
 880
 881         // Allocate
 882         gaTCP_PortBitmap[Port/32] &= ~(1 << (Port%32));
 883
 884         return 1;
 885 }
 886
 887 tTCPConnection *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State)
 888 {
 889         tTCPConnection  *conn = calloc( sizeof(tTCPConnection) + TCP_WINDOW_SIZE + TCP_WINDOW_SIZE/8, 1 );
 890
 891         conn->State = State;
 892         conn->Interface = Interface;
 893         conn->LocalPort = -1;
 894         conn->RemotePort = -1;
 895
 896         conn->Node.ReferenceCount = 1;
 897         conn->Node.ImplPtr = conn;
 898         conn->Node.NumACLs = 1;
 899         conn->Node.ACLs = &gVFS_ACL_EveryoneRW;
 900         conn->Node.Type = &gTCP_ClientNodeType;
 901         conn->Node.BufferFull = 1;      // Cleared when connection opens
 902
 903         conn->RecievedBuffer = RingBuffer_Create( TCP_RECIEVE_BUFFER_SIZE );
 904         #if 0
 905         conn->SentBuffer = RingBuffer_Create( TCP_SEND_BUFFER_SIZE );
 906         Semaphore_Init(conn->SentBufferSpace, 0, TCP_SEND_BUFFER_SIZE, "TCP SentBuffer", conn->Name);
 907         #endif
 908
 909         #if CACHE_FUTURE_PACKETS_IN_BYTES
 910         // Future recieved data (ahead of the expected sequence number)
 911         conn->FuturePacketData = (Uint8*)conn + sizeof(tTCPConnection);
 912         conn->FuturePacketValidBytes = conn->FuturePacketData + TCP_WINDOW_SIZE;
 913         #endif
 914
 915         conn->DeferredACKTimer = Time_AllocateTimer( TCP_int_SendDelayedACK, conn);
 916         return conn;
 917 }
 918
 919 // --- Server
 920 tVFS_Node *TCP_Server_Init(tInterface *Interface)
 921 {
 922         tTCPListener    *srv;
 923
 924         srv = calloc( 1, sizeof(tTCPListener) );
 925
 926         if( srv == NULL ) {
 927                 Log_Warning("TCP", "malloc failed for listener (%i) bytes", sizeof(tTCPListener));
 928                 return NULL;
 929         }
 930
 931         srv->Interface = Interface;
 932         srv->Port = 0;
 933         srv->NextID = 0;
 934         srv->Connections = NULL;
 935         srv->ConnectionsTail = NULL;
 936         srv->NewConnections = NULL;
 937         srv->Next = NULL;
 938         srv->Node.Flags = VFS_FFLAG_DIRECTORY;
 939         srv->Node.Size = -1;
 940         srv->Node.ImplPtr = srv;
 941         srv->Node.NumACLs = 1;
 942         srv->Node.ACLs = &gVFS_ACL_EveryoneRW;
 943         srv->Node.Type = &gTCP_ServerNodeType;
 944
 945         SHORTLOCK(&glTCP_Listeners);
 946         srv->Next = gTCP_Listeners;
 947         gTCP_Listeners = srv;
 948         SHORTREL(&glTCP_Listeners);
 949
 950         return &srv->Node;
 951 }
 952
 953 /**
 954  * \brief Wait for a new connection and return the connection ID
 955  * \note Blocks until a new connection is made
 956  * \param Node  Server node
 957  * \param Pos   Position (ignored)
 958  */
 959 int TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Dest[FILENAME_MAX])
 960 {
 961         tTCPListener    *srv = Node->ImplPtr;
 962         tTCPConnection  *conn;
 963
 964         ENTER("pNode iPos", Node, Pos);
 965
 966         Log_Log("TCP", "Thread %i waiting for a connection", Threads_GetTID());
 967         Semaphore_Wait( &srv->WaitingConnections, 1 );
 968
 969         SHORTLOCK(&srv->lConnections);
 970         // Increment the new list (the current connection is still on the
 971         // normal list)
 972         conn = srv->NewConnections;
 973         srv->NewConnections = conn->Next;
 974
 975         if( srv->NewConnections == NULL )
 976                 VFS_MarkAvaliable( Node, 0 );
 977
 978         SHORTREL( &srv->lConnections );
 979
 980         LOG("conn = %p", conn);
 981         LOG("srv->Connections = %p", srv->Connections);
 982         LOG("srv->NewConnections = %p", srv->NewConnections);
 983         LOG("srv->ConnectionsTail = %p", srv->ConnectionsTail);
 984
 985         itoa(Dest, conn->Node.ImplInt, 16, 8, '0');
 986         Log_Log("TCP", "Thread %i got connection '%s'", Threads_GetTID(), Dest);
 987         LEAVE('i', 0);
 988         return 0;
 989 }
 990
 991 /**
 992  * \brief Gets a client connection node
 993  * \param Node  Server node
 994  * \param Name  Hexadecimal ID of the node
 995  */
 996 tVFS_Node *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags)
 997 {
 998         tTCPConnection  *conn;
 999         tTCPListener    *srv = Node->ImplPtr;
1000         char    tmp[9];
1001          int    id = atoi(Name);
1002
1003         ENTER("pNode sName", Node, Name);
1004
1005         // Check for a non-empty name
1006         if( Name[0] )
1007         {
1008                 // Sanity Check
1009                 itoa(tmp, id, 16, 8, '0');
1010                 if(strcmp(tmp, Name) != 0) {
1011                         LOG("'%s' != '%s' (%08x)", Name, tmp, id);
1012                         LEAVE('n');
1013                         return NULL;
1014                 }
1015
1016                 Log_Debug("TCP", "srv->Connections = %p", srv->Connections);
1017                 Log_Debug("TCP", "srv->NewConnections = %p", srv->NewConnections);
1018                 Log_Debug("TCP", "srv->ConnectionsTail = %p", srv->ConnectionsTail);
1019
1020                 // Search
1021                 SHORTLOCK( &srv->lConnections );
1022                 for(conn = srv->Connections;
1023                         conn;
1024                         conn = conn->Next)
1025                 {
1026                         LOG("conn->Node.ImplInt = %i", conn->Node.ImplInt);
1027                         if(conn->Node.ImplInt == id)    break;
1028                 }
1029                 SHORTREL( &srv->lConnections );
1030
1031                 // If not found, ret NULL
1032                 if(!conn) {
1033                         LOG("Connection %i not found", id);
1034                         LEAVE('n');
1035                         return NULL;
1036                 }
1037         }
1038         // Empty Name - Check for a new connection and if it's there, open it
1039         else
1040         {
1041                 SHORTLOCK( &srv->lConnections );
1042                 conn = srv->NewConnections;
1043                 if( conn != NULL )
1044                         srv->NewConnections = conn->Next;
1045                 VFS_MarkAvaliable( Node, srv->NewConnections != NULL );
1046                 SHORTREL( &srv->lConnections );
1047                 if( !conn ) {
1048                         LOG("No new connections");
1049                         LEAVE('n');
1050                         return NULL;
1051                 }
1052         }
1053
1054         // Return node
1055         LEAVE('p', &conn->Node);
1056         return &conn->Node;
1057 }
1058
1059 /**
1060  * \brief Handle IOCtl calls
1061  */
1062 int TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data)
1063 {
1064         tTCPListener    *srv = Node->ImplPtr;
1065
1066         switch(ID)
1067         {
1068         case 4: // Get/Set Port
1069                 if(!Data)       // Get Port
1070                         return srv->Port;
1071
1072                 if(srv->Port)   // Wait, you can't CHANGE the port
1073                         return -1;
1074
1075                 if(!CheckMem(Data, sizeof(Uint16)))     // Sanity check
1076                         return -1;
1077
1078                 // Permissions check
1079                 if(Threads_GetUID() != 0
1080                 && *(Uint16*)Data != 0
1081                 && *(Uint16*)Data < 1024)
1082                         return -1;
1083
1084                 // TODO: Check if a port is in use
1085
1086                 // Set Port
1087                 srv->Port = *(Uint16*)Data;
1088                 if(srv->Port == 0)      // Allocate a random port
1089                         srv->Port = TCP_GetUnusedPort();
1090                 else    // Else, mark this as used
1091                         TCP_AllocatePort(srv->Port);
1092
1093                 Log_Log("TCP", "Server %p listening on port %i", srv, srv->Port);
1094
1095                 return srv->Port;
1096         }
1097         return 0;
1098 }
1099
1100 void TCP_Server_Close(tVFS_Node *Node)
1101 {
1102         free(Node->ImplPtr);
1103 }
1104
1105 // --- Client
1106 /**
1107  * \brief Create a client node
1108  */
1109 tVFS_Node *TCP_Client_Init(tInterface *Interface)
1110 {
1111         tTCPConnection  *conn = TCP_int_CreateConnection(Interface, TCP_ST_CLOSED);
1112
1113         SHORTLOCK(&glTCP_OutbountCons);
1114         conn->Next = gTCP_OutbountCons;
1115         gTCP_OutbountCons = conn;
1116         SHORTREL(&glTCP_OutbountCons);
1117
1118         return &conn->Node;
1119 }
1120
1121 /**
1122  * \brief Wait for a packet and return it
1123  * \note If \a Length is smaller than the size of the packet, the rest
1124  *       of the packet's data will be discarded.
1125  */
1126 size_t TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags)
1127 {
1128         tTCPConnection  *conn = Node->ImplPtr;
1129         size_t  len;
1130
1131         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1132         LOG("conn = %p {State:%i}", conn, conn->State);
1133
1134         // If the connection has been closed (state > ST_OPEN) then clear
1135         // any stale data in the buffer (until it is empty (until it is empty))
1136         if( conn->State > TCP_ST_OPEN )
1137         {
1138                 Mutex_Acquire( &conn->lRecievedPackets );
1139                 len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1140                 Mutex_Release( &conn->lRecievedPackets );
1141
1142                 if( len == 0 ) {
1143                         VFS_MarkAvaliable(Node, 0);
1144                         errno = 0;
1145                         LEAVE('i', -1);
1146                         return -1;
1147                 }
1148
1149                 LEAVE('i', len);
1150                 return len;
1151         }
1152
1153         // Wait
1154         {
1155                 tTime   *timeout = NULL;
1156                 tTime   timeout_zero = 0;
1157                 if( Flags & VFS_IOFLAG_NOBLOCK )
1158                         timeout = &timeout_zero;
1159                 if( !VFS_SelectNode(Node, VFS_SELECT_READ|VFS_SELECT_ERROR, timeout, "TCP_Client_Read") ) {
1160                         errno = EWOULDBLOCK;
1161                         LEAVE('i', -1);
1162                         return -1;
1163                 }
1164         }
1165
1166         // Lock list and read as much as possible (up to `Length`)
1167         Mutex_Acquire( &conn->lRecievedPackets );
1168         len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1169
1170         if( len == 0 || conn->RecievedBuffer->Length == 0 ) {
1171                 LOG("Marking as none avaliable (len = %i)", len);
1172                 VFS_MarkAvaliable(Node, 0);
1173         }
1174
1175         // Release the lock (we don't need it any more)
1176         Mutex_Release( &conn->lRecievedPackets );
1177
1178         LEAVE('i', len);
1179         return len;
1180 }
1181
1182 /**
1183  * \brief Send a data packet on a connection
1184  */
1185 void TCP_INT_SendDataPacket(tTCPConnection *Connection, size_t Length, const void *Data)
1186 {
1187         char    buf[sizeof(tTCPHeader)+Length];
1188         tTCPHeader      *packet = (void*)buf;
1189
1190         // - Stop Delayed ACK timer (as this data packet ACKs)
1191         Time_RemoveTimer(Connection->DeferredACKTimer);
1192
1193         packet->SourcePort = htons(Connection->LocalPort);
1194         packet->DestPort = htons(Connection->RemotePort);
1195         packet->DataOffset = (sizeof(tTCPHeader)/4)*16;
1196         packet->WindowSize = htons(TCP_WINDOW_SIZE);
1197
1198         packet->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
1199         packet->SequenceNumber = htonl(Connection->NextSequenceSend);
1200         packet->Flags = TCP_FLAG_PSH|TCP_FLAG_ACK;      // Hey, ACK if you can!
1201         packet->UrgentPointer = 0;
1202
1203         memcpy(packet->Options, Data, Length);
1204
1205         Log_Debug("TCP", "Send sequence 0x%08x", Connection->NextSequenceSend);
1206 #if HEXDUMP_OUTGOING
1207         Debug_HexDump("TCP_INT_SendDataPacket: Data = ", Data, Length);
1208 #endif
1209
1210         TCP_SendPacket( Connection, packet, Length, Data );
1211
1212         Connection->NextSequenceSend += Length;
1213 }
1214
1215 /**
1216  * \brief Send some bytes on a connection
1217  */
1218 size_t TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags)
1219 {
1220         tTCPConnection  *conn = Node->ImplPtr;
1221         size_t  rem = Length;
1222
1223         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1224
1225 //      #if DEBUG
1226 //      Debug_HexDump("TCP_Client_Write: Buffer = ",
1227 //              Buffer, Length);
1228 //      #endif
1229
1230         // Don't allow a write to a closed connection
1231         if( conn->State > TCP_ST_OPEN ) {
1232                 VFS_MarkError(Node, 1);
1233                 errno = 0;
1234                 LEAVE('i', -1);
1235                 return -1;
1236         }
1237
1238         // Wait
1239         {
1240                 tTime   *timeout = NULL;
1241                 tTime   timeout_zero = 0;
1242                 if( Flags & VFS_IOFLAG_NOBLOCK )
1243                         timeout = &timeout_zero;
1244                 if( !VFS_SelectNode(Node, VFS_SELECT_WRITE|VFS_SELECT_ERROR, timeout, "TCP_Client_Write") ) {
1245                         errno = EWOULDBLOCK;
1246                         LEAVE('i', -1);
1247                         return -1;
1248                 }
1249         }
1250
1251         do
1252         {
1253                  int    len = (rem < TCP_MAX_PACKET_SIZE) ? rem : TCP_MAX_PACKET_SIZE;
1254
1255                 #if 0
1256                 // Wait for space in the buffer
1257                 Semaphore_Signal( &Connection->SentBufferSpace, len );
1258
1259                 // Save data to buffer (and update the length read by the ammount written)
1260                 len = RingBuffer_Write( &Connection->SentBuffer, Buffer, len);
1261                 #endif
1262
1263                 // Send packet
1264                 TCP_INT_SendDataPacket(conn, len, Buffer);
1265
1266                 Buffer += len;
1267                 rem -= len;
1268         } while( rem > 0 );
1269
1270         LEAVE('i', Length);
1271         return Length;
1272 }
1273
1274 /**
1275  * \brief Open a connection to another host using TCP
1276  * \param Conn  Connection structure
1277  */
1278 void TCP_StartConnection(tTCPConnection *Conn)
1279 {
1280         tTCPHeader      hdr = {0};
1281
1282         Conn->State = TCP_ST_SYN_SENT;
1283
1284         hdr.SourcePort = htons(Conn->LocalPort);
1285         hdr.DestPort = htons(Conn->RemotePort);
1286         Conn->NextSequenceSend = rand();
1287         hdr.SequenceNumber = htonl(Conn->NextSequenceSend);
1288         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
1289         hdr.Flags = TCP_FLAG_SYN;
1290         hdr.WindowSize = htons(TCP_WINDOW_SIZE);        // Max
1291         hdr.Checksum = 0;       // TODO
1292
1293         TCP_SendPacket( Conn, &hdr, 0, NULL );
1294
1295         Conn->NextSequenceSend ++;
1296         Conn->State = TCP_ST_SYN_SENT;
1297
1298         return ;
1299 }
1300
1301 /**
1302  * \brief Control a client socket
1303  */
1304 int TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data)
1305 {
1306         tTCPConnection  *conn = Node->ImplPtr;
1307
1308         ENTER("pNode iID pData", Node, ID, Data);
1309
1310         switch(ID)
1311         {
1312         case 4: // Get/Set local port
1313                 if(!Data)
1314                         LEAVE_RET('i', conn->LocalPort);
1315                 if(conn->State != TCP_ST_CLOSED)
1316                         LEAVE_RET('i', -1);
1317                 if(!CheckMem(Data, sizeof(Uint16)))
1318                         LEAVE_RET('i', -1);
1319
1320                 if(Threads_GetUID() != 0 && *(Uint16*)Data < 1024)
1321                         LEAVE_RET('i', -1);
1322
1323                 conn->LocalPort = *(Uint16*)Data;
1324                 LEAVE_RET('i', conn->LocalPort);
1325
1326         case 5: // Get/Set remote port
1327                 if(!Data)       LEAVE_RET('i', conn->RemotePort);
1328                 if(conn->State != TCP_ST_CLOSED)        LEAVE_RET('i', -1);
1329                 if(!CheckMem(Data, sizeof(Uint16)))     LEAVE_RET('i', -1);
1330                 conn->RemotePort = *(Uint16*)Data;
1331                 LEAVE_RET('i', conn->RemotePort);
1332
1333         case 6: // Set Remote IP
1334                 if( conn->State != TCP_ST_CLOSED )
1335                         LEAVE_RET('i', -1);
1336                 if( conn->Interface->Type == 4 )
1337                 {
1338                         if(!CheckMem(Data, sizeof(tIPv4)))      LEAVE_RET('i', -1);
1339                         conn->RemoteIP.v4 = *(tIPv4*)Data;
1340                 }
1341                 else if( conn->Interface->Type == 6 )
1342                 {
1343                         if(!CheckMem(Data, sizeof(tIPv6)))      LEAVE_RET('i', -1);
1344                         conn->RemoteIP.v6 = *(tIPv6*)Data;
1345                 }
1346                 LEAVE_RET('i', 0);
1347
1348         case 7: // Connect
1349                 if(conn->LocalPort == 0xFFFF)
1350                         conn->LocalPort = TCP_GetUnusedPort();
1351                 if(conn->RemotePort == -1)
1352                         LEAVE_RET('i', 0);
1353
1354                 {
1355                         tTime   timeout = conn->Interface->TimeoutDelay;
1356
1357                         TCP_StartConnection(conn);
1358                         VFS_SelectNode(&conn->Node, VFS_SELECT_WRITE, &timeout, "TCP Connection");
1359                         if( conn->State == TCP_ST_SYN_SENT )
1360                                 LEAVE_RET('i', 0);
1361                 }
1362
1363                 LEAVE_RET('i', 1);
1364
1365         // Get recieve buffer length
1366         case 8:
1367                 LEAVE_RET('i', conn->RecievedBuffer->Length);
1368         }
1369
1370         return 0;
1371 }
1372
1373 void TCP_Client_Close(tVFS_Node *Node)
1374 {
1375         tTCPConnection  *conn = Node->ImplPtr;
1376         tTCPHeader      packet;
1377
1378         ENTER("pNode", Node);
1379
1380         ASSERT(Node->ReferenceCount != 0);
1381
1382         if( Node->ReferenceCount > 1 ) {
1383                 Node->ReferenceCount --;
1384                 LOG("Dereference only");
1385                 LEAVE('-');
1386                 return ;
1387         }
1388
1389         if( conn->State == TCP_ST_CLOSE_WAIT || conn->State == TCP_ST_OPEN )
1390         {
1391                 packet.SourcePort = htons(conn->LocalPort);
1392                 packet.DestPort = htons(conn->RemotePort);
1393                 packet.DataOffset = (sizeof(tTCPHeader)/4)*16;
1394                 packet.WindowSize = TCP_WINDOW_SIZE;
1395
1396                 packet.AcknowlegementNumber = 0;
1397                 packet.SequenceNumber = htonl(conn->NextSequenceSend);
1398                 packet.Flags = TCP_FLAG_FIN;
1399
1400                 TCP_SendPacket( conn, &packet, 0, NULL );
1401         }
1402
1403         switch( conn->State )
1404         {
1405         case TCP_ST_CLOSED:
1406                 Log_Warning("TCP", "Closing connection that was never opened");
1407                 break;
1408         case TCP_ST_CLOSE_WAIT:
1409                 conn->State = TCP_ST_LAST_ACK;
1410                 break;
1411         case TCP_ST_OPEN:
1412                 conn->State = TCP_ST_FIN_WAIT1;
1413                 while( conn->State == TCP_ST_FIN_WAIT1 )        Threads_Yield();
1414                 break;
1415         default:
1416                 Log_Warning("TCP", "Unhandled connection state %i in TCP_Client_Close",
1417                         conn->State);
1418                 break;
1419         }
1420
1421         Time_RemoveTimer(conn->DeferredACKTimer);
1422         Time_FreeTimer(conn->DeferredACKTimer);
1423         free(conn);
1424
1425         LEAVE('-');
1426 }
1427
1428 /**
1429  * \brief Checks if a value is between two others (after taking into account wrapping)
1430  */
1431 int WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue)
1432 {
1433         if( MaxValue < 0xFFFFFFFF )
1434         {
1435                 Lower %= MaxValue + 1;
1436                 Value %= MaxValue + 1;
1437                 Higher %= MaxValue + 1;
1438         }
1439
1440         // Simple Case, no wrap ?
1441         //       Lower Value Higher
1442         // | ... + ... + ... + ... |
1443
1444         if( Lower < Higher ) {
1445                 return Lower < Value && Value < Higher;
1446         }
1447         // Higher has wrapped below lower
1448
1449         // Value > Lower ?
1450         //       Higher Lower Value
1451         // | ... +  ... + ... + ... |
1452         if( Value > Lower ) {
1453                 return 1;
1454         }
1455
1456         // Value < Higher ?
1457         //       Value Higher Lower
1458         // | ... + ... +  ... + ... |
1459         if( Value < Higher ) {
1460                 return 1;
1461         }
1462
1463         return 0;
1464 }