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 USE_SELECT      1
  12 #define HEXDUMP_INCOMING        0
  13 #define HEXDUMP_OUTGOING        0
  14
  15 #define TCP_MIN_DYNPORT 0xC000
  16 #define TCP_MAX_HALFOPEN        1024    // Should be enough
  17
  18 #define TCP_MAX_PACKET_SIZE     1024
  19 #define TCP_WINDOW_SIZE 0x2000
  20 #define TCP_RECIEVE_BUFFER_SIZE 0x8000
  21 #define TCP_DACK_THRESHOLD      4096
  22 #define TCP_DACK_TIMEOUT        500
  23
  24 #define TCP_DEBUG       0       // Set to non-0 to enable TCP packet logging
  25
  26 // === PROTOTYPES ===
  27 void    TCP_Initialise(void);
  28 void    TCP_StartConnection(tTCPConnection *Conn);
  29 void    TCP_SendPacket(tTCPConnection *Conn, tTCPHeader *Header, size_t DataLen, const void *Data);
  30 void    TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer);
  31 void    TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length);
  32 int     TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length);
  33 void    TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection);
  34 void    TCP_int_SendDelayedACK(void *ConnPtr);
  35 void    TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason);
  36 Uint16  TCP_GetUnusedPort();
  37  int    TCP_AllocatePort(Uint16 Port);
  38  int    TCP_DeallocatePort(Uint16 Port);
  39 tTCPConnection  *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State);
  40 // --- Server
  41 tVFS_Node       *TCP_Server_Init(tInterface *Interface);
  42  int    TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Name[FILENAME_MAX]);
  43 tVFS_Node       *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags);
  44  int    TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data);
  45 void    TCP_Server_Close(tVFS_Node *Node);
  46 // --- Client
  47 tVFS_Node       *TCP_Client_Init(tInterface *Interface);
  48 size_t  TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags);
  49 size_t  TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags);
  50  int    TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data);
  51 void    TCP_Client_Close(tVFS_Node *Node);
  52 // --- Helpers
  53  int    WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue);
  54
  55 // === TEMPLATES ===
  56 tSocketFile     gTCP_ServerFile = {NULL, "tcps", TCP_Server_Init};
  57 tSocketFile     gTCP_ClientFile = {NULL, "tcpc", TCP_Client_Init};
  58 tVFS_NodeType   gTCP_ServerNodeType = {
  59         .TypeName = "TCP Server",
  60         .ReadDir = TCP_Server_ReadDir,
  61         .FindDir = TCP_Server_FindDir,
  62         .IOCtl   = TCP_Server_IOCtl,
  63         .Close   = TCP_Server_Close
  64         };
  65 tVFS_NodeType   gTCP_ClientNodeType = {
  66         .TypeName = "TCP Client/Connection",
  67         .Read  = TCP_Client_Read,
  68         .Write = TCP_Client_Write,
  69         .IOCtl = TCP_Client_IOCtl,
  70         .Close = TCP_Client_Close
  71         };
  72
  73 // === GLOBALS ===
  74  int    giTCP_NumHalfopen = 0;
  75 tShortSpinlock  glTCP_Listeners;
  76 tTCPListener    *gTCP_Listeners;
  77 tShortSpinlock  glTCP_OutbountCons;
  78 tTCPConnection  *gTCP_OutbountCons;
  79 Uint32  gaTCP_PortBitmap[0x800];
  80  int    giTCP_NextOutPort = TCP_MIN_DYNPORT;
  81
  82 // === CODE ===
  83 /**
  84  * \brief Initialise the TCP Layer
  85  *
  86  * Registers the client and server files and the GetPacket callback
  87  */
  88 void TCP_Initialise(void)
  89 {
  90         giTCP_NextOutPort += rand()%32;
  91         IPStack_AddFile(&gTCP_ServerFile);
  92         IPStack_AddFile(&gTCP_ClientFile);
  93         IPv4_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  94         IPv6_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  95 }
  96
  97 /**
  98  * \brief Sends a packet from the specified connection, calculating the checksums
  99  * \param Conn  Connection
 100  * \param Length        Length of data
 101  * \param Data  Packet data (cast as a TCP Header)
 102  */
 103 void TCP_SendPacket( tTCPConnection *Conn, tTCPHeader *Header, size_t Length, const void *Data )
 104 {
 105         tIPStackBuffer  *buffer;
 106         Uint16  checksum[3];
 107          int    packlen = sizeof(*Header) + Length;
 108
 109         buffer = IPStack_Buffer_CreateBuffer(2 + IPV4_BUFFERS);
 110         if( Data && Length )
 111                 IPStack_Buffer_AppendSubBuffer(buffer, Length, 0, Data, NULL, NULL);
 112         IPStack_Buffer_AppendSubBuffer(buffer, sizeof(*Header), 0, Header, NULL, NULL);
 113
 114         LOG("Sending %i+%i to %s:%i", sizeof(*Header), Length,
 115                 IPStack_PrintAddress(Conn->Interface->Type, &Conn->RemoteIP),
 116                 Conn->RemotePort
 117                 );
 118
 119         Header->Checksum = 0;
 120         checksum[1] = htons( ~IPv4_Checksum(Header, sizeof(tTCPHeader)) );
 121         checksum[2] = htons( ~IPv4_Checksum(Data, Length) );
 122
 123         // TODO: Fragment packet
 124
 125         switch( Conn->Interface->Type )
 126         {
 127         case 4:
 128                 // Get IPv4 pseudo-header checksum
 129                 {
 130                         Uint32  buf[3];
 131                         buf[0] = ((tIPv4*)Conn->Interface->Address)->L;
 132                         buf[1] = Conn->RemoteIP.v4.L;
 133                         buf[2] = (htons(packlen)<<16) | (6<<8) | 0;
 134                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 135                 }
 136                 // - Combine checksums
 137                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );
 138                 IPv4_SendPacket(Conn->Interface, Conn->RemoteIP.v4, IP4PROT_TCP, 0, buffer);
 139                 break;
 140
 141         case 6:
 142                 // Append IPv6 Pseudo Header
 143                 {
 144                         Uint32  buf[4+4+1+1];
 145                         memcpy(buf, Conn->Interface->Address, 16);
 146                         memcpy(&buf[4], &Conn->RemoteIP, 16);
 147                         buf[8] = htonl(packlen);
 148                         buf[9] = htonl(6);
 149                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 150                 }
 151                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );  // Combine the two
 152                 IPv6_SendPacket(Conn->Interface, Conn->RemoteIP.v6, IP4PROT_TCP, Length, Data);
 153                 break;
 154         }
 155 }
 156
 157 /**
 158  * \brief Handles a packet from the IP Layer
 159  * \param Interface     Interface the packet arrived from
 160  * \param Address       Pointer to the addres structure
 161  * \param Length        Size of packet in bytes
 162  * \param Buffer        Packet data
 163  */
 164 void TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer)
 165 {
 166         tTCPHeader      *hdr = Buffer;
 167         tTCPListener    *srv;
 168         tTCPConnection  *conn;
 169
 170         #if TCP_DEBUG
 171         Log_Log("TCP", "TCP_GetPacket: <Local>:%i from [%s]:%i, Flags = %s%s%s%s%s%s%s%s",
 172                 ntohs(hdr->DestPort),
 173                 IPStack_PrintAddress(Interface->Type, Address),
 174                 ntohs(hdr->SourcePort),
 175                 (hdr->Flags & TCP_FLAG_CWR) ? "CWR " : "",
 176                 (hdr->Flags & TCP_FLAG_ECE) ? "ECE " : "",
 177                 (hdr->Flags & TCP_FLAG_URG) ? "URG " : "",
 178                 (hdr->Flags & TCP_FLAG_ACK) ? "ACK " : "",
 179                 (hdr->Flags & TCP_FLAG_PSH) ? "PSH " : "",
 180                 (hdr->Flags & TCP_FLAG_RST) ? "RST " : "",
 181                 (hdr->Flags & TCP_FLAG_SYN) ? "SYN " : "",
 182                 (hdr->Flags & TCP_FLAG_FIN) ? "FIN " : ""
 183                 );
 184         #endif
 185
 186         if( Length > (hdr->DataOffset >> 4)*4 )
 187         {
 188                 LOG("SequenceNumber = 0x%x", ntohl(hdr->SequenceNumber));
 189 #if HEXDUMP_INCOMING
 190                 Debug_HexDump(
 191                         "TCP_GetPacket: Packet Data = ",
 192                         (Uint8*)hdr + (hdr->DataOffset >> 4)*4,
 193                         Length - (hdr->DataOffset >> 4)*4
 194                         );
 195 #endif
 196         }
 197
 198         // Check Servers
 199         for( srv = gTCP_Listeners; srv; srv = srv->Next )
 200         {
 201                 // Check if the server is active
 202                 if(srv->Port == 0)      continue;
 203                 // Check the interface
 204                 if(srv->Interface && srv->Interface != Interface)       continue;
 205                 // Check the destination port
 206                 if(srv->Port != htons(hdr->DestPort))   continue;
 207
 208                 Log_Log("TCP", "TCP_GetPacket: Matches server %p", srv);
 209                 // Is this in an established connection?
 210                 for( conn = srv->Connections; conn; conn = conn->Next )
 211                 {
 212                         // Check that it is coming in on the same interface
 213                         if(conn->Interface != Interface)        continue;
 214
 215                         // Check Source Port
 216                         Log_Log("TCP", "TCP_GetPacket: conn->RemotePort(%i) == hdr->SourcePort(%i)",
 217                                 conn->RemotePort, ntohs(hdr->SourcePort));
 218                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 219
 220                         // Check Source IP
 221                         Log_Debug("TCP", "TCP_GetPacket: conn->RemoteIP(%s)",
 222                                 IPStack_PrintAddress(conn->Interface->Type, &conn->RemoteIP));
 223                         Log_Debug("TCP", "                == Address(%s)",
 224                                 IPStack_PrintAddress(conn->Interface->Type, Address));
 225                         if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 226                                 continue ;
 227
 228                         Log_Log("TCP", "TCP_GetPacket: Matches connection %p", conn);
 229                         // We have a response!
 230                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 231
 232                         return;
 233                 }
 234
 235                 Log_Log("TCP", "TCP_GetPacket: Opening Connection");
 236                 // Open a new connection (well, check that it's a SYN)
 237                 if(hdr->Flags != TCP_FLAG_SYN) {
 238                         Log_Log("TCP", "TCP_GetPacket: Packet is not a SYN");
 239                         return ;
 240                 }
 241
 242                 // TODO: Check for halfopen max
 243
 244                 conn = TCP_int_CreateConnection(Interface, TCP_ST_SYN_RCVD);
 245                 conn->LocalPort = srv->Port;
 246                 conn->RemotePort = ntohs(hdr->SourcePort);
 247
 248                 switch(Interface->Type)
 249                 {
 250                 case 4: conn->RemoteIP.v4 = *(tIPv4*)Address;   break;
 251                 case 6: conn->RemoteIP.v6 = *(tIPv6*)Address;   break;
 252                 }
 253
 254                 conn->NextSequenceRcv = ntohl( hdr->SequenceNumber ) + 1;
 255                 conn->HighestSequenceRcvd = conn->NextSequenceRcv;
 256                 conn->NextSequenceSend = rand();
 257
 258                 conn->Node.ImplInt = srv->NextID ++;
 259
 260                 // Hmm... Theoretically, this lock will never have to wait,
 261                 // as the interface is locked to the watching thread, and this
 262                 // runs in the watching thread. But, it's a good idea to have
 263                 // it, just in case
 264                 // Oh, wait, there is a case where a wildcard can be used
 265                 // (srv->Interface == NULL) so having the lock is a good idea
 266                 SHORTLOCK(&srv->lConnections);
 267                 if( !srv->Connections )
 268                         srv->Connections = conn;
 269                 else
 270                         srv->ConnectionsTail->Next = conn;
 271                 srv->ConnectionsTail = conn;
 272                 if(!srv->NewConnections)
 273                         srv->NewConnections = conn;
 274                 VFS_MarkAvaliable( &srv->Node, 1 );
 275                 SHORTREL(&srv->lConnections);
 276                 Semaphore_Signal(&srv->WaitingConnections, 1);
 277
 278                 // Send the SYN ACK
 279                 hdr->Flags |= TCP_FLAG_ACK;
 280                 hdr->AcknowlegementNumber = htonl(conn->NextSequenceRcv);
 281                 hdr->SequenceNumber = htonl(conn->NextSequenceSend);
 282                 hdr->DestPort = hdr->SourcePort;
 283                 hdr->SourcePort = htons(srv->Port);
 284                 hdr->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 285                 TCP_SendPacket( conn, hdr, 0, NULL );
 286                 conn->NextSequenceSend ++;
 287                 return ;
 288         }
 289
 290         // Check Open Connections
 291         {
 292                 for( conn = gTCP_OutbountCons; conn; conn = conn->Next )
 293                 {
 294                         // Check that it is coming in on the same interface
 295                         if(conn->Interface != Interface)        continue;
 296
 297                         // Check Source Port
 298                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 299
 300                         // Check Source IP
 301                         if(conn->Interface->Type == 6 && !IP6_EQU(conn->RemoteIP.v6, *(tIPv6*)Address))
 302                                 continue;
 303                         if(conn->Interface->Type == 4 && !IP4_EQU(conn->RemoteIP.v4, *(tIPv4*)Address))
 304                                 continue;
 305
 306                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 307                         return ;
 308                 }
 309         }
 310
 311         Log_Log("TCP", "TCP_GetPacket: No Match");
 312 }
 313
 314 /**
 315  * \brief Handles a packet sent to a specific connection
 316  * \param Connection    TCP Connection pointer
 317  * \param Header        TCP Packet pointer
 318  * \param Length        Length of the packet
 319  */
 320 void TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length)
 321 {
 322          int    dataLen;
 323         Uint32  sequence_num;
 324
 325         // Silently drop once finished
 326         // TODO: Check if this needs to be here
 327         if( Connection->State == TCP_ST_FINISHED ) {
 328                 Log_Log("TCP", "Packet ignored - connection finnished");
 329                 return ;
 330         }
 331
 332         // Syncronise sequence values
 333         if(Header->Flags & TCP_FLAG_SYN) {
 334                 // TODO: What if the packet also has data?
 335                 if( Connection->LastACKSequence != Connection->NextSequenceRcv )
 336                         TCP_INT_SendACK(Connection, "SYN");
 337                 Connection->NextSequenceRcv = ntohl(Header->SequenceNumber);
 338                 // TODO: Process HighestSequenceRcvd
 339                 // HACK!
 340                 if( Connection->HighestSequenceRcvd == 0 )
 341                         Connection->HighestSequenceRcvd = Connection->NextSequenceRcv;
 342                 Connection->LastACKSequence = Connection->NextSequenceRcv;
 343         }
 344
 345         // Ackowledge a sent packet
 346         if(Header->Flags & TCP_FLAG_ACK) {
 347                 // TODO: Process an ACKed Packet
 348                 LOG("Conn %p, Sent packet 0x%x ACKed", Connection, Header->AcknowlegementNumber);
 349         }
 350
 351         // Get length of data
 352         dataLen = Length - (Header->DataOffset>>4)*4;
 353         LOG("dataLen = %i", dataLen);
 354         #if TCP_DEBUG
 355         Log_Debug("TCP", "State %i, dataLen = %x", Connection->State, dataLen);
 356         #endif
 357
 358         //
 359         // State Machine
 360         //
 361         switch( Connection->State )
 362         {
 363         // Pre-init connection?
 364         case TCP_ST_CLOSED:
 365                 Log_Log("TCP", "Packets to a closed connection?!");
 366                 break;
 367
 368         // --- Init States ---
 369         // SYN sent, expecting SYN-ACK Connection Opening
 370         case TCP_ST_SYN_SENT:
 371                 if( Header->Flags & TCP_FLAG_SYN )
 372                 {
 373                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 374                                 Connection->HighestSequenceRcvd ++;
 375                         Connection->NextSequenceRcv ++;
 376
 377                         if( Header->Flags & TCP_FLAG_ACK )
 378                         {
 379                                 Log_Log("TCP", "ACKing SYN-ACK");
 380                                 Connection->State = TCP_ST_OPEN;
 381                                 VFS_MarkFull(&Connection->Node, 0);
 382                         }
 383                         else
 384                         {
 385                                 Log_Log("TCP", "ACKing SYN");
 386                                 Connection->State = TCP_ST_SYN_RCVD;
 387                         }
 388                         Header->DestPort = Header->SourcePort;
 389                         Header->SourcePort = htons(Connection->LocalPort);
 390                         Header->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 391                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 392                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 393                         Header->Flags = TCP_FLAG_ACK;
 394                         Header->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 395                         TCP_SendPacket( Connection, Header, 0, NULL );
 396                 }
 397                 break;
 398
 399         // SYN-ACK sent, expecting ACK
 400         case TCP_ST_SYN_RCVD:
 401                 if( Header->Flags & TCP_FLAG_ACK )
 402                 {
 403                         // TODO: Handle max half-open limit
 404                         Log_Log("TCP", "Connection fully opened");
 405                         Connection->State = TCP_ST_OPEN;
 406                         VFS_MarkFull(&Connection->Node, 0);
 407                 }
 408                 break;
 409
 410         // --- Established State ---
 411         case TCP_ST_OPEN:
 412                 // - Handle State changes
 413                 //
 414                 if( Header->Flags & TCP_FLAG_FIN ) {
 415                         Log_Log("TCP", "Conn %p closed, recieved FIN", Connection);
 416                         VFS_MarkError(&Connection->Node, 1);
 417                         Connection->State = TCP_ST_CLOSE_WAIT;
 418 //                      Header->Flags &= ~TCP_FLAG_FIN;
 419                         // CLOSE WAIT requires the client to close (or does it?)
 420                         #if 0
 421
 422                         #endif
 423                 }
 424
 425                 // Check for an empty packet
 426                 if(dataLen == 0) {
 427                         if( Header->Flags == TCP_FLAG_ACK )
 428                         {
 429                                 Log_Log("TCP", "ACK only packet");
 430                                 return ;
 431                         }
 432                         // TODO: Is this right? (empty packet counts as one byte)
 433                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 434                                 Connection->HighestSequenceRcvd ++;
 435                         Connection->NextSequenceRcv ++;
 436                         Log_Log("TCP", "Empty Packet, inc and ACK the current sequence number");
 437                         TCP_INT_SendACK(Connection, "Empty");
 438                         return ;
 439                 }
 440
 441                 // NOTES:
 442                 // Flags
 443                 //    PSH - Has Data?
 444                 // /NOTES
 445
 446                 sequence_num = ntohl(Header->SequenceNumber);
 447
 448                 LOG("0x%08x <= 0x%08x < 0x%08x",
 449                         Connection->NextSequenceRcv,
 450                         ntohl(Header->SequenceNumber),
 451                         Connection->NextSequenceRcv + TCP_WINDOW_SIZE
 452                         );
 453
 454                 // Is this packet the next expected packet?
 455                 if( sequence_num == Connection->NextSequenceRcv )
 456                 {
 457                          int    rv;
 458                         // Ooh, Goodie! Add it to the recieved list
 459                         rv = TCP_INT_AppendRecieved(Connection,
 460                                 (Uint8*)Header + (Header->DataOffset>>4)*4,
 461                                 dataLen
 462                                 );
 463                         if(rv != 0) {
 464                                 Log_Notice("TCP", "TCP_INT_AppendRecieved rv %i", rv);
 465                                 break;
 466                         }
 467                         LOG("0x%08x += %i", Connection->NextSequenceRcv, dataLen);
 468                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 469                                 Connection->HighestSequenceRcvd += dataLen;
 470                         Connection->NextSequenceRcv += dataLen;
 471
 472                         // TODO: This should be moved out of the watcher thread,
 473                         // so that a single lost packet on one connection doesn't cause
 474                         // all connections on the interface to lag.
 475                         // - Meh, no real issue, as the cache shouldn't be that large
 476                         TCP_INT_UpdateRecievedFromFuture(Connection);
 477
 478                         #if 1
 479                         // - Only send an ACK if we've had a burst
 480                         if( Connection->NextSequenceRcv > (Uint32)(TCP_DACK_THRESHOLD + Connection->LastACKSequence) )
 481                         {
 482                                 TCP_INT_SendACK(Connection, "DACK Burst");
 483                                 // - Extend TCP deferred ACK timer
 484                                 Time_RemoveTimer(Connection->DeferredACKTimer);
 485                         }
 486                         // - Schedule the deferred ACK timer (if already scheduled, this is a NOP)
 487                         Time_ScheduleTimer(Connection->DeferredACKTimer, TCP_DACK_TIMEOUT);
 488                         #else
 489                         TCP_INT_SendACK(Connection, "RX");
 490                         #endif
 491                 }
 492                 // Check if the packet is in window
 493                 else if( WrapBetween(Connection->NextSequenceRcv, sequence_num,
 494                                 Connection->NextSequenceRcv+TCP_WINDOW_SIZE, 0xFFFFFFFF) )
 495                 {
 496                         Uint8   *dataptr = (Uint8*)Header + (Header->DataOffset>>4)*4;
 497                         #if CACHE_FUTURE_PACKETS_IN_BYTES
 498                         Uint32  index;
 499
 500                         index = sequence_num % TCP_WINDOW_SIZE;
 501                         for( int i = 0; i < dataLen; i ++ )
 502                         {
 503                                 Connection->FuturePacketValidBytes[index/8] |= 1 << (index%8);
 504                                 Connection->FuturePacketData[index] = dataptr[i];
 505                                 // Do a wrap increment
 506                                 index ++;
 507                                 if(index == TCP_WINDOW_SIZE)    index = 0;
 508                         }
 509                         #else
 510                         tTCPStoredPacket        *pkt, *tmp, *prev = NULL;
 511
 512                         // Allocate and fill cached packet
 513                         pkt = malloc( sizeof(tTCPStoredPacket) + dataLen );
 514                         pkt->Next = NULL;
 515                         pkt->Sequence = ntohl(Header->SequenceNumber);
 516                         pkt->Length = dataLen;
 517                         memcpy(pkt->Data, dataptr, dataLen);
 518
 519                         Log_Log("TCP", "We missed a packet, caching",
 520                                 pkt->Sequence, Connection->NextSequenceRcv);
 521
 522                         // No? Well, let's cache it and look at it later
 523                         SHORTLOCK( &Connection->lFuturePackets );
 524                         for(tmp = Connection->FuturePackets;
 525                                 tmp;
 526                                 prev = tmp, tmp = tmp->Next)
 527                         {
 528                                 if(tmp->Sequence >= pkt->Sequence)      break;
 529                         }
 530
 531                         // Add if before first, or sequences don't match
 532                         if( !tmp || tmp->Sequence != pkt->Sequence )
 533                         {
 534                                 if(prev)
 535                                         prev->Next = pkt;
 536                                 else
 537                                         Connection->FuturePackets = pkt;
 538                                 pkt->Next = tmp;
 539                         }
 540                         // Replace if larger
 541                         else if(pkt->Length > tmp->Length)
 542                         {
 543                                 if(prev)
 544                                         prev->Next = pkt;
 545                                 pkt->Next = tmp->Next;
 546                                 free(tmp);
 547                         }
 548                         else
 549                         {
 550                                 free(pkt);      // TODO: Find some way to remove this
 551                         }
 552                         SHORTREL( &Connection->lFuturePackets );
 553                         #endif
 554                 }
 555                 // Badly out of sequence packet
 556                 else
 557                 {
 558                         Log_Log("TCP", "Fully out of sequence packet (0x%08x not between 0x%08x and 0x%08x), dropped",
 559                                 sequence_num, Connection->NextSequenceRcv, Connection->NextSequenceRcv+TCP_WINDOW_SIZE);
 560                         // Spec says we should send an empty ACK with the current state
 561                         TCP_INT_SendACK(Connection, "Bad Seq");
 562                 }
 563                 break;
 564
 565         // --- Remote close states
 566         case TCP_ST_CLOSE_WAIT:
 567
 568                 // Ignore everything, CLOSE_WAIT is terminated by the client
 569                 Log_Debug("TCP", "CLOSE WAIT - Ignoring packets");
 570
 571                 break;
 572
 573         // LAST-ACK - Waiting for the ACK of FIN (from CLOSE WAIT)
 574         case TCP_ST_LAST_ACK:
 575                 if( Header->Flags & TCP_FLAG_ACK )
 576                 {
 577                         Connection->State = TCP_ST_FINISHED;    // Connection completed
 578                         Log_Log("TCP", "LAST-ACK to CLOSED - Connection remote closed");
 579                         // TODO: Destrory the TCB
 580                 }
 581                 break;
 582
 583         // --- Local close States
 584         case TCP_ST_FIN_WAIT1:
 585                 if( Header->Flags & TCP_FLAG_FIN )
 586                 {
 587                         Connection->State = TCP_ST_CLOSING;
 588                         Log_Debug("TCP", "Conn %p closed, sent FIN and recieved FIN", Connection);
 589                         VFS_MarkError(&Connection->Node, 1);
 590
 591                         // ACK Packet
 592                         Header->DestPort = Header->SourcePort;
 593                         Header->SourcePort = htons(Connection->LocalPort);
 594                         Header->AcknowlegementNumber = Header->SequenceNumber;
 595                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 596                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 597                         Header->Flags = TCP_FLAG_ACK;
 598                         TCP_SendPacket( Connection, Header, 0, NULL );
 599                         break ;
 600                 }
 601
 602                 // TODO: Make sure that the packet is actually ACKing the FIN
 603                 if( Header->Flags & TCP_FLAG_ACK )
 604                 {
 605                         Connection->State = TCP_ST_FIN_WAIT2;
 606                         Log_Debug("TCP", "Conn %p closed, sent FIN ACKed", Connection);
 607                         VFS_MarkError(&Connection->Node, 1);
 608                         return ;
 609                 }
 610                 break;
 611
 612         case TCP_ST_FIN_WAIT2:
 613                 if( Header->Flags & TCP_FLAG_FIN )
 614                 {
 615                         Connection->State = TCP_ST_TIME_WAIT;
 616                         Log_Debug("TCP", "FIN sent and recieved, ACKing and going into TIME WAIT %p FINWAIT-2 -> TIME WAIT", Connection);
 617                         // Send ACK
 618                         Header->DestPort = Header->SourcePort;
 619                         Header->SourcePort = htons(Connection->LocalPort);
 620                         Header->AcknowlegementNumber = Header->SequenceNumber;
 621                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 622                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 623                         Header->Flags = TCP_FLAG_ACK;
 624                         TCP_SendPacket( Connection, Header, 0, NULL );
 625                 }
 626                 break;
 627
 628         case TCP_ST_CLOSING:
 629                 // TODO: Make sure that the packet is actually ACKing the FIN
 630                 if( Header->Flags & TCP_FLAG_ACK )
 631                 {
 632                         Connection->State = TCP_ST_TIME_WAIT;
 633                         Log_Debug("TCP", "Conn %p CLOSING -> TIME WAIT", Connection);
 634                         VFS_MarkError(&Connection->Node, 1);
 635                         return ;
 636                 }
 637                 break;
 638
 639         // --- Closed (or near closed) states) ---
 640         case TCP_ST_TIME_WAIT:
 641                 Log_Log("TCP", "Packets on Time-Wait, ignored");
 642                 break;
 643
 644         case TCP_ST_FINISHED:
 645                 Log_Log("TCP", "Packets when CLOSED, ignoring");
 646                 break;
 647
 648         //default:
 649         //      Log_Warning("TCP", "Unhandled TCP state %i", Connection->State);
 650         //      break;
 651         }
 652
 653 }
 654
 655 /**
 656  * \brief Appends a packet to the recieved list
 657  * \param Connection    Connection structure
 658  * \param Data  Packet contents
 659  * \param Length        Length of \a Data
 660  */
 661 int TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length)
 662 {
 663         Mutex_Acquire( &Connection->lRecievedPackets );
 664
 665         if(Connection->RecievedBuffer->Length + Length > Connection->RecievedBuffer->Space )
 666         {
 667                 VFS_MarkAvaliable(&Connection->Node, 1);
 668                 Log_Error("TCP", "Buffer filled, packet dropped (:%i) - %i + %i > %i",
 669                         Connection->LocalPort, Connection->RecievedBuffer->Length, Length,
 670                         Connection->RecievedBuffer->Space
 671                         );
 672                 Mutex_Release( &Connection->lRecievedPackets );
 673                 return 1;
 674         }
 675
 676         RingBuffer_Write( Connection->RecievedBuffer, Data, Length );
 677
 678         VFS_MarkAvaliable(&Connection->Node, 1);
 679
 680         Mutex_Release( &Connection->lRecievedPackets );
 681         return 0;
 682 }
 683
 684 /**
 685  * \brief Updates the connections recieved list from the future list
 686  * \param Connection    Connection structure
 687  *
 688  * Updates the recieved packets list with packets from the future (out
 689  * of order) packets list that are now able to be added in direct
 690  * sequence.
 691  */
 692 void TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection)
 693 {
 694         #if CACHE_FUTURE_PACKETS_IN_BYTES
 695         // Calculate length of contiguous bytes
 696          int    length = Connection->HighestSequenceRcvd - Connection->NextSequenceRcv;
 697         Uint32  index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 698         LOG("length=%i, index=%i", length, index);
 699         for( int i = 0; i < length; i ++ )
 700         {
 701                  int    bit = index % 8;
 702                 Uint8   bitfield_byte = Connection->FuturePacketValidBytes[index / 8];
 703                 if( (bitfield_byte & (1 << bit)) == 0 ) {
 704                         length = i;
 705                         break;
 706                 }
 707
 708                 if( bitfield_byte == 0xFF ) {
 709                          int    inc = 8 - bit;
 710                         i += inc - 1;
 711                         index += inc;
 712                 }
 713                 else {
 714                         index ++;
 715                 }
 716                 if(index > TCP_WINDOW_SIZE)
 717                         index -= TCP_WINDOW_SIZE;
 718         }
 719
 720         index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 721
 722         // Write data to to the ring buffer
 723         if( TCP_WINDOW_SIZE - index > length )
 724         {
 725                 // Simple case
 726                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, length );
 727         }
 728         else
 729         {
 730                  int    endLen = TCP_WINDOW_SIZE - index;
 731                 // 2-part case
 732                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, endLen );
 733                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData, endLen - length );
 734         }
 735
 736         // Mark (now saved) bytes as invalid
 737         // - Align index
 738         while(index % 8 && length > 0)
 739         {
 740                 Connection->FuturePacketData[index] = 0;
 741                 Connection->FuturePacketValidBytes[index/8] &= ~(1 << (index%8));
 742                 index ++;
 743                 if(index > TCP_WINDOW_SIZE)
 744                         index -= TCP_WINDOW_SIZE;
 745                 length --;
 746         }
 747         while( length > 7 )
 748         {
 749                 Connection->FuturePacketData[index] = 0;
 750                 Connection->FuturePacketValidBytes[index/8] = 0;
 751                 length -= 8;
 752                 index += 8;
 753                 if(index > TCP_WINDOW_SIZE)
 754                         index -= TCP_WINDOW_SIZE;
 755         }
 756         while(length)
 757         {
 758                 Connection->FuturePacketData[index] = 0;
 759                 Connection->FuturePacketData[index/8] &= ~(1 << (index%8));
 760                 index ++;
 761                 if(index > TCP_WINDOW_SIZE)
 762                         index -= TCP_WINDOW_SIZE;
 763                 length --;
 764         }
 765
 766         #else
 767         tTCPStoredPacket        *pkt;
 768         for(;;)
 769         {
 770                 SHORTLOCK( &Connection->lFuturePackets );
 771
 772                 // Clear out duplicates from cache
 773                 // - If a packet has just been recieved, and it is expected, then
 774                 //   (since NextSequenceRcv = rcvd->Sequence + rcvd->Length) all
 775                 //   packets in cache that are smaller than the next expected
 776                 //   are now defunct.
 777                 pkt = Connection->FuturePackets;
 778                 while(pkt && pkt->Sequence < Connection->NextSequenceRcv)
 779                 {
 780                         tTCPStoredPacket        *next = pkt->Next;
 781                         free(pkt);
 782                         pkt = next;
 783                 }
 784
 785                 // If there's no packets left in cache, stop looking
 786                 if(!pkt || pkt->Sequence > Connection->NextSequenceRcv) {
 787                         SHORTREL( &Connection->lFuturePackets );
 788                         return;
 789                 }
 790
 791                 // Delete packet from future list
 792                 Connection->FuturePackets = pkt->Next;
 793
 794                 // Release list
 795                 SHORTREL( &Connection->lFuturePackets );
 796
 797                 // Looks like we found one
 798                 TCP_INT_AppendRecieved(Connection, pkt->Data, pkt->Length);
 799                 if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 800                         Connection->HighestSequenceRcvd += pkt->Length;
 801                 Connection->NextSequenceRcv += pkt->Length;
 802                 free(pkt);
 803         }
 804         #endif
 805 }
 806
 807 void TCP_int_SendDelayedACK(void *ConnPtr)
 808 {
 809         TCP_INT_SendACK(ConnPtr, "DACK Timeout");
 810 }
 811
 812 void TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason)
 813 {
 814         tTCPHeader      hdr;
 815         // ACK Packet
 816         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
 817         hdr.DestPort = htons(Connection->RemotePort);
 818         hdr.SourcePort = htons(Connection->LocalPort);
 819         hdr.AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 820         hdr.SequenceNumber = htonl(Connection->NextSequenceSend);
 821         hdr.WindowSize = htons(TCP_WINDOW_SIZE);
 822         hdr.Flags = TCP_FLAG_ACK;       // TODO: Determine if SYN is wanted too
 823         hdr.Checksum = 0;       // TODO: Checksum
 824         hdr.UrgentPointer = 0;
 825         Log_Debug("TCP", "Sending ACK for 0x%08x (%s)", Connection->NextSequenceRcv, Reason);
 826         TCP_SendPacket( Connection, &hdr, 0, NULL );
 827         //Connection->NextSequenceSend ++;
 828         Connection->LastACKSequence = Connection->NextSequenceRcv;
 829 }
 830
 831 /**
 832  * \fn Uint16 TCP_GetUnusedPort()
 833  * \brief Gets an unused port and allocates it
 834  */
 835 Uint16 TCP_GetUnusedPort()
 836 {
 837         Uint16  ret;
 838
 839         // Get Next outbound port
 840         ret = giTCP_NextOutPort++;
 841         while( gaTCP_PortBitmap[ret/32] & (1UL << (ret%32)) )
 842         {
 843                 ret ++;
 844                 giTCP_NextOutPort++;
 845                 if(giTCP_NextOutPort == 0x10000) {
 846                         ret = giTCP_NextOutPort = TCP_MIN_DYNPORT;
 847                 }
 848         }
 849
 850         // Mark the new port as used
 851         gaTCP_PortBitmap[ret/32] |= 1 << (ret%32);
 852
 853         return ret;
 854 }
 855
 856 /**
 857  * \fn int TCP_AllocatePort(Uint16 Port)
 858  * \brief Marks a port as used
 859  */
 860 int TCP_AllocatePort(Uint16 Port)
 861 {
 862         // Check if the port has already been allocated
 863         if( gaTCP_PortBitmap[Port/32] & (1 << (Port%32)) )
 864                 return 0;
 865
 866         // Allocate
 867         gaTCP_PortBitmap[Port/32] |= 1 << (Port%32);
 868
 869         return 1;
 870 }
 871
 872 /**
 873  * \fn int TCP_DeallocatePort(Uint16 Port)
 874  * \brief Marks a port as unused
 875  */
 876 int TCP_DeallocatePort(Uint16 Port)
 877 {
 878         // Check if the port has already been allocated
 879         if( !(gaTCP_PortBitmap[Port/32] & (1 << (Port%32))) )
 880                 return 0;
 881
 882         // Allocate
 883         gaTCP_PortBitmap[Port/32] &= ~(1 << (Port%32));
 884
 885         return 1;
 886 }
 887
 888 tTCPConnection *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State)
 889 {
 890         tTCPConnection  *conn = calloc( sizeof(tTCPConnection) + TCP_WINDOW_SIZE + TCP_WINDOW_SIZE/8, 1 );
 891
 892         conn->State = State;
 893         conn->Interface = Interface;
 894         conn->LocalPort = -1;
 895         conn->RemotePort = -1;
 896
 897         conn->Node.ReferenceCount = 1;
 898         conn->Node.ImplPtr = conn;
 899         conn->Node.NumACLs = 1;
 900         conn->Node.ACLs = &gVFS_ACL_EveryoneRW;
 901         conn->Node.Type = &gTCP_ClientNodeType;
 902         conn->Node.BufferFull = 1;      // Cleared when connection opens
 903
 904         conn->RecievedBuffer = RingBuffer_Create( TCP_RECIEVE_BUFFER_SIZE );
 905         #if 0
 906         conn->SentBuffer = RingBuffer_Create( TCP_SEND_BUFFER_SIZE );
 907         Semaphore_Init(conn->SentBufferSpace, 0, TCP_SEND_BUFFER_SIZE, "TCP SentBuffer", conn->Name);
 908         #endif
 909
 910         #if CACHE_FUTURE_PACKETS_IN_BYTES
 911         // Future recieved data (ahead of the expected sequence number)
 912         conn->FuturePacketData = (Uint8*)conn + sizeof(tTCPConnection);
 913         conn->FuturePacketValidBytes = conn->FuturePacketData + TCP_WINDOW_SIZE;
 914         #endif
 915
 916         conn->DeferredACKTimer = Time_AllocateTimer( TCP_int_SendDelayedACK, conn);
 917         return conn;
 918 }
 919
 920 // --- Server
 921 tVFS_Node *TCP_Server_Init(tInterface *Interface)
 922 {
 923         tTCPListener    *srv;
 924
 925         srv = calloc( 1, sizeof(tTCPListener) );
 926
 927         if( srv == NULL ) {
 928                 Log_Warning("TCP", "malloc failed for listener (%i) bytes", sizeof(tTCPListener));
 929                 return NULL;
 930         }
 931
 932         srv->Interface = Interface;
 933         srv->Port = 0;
 934         srv->NextID = 0;
 935         srv->Connections = NULL;
 936         srv->ConnectionsTail = NULL;
 937         srv->NewConnections = NULL;
 938         srv->Next = NULL;
 939         srv->Node.Flags = VFS_FFLAG_DIRECTORY;
 940         srv->Node.Size = -1;
 941         srv->Node.ImplPtr = srv;
 942         srv->Node.NumACLs = 1;
 943         srv->Node.ACLs = &gVFS_ACL_EveryoneRW;
 944         srv->Node.Type = &gTCP_ServerNodeType;
 945
 946         SHORTLOCK(&glTCP_Listeners);
 947         srv->Next = gTCP_Listeners;
 948         gTCP_Listeners = srv;
 949         SHORTREL(&glTCP_Listeners);
 950
 951         return &srv->Node;
 952 }
 953
 954 /**
 955  * \brief Wait for a new connection and return the connection ID
 956  * \note Blocks until a new connection is made
 957  * \param Node  Server node
 958  * \param Pos   Position (ignored)
 959  */
 960 int TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Dest[FILENAME_MAX])
 961 {
 962         tTCPListener    *srv = Node->ImplPtr;
 963         tTCPConnection  *conn;
 964
 965         ENTER("pNode iPos", Node, Pos);
 966
 967         Log_Log("TCP", "Thread %i waiting for a connection", Threads_GetTID());
 968         Semaphore_Wait( &srv->WaitingConnections, 1 );
 969
 970         SHORTLOCK(&srv->lConnections);
 971         // Increment the new list (the current connection is still on the
 972         // normal list)
 973         conn = srv->NewConnections;
 974         srv->NewConnections = conn->Next;
 975
 976         if( srv->NewConnections == NULL )
 977                 VFS_MarkAvaliable( Node, 0 );
 978
 979         SHORTREL( &srv->lConnections );
 980
 981         LOG("conn = %p", conn);
 982         LOG("srv->Connections = %p", srv->Connections);
 983         LOG("srv->NewConnections = %p", srv->NewConnections);
 984         LOG("srv->ConnectionsTail = %p", srv->ConnectionsTail);
 985
 986         itoa(Dest, conn->Node.ImplInt, 16, 8, '0');
 987         Log_Log("TCP", "Thread %i got connection '%s'", Threads_GetTID(), Dest);
 988         LEAVE('i', 0);
 989         return 0;
 990 }
 991
 992 /**
 993  * \brief Gets a client connection node
 994  * \param Node  Server node
 995  * \param Name  Hexadecimal ID of the node
 996  */
 997 tVFS_Node *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags)
 998 {
 999         tTCPConnection  *conn;
1000         tTCPListener    *srv = Node->ImplPtr;
1001         char    tmp[9];
1002          int    id = atoi(Name);
1003
1004         ENTER("pNode sName", Node, Name);
1005
1006         // Check for a non-empty name
1007         if( Name[0] )
1008         {
1009                 // Sanity Check
1010                 itoa(tmp, id, 16, 8, '0');
1011                 if(strcmp(tmp, Name) != 0) {
1012                         LOG("'%s' != '%s' (%08x)", Name, tmp, id);
1013                         LEAVE('n');
1014                         return NULL;
1015                 }
1016
1017                 Log_Debug("TCP", "srv->Connections = %p", srv->Connections);
1018                 Log_Debug("TCP", "srv->NewConnections = %p", srv->NewConnections);
1019                 Log_Debug("TCP", "srv->ConnectionsTail = %p", srv->ConnectionsTail);
1020
1021                 // Search
1022                 SHORTLOCK( &srv->lConnections );
1023                 for(conn = srv->Connections;
1024                         conn;
1025                         conn = conn->Next)
1026                 {
1027                         LOG("conn->Node.ImplInt = %i", conn->Node.ImplInt);
1028                         if(conn->Node.ImplInt == id)    break;
1029                 }
1030                 SHORTREL( &srv->lConnections );
1031
1032                 // If not found, ret NULL
1033                 if(!conn) {
1034                         LOG("Connection %i not found", id);
1035                         LEAVE('n');
1036                         return NULL;
1037                 }
1038         }
1039         // Empty Name - Check for a new connection and if it's there, open it
1040         else
1041         {
1042                 SHORTLOCK( &srv->lConnections );
1043                 conn = srv->NewConnections;
1044                 if( conn != NULL )
1045                         srv->NewConnections = conn->Next;
1046                 VFS_MarkAvaliable( Node, srv->NewConnections != NULL );
1047                 SHORTREL( &srv->lConnections );
1048                 if( !conn ) {
1049                         LOG("No new connections");
1050                         LEAVE('n');
1051                         return NULL;
1052                 }
1053         }
1054
1055         // Return node
1056         LEAVE('p', &conn->Node);
1057         return &conn->Node;
1058 }
1059
1060 /**
1061  * \brief Handle IOCtl calls
1062  */
1063 int TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data)
1064 {
1065         tTCPListener    *srv = Node->ImplPtr;
1066
1067         switch(ID)
1068         {
1069         case 4: // Get/Set Port
1070                 if(!Data)       // Get Port
1071                         return srv->Port;
1072
1073                 if(srv->Port)   // Wait, you can't CHANGE the port
1074                         return -1;
1075
1076                 if(!CheckMem(Data, sizeof(Uint16)))     // Sanity check
1077                         return -1;
1078
1079                 // Permissions check
1080                 if(Threads_GetUID() != 0
1081                 && *(Uint16*)Data != 0
1082                 && *(Uint16*)Data < 1024)
1083                         return -1;
1084
1085                 // TODO: Check if a port is in use
1086
1087                 // Set Port
1088                 srv->Port = *(Uint16*)Data;
1089                 if(srv->Port == 0)      // Allocate a random port
1090                         srv->Port = TCP_GetUnusedPort();
1091                 else    // Else, mark this as used
1092                         TCP_AllocatePort(srv->Port);
1093
1094                 Log_Log("TCP", "Server %p listening on port %i", srv, srv->Port);
1095
1096                 return srv->Port;
1097         }
1098         return 0;
1099 }
1100
1101 void TCP_Server_Close(tVFS_Node *Node)
1102 {
1103         free(Node->ImplPtr);
1104 }
1105
1106 // --- Client
1107 /**
1108  * \brief Create a client node
1109  */
1110 tVFS_Node *TCP_Client_Init(tInterface *Interface)
1111 {
1112         tTCPConnection  *conn = TCP_int_CreateConnection(Interface, TCP_ST_CLOSED);
1113
1114         SHORTLOCK(&glTCP_OutbountCons);
1115         conn->Next = gTCP_OutbountCons;
1116         gTCP_OutbountCons = conn;
1117         SHORTREL(&glTCP_OutbountCons);
1118
1119         return &conn->Node;
1120 }
1121
1122 /**
1123  * \brief Wait for a packet and return it
1124  * \note If \a Length is smaller than the size of the packet, the rest
1125  *       of the packet's data will be discarded.
1126  */
1127 size_t TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags)
1128 {
1129         tTCPConnection  *conn = Node->ImplPtr;
1130         size_t  len;
1131
1132         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1133         LOG("conn = %p {State:%i}", conn, conn->State);
1134
1135         // If the connection has been closed (state > ST_OPEN) then clear
1136         // any stale data in the buffer (until it is empty (until it is empty))
1137         if( conn->State > TCP_ST_OPEN )
1138         {
1139                 Mutex_Acquire( &conn->lRecievedPackets );
1140                 len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1141                 Mutex_Release( &conn->lRecievedPackets );
1142
1143                 if( len == 0 ) {
1144                         VFS_MarkAvaliable(Node, 0);
1145                         errno = 0;
1146                         LEAVE('i', -1);
1147                         return -1;
1148                 }
1149
1150                 LEAVE('i', len);
1151                 return len;
1152         }
1153
1154         // Wait
1155         {
1156                 tTime   *timeout = NULL;
1157                 tTime   timeout_zero = 0;
1158                 if( Flags & VFS_IOFLAG_NOBLOCK )
1159                         timeout = &timeout_zero;
1160                 if( !VFS_SelectNode(Node, VFS_SELECT_READ|VFS_SELECT_ERROR, timeout, "TCP_Client_Read") ) {
1161                         errno = EWOULDBLOCK;
1162                         LEAVE('i', -1);
1163                         return -1;
1164                 }
1165         }
1166
1167         // Lock list and read as much as possible (up to `Length`)
1168         Mutex_Acquire( &conn->lRecievedPackets );
1169         len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1170
1171         if( len == 0 || conn->RecievedBuffer->Length == 0 ) {
1172                 LOG("Marking as none avaliable (len = %i)", len);
1173                 VFS_MarkAvaliable(Node, 0);
1174         }
1175
1176         // Release the lock (we don't need it any more)
1177         Mutex_Release( &conn->lRecievedPackets );
1178
1179         LEAVE('i', len);
1180         return len;
1181 }
1182
1183 /**
1184  * \brief Send a data packet on a connection
1185  */
1186 void TCP_INT_SendDataPacket(tTCPConnection *Connection, size_t Length, const void *Data)
1187 {
1188         char    buf[sizeof(tTCPHeader)+Length];
1189         tTCPHeader      *packet = (void*)buf;
1190
1191         // - Stop Delayed ACK timer (as this data packet ACKs)
1192         Time_RemoveTimer(Connection->DeferredACKTimer);
1193
1194         packet->SourcePort = htons(Connection->LocalPort);
1195         packet->DestPort = htons(Connection->RemotePort);
1196         packet->DataOffset = (sizeof(tTCPHeader)/4)*16;
1197         packet->WindowSize = htons(TCP_WINDOW_SIZE);
1198
1199         packet->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
1200         packet->SequenceNumber = htonl(Connection->NextSequenceSend);
1201         packet->Flags = TCP_FLAG_PSH|TCP_FLAG_ACK;      // Hey, ACK if you can!
1202         packet->UrgentPointer = 0;
1203
1204         memcpy(packet->Options, Data, Length);
1205
1206         Log_Debug("TCP", "Send sequence 0x%08x", Connection->NextSequenceSend);
1207 #if HEXDUMP_OUTGOING
1208         Debug_HexDump("TCP_INT_SendDataPacket: Data = ", Data, Length);
1209 #endif
1210
1211         TCP_SendPacket( Connection, packet, Length, Data );
1212
1213         Connection->NextSequenceSend += Length;
1214 }
1215
1216 /**
1217  * \brief Send some bytes on a connection
1218  */
1219 size_t TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags)
1220 {
1221         tTCPConnection  *conn = Node->ImplPtr;
1222         size_t  rem = Length;
1223
1224         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1225
1226 //      #if DEBUG
1227 //      Debug_HexDump("TCP_Client_Write: Buffer = ",
1228 //              Buffer, Length);
1229 //      #endif
1230
1231         // Don't allow a write to a closed connection
1232         if( conn->State > TCP_ST_OPEN ) {
1233                 VFS_MarkError(Node, 1);
1234                 errno = 0;
1235                 LEAVE('i', -1);
1236                 return -1;
1237         }
1238
1239         // Wait
1240         {
1241                 tTime   *timeout = NULL;
1242                 tTime   timeout_zero = 0;
1243                 if( Flags & VFS_IOFLAG_NOBLOCK )
1244                         timeout = &timeout_zero;
1245                 if( !VFS_SelectNode(Node, VFS_SELECT_WRITE|VFS_SELECT_ERROR, timeout, "TCP_Client_Write") ) {
1246                         errno = EWOULDBLOCK;
1247                         LEAVE('i', -1);
1248                         return -1;
1249                 }
1250         }
1251
1252         do
1253         {
1254                  int    len = (rem < TCP_MAX_PACKET_SIZE) ? rem : TCP_MAX_PACKET_SIZE;
1255
1256                 #if 0
1257                 // Wait for space in the buffer
1258                 Semaphore_Signal( &Connection->SentBufferSpace, len );
1259
1260                 // Save data to buffer (and update the length read by the ammount written)
1261                 len = RingBuffer_Write( &Connection->SentBuffer, Buffer, len);
1262                 #endif
1263
1264                 // Send packet
1265                 TCP_INT_SendDataPacket(conn, len, Buffer);
1266
1267                 Buffer += len;
1268                 rem -= len;
1269         } while( rem > 0 );
1270
1271         LEAVE('i', Length);
1272         return Length;
1273 }
1274
1275 /**
1276  * \brief Open a connection to another host using TCP
1277  * \param Conn  Connection structure
1278  */
1279 void TCP_StartConnection(tTCPConnection *Conn)
1280 {
1281         tTCPHeader      hdr = {0};
1282
1283         Conn->State = TCP_ST_SYN_SENT;
1284
1285         hdr.SourcePort = htons(Conn->LocalPort);
1286         hdr.DestPort = htons(Conn->RemotePort);
1287         Conn->NextSequenceSend = rand();
1288         hdr.SequenceNumber = htonl(Conn->NextSequenceSend);
1289         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
1290         hdr.Flags = TCP_FLAG_SYN;
1291         hdr.WindowSize = htons(TCP_WINDOW_SIZE);        // Max
1292         hdr.Checksum = 0;       // TODO
1293
1294         TCP_SendPacket( Conn, &hdr, 0, NULL );
1295
1296         Conn->NextSequenceSend ++;
1297         Conn->State = TCP_ST_SYN_SENT;
1298
1299         return ;
1300 }
1301
1302 /**
1303  * \brief Control a client socket
1304  */
1305 int TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data)
1306 {
1307         tTCPConnection  *conn = Node->ImplPtr;
1308
1309         ENTER("pNode iID pData", Node, ID, Data);
1310
1311         switch(ID)
1312         {
1313         case 4: // Get/Set local port
1314                 if(!Data)
1315                         LEAVE_RET('i', conn->LocalPort);
1316                 if(conn->State != TCP_ST_CLOSED)
1317                         LEAVE_RET('i', -1);
1318                 if(!CheckMem(Data, sizeof(Uint16)))
1319                         LEAVE_RET('i', -1);
1320
1321                 if(Threads_GetUID() != 0 && *(Uint16*)Data < 1024)
1322                         LEAVE_RET('i', -1);
1323
1324                 conn->LocalPort = *(Uint16*)Data;
1325                 LEAVE_RET('i', conn->LocalPort);
1326
1327         case 5: // Get/Set remote port
1328                 if(!Data)       LEAVE_RET('i', conn->RemotePort);
1329                 if(conn->State != TCP_ST_CLOSED)        LEAVE_RET('i', -1);
1330                 if(!CheckMem(Data, sizeof(Uint16)))     LEAVE_RET('i', -1);
1331                 conn->RemotePort = *(Uint16*)Data;
1332                 LEAVE_RET('i', conn->RemotePort);
1333
1334         case 6: // Set Remote IP
1335                 if( conn->State != TCP_ST_CLOSED )
1336                         LEAVE_RET('i', -1);
1337                 if( conn->Interface->Type == 4 )
1338                 {
1339                         if(!CheckMem(Data, sizeof(tIPv4)))      LEAVE_RET('i', -1);
1340                         conn->RemoteIP.v4 = *(tIPv4*)Data;
1341                 }
1342                 else if( conn->Interface->Type == 6 )
1343                 {
1344                         if(!CheckMem(Data, sizeof(tIPv6)))      LEAVE_RET('i', -1);
1345                         conn->RemoteIP.v6 = *(tIPv6*)Data;
1346                 }
1347                 LEAVE_RET('i', 0);
1348
1349         case 7: // Connect
1350                 if(conn->LocalPort == 0xFFFF)
1351                         conn->LocalPort = TCP_GetUnusedPort();
1352                 if(conn->RemotePort == -1)
1353                         LEAVE_RET('i', 0);
1354
1355                 {
1356                         tTime   timeout = conn->Interface->TimeoutDelay;
1357
1358                         TCP_StartConnection(conn);
1359                         VFS_SelectNode(&conn->Node, VFS_SELECT_WRITE, &timeout, "TCP Connection");
1360                         if( conn->State == TCP_ST_SYN_SENT )
1361                                 LEAVE_RET('i', 0);
1362                 }
1363
1364                 LEAVE_RET('i', 1);
1365
1366         // Get recieve buffer length
1367         case 8:
1368                 LEAVE_RET('i', conn->RecievedBuffer->Length);
1369         }
1370
1371         return 0;
1372 }
1373
1374 void TCP_Client_Close(tVFS_Node *Node)
1375 {
1376         tTCPConnection  *conn = Node->ImplPtr;
1377         tTCPHeader      packet;
1378
1379         ENTER("pNode", Node);
1380
1381         ASSERT(Node->ReferenceCount != 0);
1382
1383         if( Node->ReferenceCount > 1 ) {
1384                 Node->ReferenceCount --;
1385                 LOG("Dereference only");
1386                 LEAVE('-');
1387                 return ;
1388         }
1389
1390         if( conn->State == TCP_ST_CLOSE_WAIT || conn->State == TCP_ST_OPEN )
1391         {
1392                 packet.SourcePort = htons(conn->LocalPort);
1393                 packet.DestPort = htons(conn->RemotePort);
1394                 packet.DataOffset = (sizeof(tTCPHeader)/4)*16;
1395                 packet.WindowSize = TCP_WINDOW_SIZE;
1396
1397                 packet.AcknowlegementNumber = 0;
1398                 packet.SequenceNumber = htonl(conn->NextSequenceSend);
1399                 packet.Flags = TCP_FLAG_FIN;
1400
1401                 TCP_SendPacket( conn, &packet, 0, NULL );
1402         }
1403
1404         switch( conn->State )
1405         {
1406         case TCP_ST_CLOSED:
1407                 Log_Warning("TCP", "Closing connection that was never opened");
1408                 break;
1409         case TCP_ST_CLOSE_WAIT:
1410                 conn->State = TCP_ST_LAST_ACK;
1411                 break;
1412         case TCP_ST_OPEN:
1413                 conn->State = TCP_ST_FIN_WAIT1;
1414                 while( conn->State == TCP_ST_FIN_WAIT1 )        Threads_Yield();
1415                 break;
1416         default:
1417                 Log_Warning("TCP", "Unhandled connection state %i in TCP_Client_Close",
1418                         conn->State);
1419                 break;
1420         }
1421
1422         Time_RemoveTimer(conn->DeferredACKTimer);
1423         Time_FreeTimer(conn->DeferredACKTimer);
1424         free(conn);
1425
1426         LEAVE('-');
1427 }
1428
1429 /**
1430  * \brief Checks if a value is between two others (after taking into account wrapping)
1431  */
1432 int WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue)
1433 {
1434         if( MaxValue < 0xFFFFFFFF )
1435         {
1436                 Lower %= MaxValue + 1;
1437                 Value %= MaxValue + 1;
1438                 Higher %= MaxValue + 1;
1439         }
1440
1441         // Simple Case, no wrap ?
1442         //       Lower Value Higher
1443         // | ... + ... + ... + ... |
1444
1445         if( Lower < Higher ) {
1446                 return Lower < Value && Value < Higher;
1447         }
1448         // Higher has wrapped below lower
1449
1450         // Value > Lower ?
1451         //       Higher Lower Value
1452         // | ... +  ... + ... + ... |
1453         if( Value > Lower ) {
1454                 return 1;
1455         }
1456
1457         // Value < Higher ?
1458         //       Value Higher Lower
1459         // | ... + ... +  ... + ... |
1460         if( Value < Higher ) {
1461                 return 1;
1462         }
1463
1464         return 0;
1465 }