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        100
  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_int_SendPacket(tInterface *Interface, const void *Dest, tTCPHeader *Header, size_t Length, const void *Data);
  30 void    TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer);
  31 void    TCP_IPError(tInterface *Interface, tIPErrorMode Mode, const void *Address, int Length, const void *Buffer);
  32  int    TCP_INT_HandleServerPacket(tInterface *Interface, tTCPListener *Server, const void *Address, tTCPHeader *Header, size_t Length);
  33  int    TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length);
  34 int     TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length);
  35 void    TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection);
  36 void    TCP_int_SendDelayedACK(void *ConnPtr);
  37 void    TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason);
  38 Uint16  TCP_GetUnusedPort();
  39  int    TCP_AllocatePort(Uint16 Port);
  40  int    TCP_DeallocatePort(Uint16 Port);
  41 tTCPConnection  *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State);
  42 void    TCP_int_FreeTCB(tTCPConnection *Connection);
  43 // --- Server
  44 tVFS_Node       *TCP_Server_Init(tInterface *Interface);
  45  int    TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Name[FILENAME_MAX]);
  46 tVFS_Node       *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags);
  47  int    TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data);
  48 void    TCP_Server_Close(tVFS_Node *Node);
  49 // --- Client
  50 tVFS_Node       *TCP_Client_Init(tInterface *Interface);
  51 size_t  TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags);
  52 size_t  TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags);
  53  int    TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data);
  54 void    TCP_Client_Close(tVFS_Node *Node);
  55 // --- Helpers
  56  int    WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue);
  57 Uint32  GetRelative(Uint32 Base, Uint32 Value);
  58
  59 // === TEMPLATES ===
  60 tSocketFile     gTCP_ServerFile = {NULL, "tcps", TCP_Server_Init};
  61 tSocketFile     gTCP_ClientFile = {NULL, "tcpc", TCP_Client_Init};
  62 tVFS_NodeType   gTCP_ServerNodeType = {
  63         .TypeName = "TCP Server",
  64         .ReadDir = TCP_Server_ReadDir,
  65         .FindDir = TCP_Server_FindDir,
  66         .IOCtl   = TCP_Server_IOCtl,
  67         .Close   = TCP_Server_Close
  68         };
  69 tVFS_NodeType   gTCP_ClientNodeType = {
  70         .TypeName = "TCP Client/Connection",
  71         .Flags = VFS_NODETYPEFLAG_STREAM,
  72         .Read  = TCP_Client_Read,
  73         .Write = TCP_Client_Write,
  74         .IOCtl = TCP_Client_IOCtl,
  75         .Close = TCP_Client_Close
  76         };
  77
  78 // === GLOBALS ===
  79  int    giTCP_NumHalfopen = 0;
  80 tShortSpinlock  glTCP_Listeners;
  81 tTCPListener    *gTCP_Listeners;
  82 tShortSpinlock  glTCP_OutbountCons;
  83 tTCPConnection  *gTCP_OutbountCons;
  84 Uint32  gaTCP_PortBitmap[0x800];
  85  int    giTCP_NextOutPort = TCP_MIN_DYNPORT;
  86
  87 // === CODE ===
  88 /**
  89  * \brief Initialise the TCP Layer
  90  *
  91  * Registers the client and server files and the GetPacket callback
  92  */
  93 void TCP_Initialise(void)
  94 {
  95         giTCP_NextOutPort += rand()%128;
  96         IPStack_AddFile(&gTCP_ServerFile);
  97         IPStack_AddFile(&gTCP_ClientFile);
  98         IPv4_RegisterCallback(IP4PROT_TCP, TCP_GetPacket, TCP_IPError);
  99         IPv6_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
 100 }
 101
 102 /**
 103  * \brief Sends a packet from the specified connection, calculating the checksums
 104  * \param Conn  Connection
 105  * \param Length        Length of data
 106  * \param Data  Packet data (cast as a TCP Header)
 107  */
 108 void TCP_SendPacket( tTCPConnection *Conn, tTCPHeader *Header, size_t Length, const void *Data )
 109 {
 110         TCP_int_SendPacket(Conn->Interface, &Conn->RemoteIP, Header, Length, Data);
 111 }
 112
 113 Uint16 TCP_int_CalculateChecksum(int AddrType, const void *LAddr, const void *RAddr,
 114         size_t HeaderLength, const tTCPHeader *Header, size_t DataLength, const void *Data)
 115 {
 116         size_t packlen = HeaderLength + DataLength;
 117         Uint16  checksum[3];
 118
 119         switch(AddrType)
 120         {
 121         case 4: {
 122                 Uint32  buf[3];
 123                 buf[0] = ((tIPv4*)LAddr)->L;
 124                 buf[1] = ((tIPv4*)RAddr)->L;
 125                 buf[2] = htonl( (packlen) | (IP4PROT_TCP<<16) | (0<<24) );
 126                 checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 127                 break; }
 128         case 6: {
 129                 Uint32  buf[4+4+1+1];
 130                 memcpy(&buf[0], LAddr, 16);
 131                 memcpy(&buf[4], RAddr, 16);
 132                 buf[8] = htonl(packlen);
 133                 buf[9] = htonl(IP4PROT_TCP);
 134                 checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 135                 break; }
 136         default:
 137                 return 0;
 138         }
 139         checksum[1] = htons( ~IPv4_Checksum(Header, HeaderLength) );
 140         checksum[2] = htons( ~IPv4_Checksum(Data, DataLength) );
 141
 142         return htons( IPv4_Checksum(checksum, sizeof(checksum)) );
 143 }
 144
 145 void TCP_int_SendPacket(tInterface *Interface, const void *Dest, tTCPHeader *Header, size_t Length, const void *Data )
 146 {
 147         tIPStackBuffer  *buffer = IPStack_Buffer_CreateBuffer(2 + IPV4_BUFFERS);
 148         if( Length > 0 )
 149                 IPStack_Buffer_AppendSubBuffer(buffer, Length, 0, Data, NULL, NULL);
 150         IPStack_Buffer_AppendSubBuffer(buffer, sizeof(*Header), 0, Header, NULL, NULL);
 151
 152         #if TCP_DEBUG
 153         Log_Log("TCP", "TCP_int_SendPacket: <Local>:%i to [%s]:%i (%i data), Flags = %s%s%s%s%s%s%s%s",
 154                 ntohs(Header->SourcePort),
 155                 IPStack_PrintAddress(Interface->Type, Dest),
 156                 ntohs(Header->DestPort),
 157                 Length,
 158                 (Header->Flags & TCP_FLAG_CWR) ? "CWR " : "",
 159                 (Header->Flags & TCP_FLAG_ECE) ? "ECE " : "",
 160                 (Header->Flags & TCP_FLAG_URG) ? "URG " : "",
 161                 (Header->Flags & TCP_FLAG_ACK) ? "ACK " : "",
 162                 (Header->Flags & TCP_FLAG_PSH) ? "PSH " : "",
 163                 (Header->Flags & TCP_FLAG_RST) ? "RST " : "",
 164                 (Header->Flags & TCP_FLAG_SYN) ? "SYN " : "",
 165                 (Header->Flags & TCP_FLAG_FIN) ? "FIN " : ""
 166                 );
 167         #endif
 168
 169         Header->Checksum = 0;
 170         Header->Checksum = TCP_int_CalculateChecksum(Interface->Type, Interface->Address, Dest,
 171                 sizeof(tTCPHeader), Header, Length, Data);
 172
 173         // TODO: Fragment packet
 174
 175         switch( Interface->Type )
 176         {
 177         case 4:
 178                 IPv4_SendPacket(Interface, *(tIPv4*)Dest, IP4PROT_TCP, 0, buffer);
 179                 break;
 180         case 6:
 181                 IPv6_SendPacket(Interface, *(tIPv6*)Dest, IP4PROT_TCP, buffer);
 182                 break;
 183         }
 184 }
 185
 186 void TCP_int_SendRSTTo(tInterface *Interface, const void *Address, size_t Length, const tTCPHeader *Header)
 187 {
 188         tTCPHeader      out_hdr = {0};
 189
 190         out_hdr.DataOffset = (sizeof(out_hdr)/4) << 4;
 191         out_hdr.DestPort = Header->SourcePort;
 192         out_hdr.SourcePort = Header->DestPort;
 193
 194         size_t  data_len = Length - (Header->DataOffset>>4)*4;
 195         out_hdr.AcknowlegementNumber = htonl( ntohl(Header->SequenceNumber) + data_len );
 196         if( Header->Flags & TCP_FLAG_ACK ) {
 197                 out_hdr.Flags = TCP_FLAG_RST;
 198                 out_hdr.SequenceNumber = Header->AcknowlegementNumber;
 199         }
 200         else {
 201                 out_hdr.Flags = TCP_FLAG_RST|TCP_FLAG_ACK;
 202                 out_hdr.SequenceNumber = 0;
 203         }
 204         TCP_int_SendPacket(Interface, Address, &out_hdr, 0, NULL);
 205 }
 206
 207 /**
 208  * \brief Handles a packet from the IP Layer
 209  * \param Interface     Interface the packet arrived from
 210  * \param Address       Pointer to the addres structure
 211  * \param Length        Size of packet in bytes
 212  * \param Buffer        Packet data
 213  */
 214 void TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer)
 215 {
 216         tTCPHeader      *hdr = Buffer;
 217
 218         #if TCP_DEBUG
 219         Log_Log("TCP", "TCP_GetPacket: <Local>:%i from [%s]:%i, Flags = %s%s%s%s%s%s%s%s",
 220                 ntohs(hdr->DestPort),
 221                 IPStack_PrintAddress(Interface->Type, Address),
 222                 ntohs(hdr->SourcePort),
 223                 (hdr->Flags & TCP_FLAG_CWR) ? "CWR " : "",
 224                 (hdr->Flags & TCP_FLAG_ECE) ? "ECE " : "",
 225                 (hdr->Flags & TCP_FLAG_URG) ? "URG " : "",
 226                 (hdr->Flags & TCP_FLAG_ACK) ? "ACK " : "",
 227                 (hdr->Flags & TCP_FLAG_PSH) ? "PSH " : "",
 228                 (hdr->Flags & TCP_FLAG_RST) ? "RST " : "",
 229                 (hdr->Flags & TCP_FLAG_SYN) ? "SYN " : "",
 230                 (hdr->Flags & TCP_FLAG_FIN) ? "FIN " : ""
 231                 );
 232         #endif
 233
 234         if( Length > (hdr->DataOffset >> 4)*4 )
 235         {
 236                 LOG("SequenceNumber = 0x%x", ntohl(hdr->SequenceNumber));
 237 #if HEXDUMP_INCOMING
 238                 Debug_HexDump(
 239                         "TCP_GetPacket: Packet Data = ",
 240                         (Uint8*)hdr + (hdr->DataOffset >> 4)*4,
 241                         Length - (hdr->DataOffset >> 4)*4
 242                         );
 243 #endif
 244         }
 245
 246         // Check Servers
 247         for( tTCPListener *srv = gTCP_Listeners; srv; srv = srv->Next )
 248         {
 249                 // Check if the server is active
 250                 if(srv->Port == 0)      continue;
 251                 // Check the interface
 252                 if(srv->Interface && srv->Interface != Interface)       continue;
 253                 // Check the destination port
 254                 if(srv->Port != htons(hdr->DestPort))   continue;
 255
 256                 Log_Log("TCP", "TCP_GetPacket: Matches server %p", srv);
 257                 // Is this in an established connection?
 258                 for( tTCPConnection *conn = srv->Connections; conn; conn = conn->Next )
 259                 {
 260                         // Check that it is coming in on the same interface
 261                         if(conn->Interface != Interface)        continue;
 262                         // Check Source Port
 263                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 264                         // Check Source IP
 265                         if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 266                                 continue ;
 267
 268                         Log_Log("TCP", "TCP_GetPacket: Matches connection %p", conn);
 269                         // We have a response!
 270                         if( TCP_INT_HandleConnectionPacket(conn, hdr, Length) == 0 )
 271                                 return;
 272                         break ;
 273                 }
 274
 275                 TCP_INT_HandleServerPacket(Interface, srv, Address, hdr, Length);
 276                 return ;
 277         }
 278
 279         // Check Open Connections
 280         {
 281                 for( tTCPConnection *conn = gTCP_OutbountCons; conn; conn = conn->Next )
 282                 {
 283                         // Check that it is coming in on the same interface
 284                         if(conn->Interface != Interface)        continue;
 285
 286                         // Check Source Port
 287                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 288
 289                         // Check Source IP
 290                         if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 291                                 continue ;
 292
 293                         // Handle or fall through
 294                         if( TCP_INT_HandleConnectionPacket(conn, hdr, Length) == 0 )
 295                                 return ;
 296                         break;
 297                 }
 298         }
 299
 300         Log_Log("TCP", "TCP_GetPacket: No Match");
 301         // If not a RST, send a RST
 302         if( !(hdr->Flags & TCP_FLAG_RST) )
 303         {
 304                 TCP_int_SendRSTTo(Interface, Address, Length, hdr);
 305         }
 306 }
 307
 308 void TCP_IPError(tInterface *Interface, tIPErrorMode Mode, const void *Address, int Length, const void *Buffer)
 309 {
 310         if( Length < sizeof(tTCPHeader) )       return ;
 311
 312         const tTCPHeader        *hdr = Buffer;
 313
 314         // TODO: Handle errors for server connections
 315
 316         for( tTCPConnection *conn = gTCP_OutbountCons; conn; conn = conn->Next )
 317         {
 318                 if(conn->Interface != Interface)
 319                         continue;
 320                 if(conn->RemotePort != ntohs(hdr->SourcePort))
 321                         continue;
 322                 if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 323                         continue ;
 324
 325                 // Mark an error on the interface
 326                 VFS_MarkError(&conn->Node, 1);
 327                 return ;
 328         }
 329 }
 330
 331 /*
 332  * Handle packets in LISTEN state
 333  */
 334 int TCP_INT_HandleServerPacket(tInterface *Interface, tTCPListener *Server, const void *Address, tTCPHeader *Header, size_t Length)
 335 {
 336         if( Header->Flags & TCP_FLAG_RST ) {
 337                 LOG("RST, ignore");
 338                 return 0;
 339         }
 340         else if( Header->Flags & TCP_FLAG_ACK ) {
 341                 LOG("ACK, send RST");
 342                 TCP_int_SendRSTTo(Interface, Address, Length, Header);
 343                 return 0;
 344         }
 345         else if( !(Header->Flags & TCP_FLAG_SYN) ) {
 346                 LOG("Other, ignore");
 347                 return 0;
 348         }
 349
 350         Log_Log("TCP", "TCP_GetPacket: Opening Connection");
 351
 352         // TODO: Check security (a TCP Option)
 353         // TODO: Check SEG.PRC
 354         // TODO: Check for halfopen max
 355
 356         tTCPConnection *conn = TCP_int_CreateConnection(Interface, TCP_ST_SYN_RCVD);
 357         conn->LocalPort = Server->Port;
 358         conn->RemotePort = ntohs(Header->SourcePort);
 359
 360         switch(Interface->Type)
 361         {
 362         case 4: conn->RemoteIP.v4 = *(tIPv4*)Address;   break;
 363         case 6: conn->RemoteIP.v6 = *(tIPv6*)Address;   break;
 364         default:        ASSERTC(Interface->Type,==,4);  return 0;
 365         }
 366
 367         conn->NextSequenceRcv = ntohl( Header->SequenceNumber ) + 1;
 368         conn->HighestSequenceRcvd = conn->NextSequenceRcv;
 369         conn->NextSequenceSend = rand();
 370         conn->LastACKSequence = ntohl( Header->SequenceNumber );
 371
 372         conn->Node.ImplInt = Server->NextID ++;
 373         conn->Node.Size = -1;
 374
 375         // Hmm... Theoretically, this lock will never have to wait,
 376         // as the interface is locked to the watching thread, and this
 377         // runs in the watching thread. But, it's a good idea to have
 378         // it, just in case
 379         // Oh, wait, there is a case where a wildcard can be used
 380         // (Server->Interface == NULL) so having the lock is a good idea
 381         SHORTLOCK(&Server->lConnections);
 382         conn->Server = Server;
 383         conn->Prev = Server->ConnectionsTail;
 384         if(Server->Connections) {
 385                 ASSERT(Server->ConnectionsTail);
 386                 Server->ConnectionsTail->Next = conn;
 387         }
 388         else {
 389                 ASSERT(!Server->ConnectionsTail);
 390                 Server->Connections = conn;
 391         }
 392         Server->ConnectionsTail = conn;
 393         if(!Server->NewConnections)
 394                 Server->NewConnections = conn;
 395         VFS_MarkAvaliable( &Server->Node, 1 );
 396         SHORTREL(&Server->lConnections);
 397         Semaphore_Signal(&Server->WaitingConnections, 1);
 398
 399         // Send the SYN ACK
 400         Header->Flags = TCP_FLAG_ACK|TCP_FLAG_SYN;
 401         Header->AcknowlegementNumber = htonl(conn->NextSequenceRcv);
 402         Header->SequenceNumber = htonl(conn->NextSequenceSend);
 403         Header->DestPort = Header->SourcePort;
 404         Header->SourcePort = htons(Server->Port);
 405         Header->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 406         TCP_SendPacket( conn, Header, 0, NULL );
 407         conn->NextSequenceSend ++;
 408         return 0;
 409 }
 410
 411 /**
 412  * \brief Handles a packet sent to a specific connection
 413  * \param Connection    TCP Connection pointer
 414  * \param Header        TCP Packet pointer
 415  * \param Length        Length of the packet
 416  */
 417 int TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length)
 418 {
 419          int    dataLen;
 420         Uint32  sequence_num;
 421
 422         // Silently drop once finished
 423         // TODO: Check if this needs to be here
 424         if( Connection->State == TCP_ST_FINISHED ) {
 425                 Log_Log("TCP", "Packet ignored - connection finnished");
 426                 return 1;
 427         }
 428         if( Connection->State == TCP_ST_FORCE_CLOSE ) {
 429                 Log_Log("TCP", "Packet ignored - connection reset");
 430                 return 1;
 431         }
 432
 433         // Syncronise sequence values
 434         if(Header->Flags & TCP_FLAG_SYN) {
 435                 // TODO: What if the packet also has data?
 436                 if( Connection->LastACKSequence != Connection->NextSequenceRcv )
 437                         TCP_INT_SendACK(Connection, "SYN");
 438                 Connection->NextSequenceRcv = ntohl(Header->SequenceNumber);
 439                 // TODO: Process HighestSequenceRcvd
 440                 // HACK!
 441                 if( Connection->HighestSequenceRcvd == 0 )
 442                         Connection->HighestSequenceRcvd = Connection->NextSequenceRcv;
 443                 Connection->LastACKSequence = Connection->NextSequenceRcv;
 444         }
 445
 446         // Ackowledge a sent packet
 447         if(Header->Flags & TCP_FLAG_ACK) {
 448                 // TODO: Process an ACKed Packet
 449                 LOG("Conn %p, Sent packet 0x%x ACKed", Connection, Header->AcknowlegementNumber);
 450         }
 451
 452         // Get length of data
 453         dataLen = Length - (Header->DataOffset>>4)*4;
 454         LOG("dataLen = %i", dataLen);
 455         #if TCP_DEBUG
 456         Log_Debug("TCP", "State %i, dataLen = %x", Connection->State, dataLen);
 457         #endif
 458
 459         //
 460         // State Machine
 461         //
 462         switch( Connection->State )
 463         {
 464         // Pre-init connection?
 465         case TCP_ST_CLOSED:
 466                 Log_Log("TCP", "Packets to a closed connection?!");
 467                 break;
 468
 469         // --- Init States ---
 470         // SYN sent, expecting SYN-ACK Connection Opening
 471         case TCP_ST_SYN_SENT:
 472                 if( Header->Flags & TCP_FLAG_SYN )
 473                 {
 474                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 475                                 Connection->HighestSequenceRcvd ++;
 476                         Connection->NextSequenceRcv ++;
 477
 478                         if( Header->Flags & TCP_FLAG_ACK )
 479                         {
 480                                 Log_Log("TCP", "ACKing SYN-ACK");
 481                                 Connection->State = TCP_ST_ESTABLISHED;
 482                                 VFS_MarkFull(&Connection->Node, 0);
 483                                 TCP_INT_SendACK(Connection, "SYN-ACK");
 484                         }
 485                         else
 486                         {
 487                                 Log_Log("TCP", "ACKing SYN");
 488                                 Connection->State = TCP_ST_SYN_RCVD;
 489                                 TCP_INT_SendACK(Connection, "SYN");
 490                         }
 491                 }
 492                 break;
 493
 494         // SYN-ACK sent, expecting ACK
 495         case TCP_ST_SYN_RCVD:
 496                 if( Header->Flags & TCP_FLAG_RST )
 497                 {
 498                         Log_Log("TCP", "RST Received, closing");
 499                         Connection->State = TCP_ST_FORCE_CLOSE;
 500                         VFS_MarkError(&Connection->Node, 1);
 501                         return 0;
 502                 }
 503                 if( Header->Flags & TCP_FLAG_ACK )
 504                 {
 505                         // TODO: Handle max half-open limit
 506                         Log_Log("TCP", "Connection fully opened");
 507                         Connection->State = TCP_ST_ESTABLISHED;
 508                         VFS_MarkFull(&Connection->Node, 0);
 509                 }
 510                 break;
 511
 512         // --- Established State ---
 513         case TCP_ST_ESTABLISHED:
 514                 // - Handle State changes
 515                 //
 516                 if( Header->Flags & TCP_FLAG_RST )
 517                 {
 518                         Log_Log("TCP", "Conn %p closed, received RST");
 519                         // Error outstanding transactions
 520                         Connection->State = TCP_ST_FORCE_CLOSE;
 521                         VFS_MarkError(&Connection->Node, 1);
 522                         return 0;
 523                 }
 524                 if( Header->Flags & TCP_FLAG_FIN ) {
 525                         Log_Log("TCP", "Conn %p closed, recieved FIN", Connection);
 526                         VFS_MarkError(&Connection->Node, 1);
 527                         Connection->NextSequenceRcv ++;
 528                         TCP_INT_SendACK(Connection, "FIN Received");
 529                         Connection->State = TCP_ST_CLOSE_WAIT;
 530                         // CLOSE WAIT requires the client to close
 531                         return 0;
 532                 }
 533
 534                 // Check for an empty packet
 535                 if(dataLen == 0) {
 536                         if( Header->Flags == TCP_FLAG_ACK )
 537                         {
 538                                 Log_Log("TCP", "ACK only packet");
 539                                 return 0;
 540                         }
 541                         // TODO: Is this right? (empty packet counts as one byte)
 542                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 543                                 Connection->HighestSequenceRcvd ++;
 544                         Connection->NextSequenceRcv ++;
 545                         Log_Log("TCP", "Empty Packet, inc and ACK the current sequence number");
 546                         TCP_INT_SendACK(Connection, "Empty");
 547                         return 0;
 548                 }
 549
 550                 // NOTES:
 551                 // Flags
 552                 //    PSH - Has Data?
 553                 // /NOTES
 554
 555                 sequence_num = ntohl(Header->SequenceNumber);
 556
 557                 LOG("0x%08x <= 0x%08x < 0x%08x",
 558                         Connection->NextSequenceRcv,
 559                         ntohl(Header->SequenceNumber),
 560                         Connection->NextSequenceRcv + TCP_WINDOW_SIZE
 561                         );
 562
 563                 // Is this packet the next expected packet?
 564                 if( sequence_num == Connection->NextSequenceRcv )
 565                 {
 566                          int    rv;
 567                         // Ooh, Goodie! Add it to the recieved list
 568                         rv = TCP_INT_AppendRecieved(Connection,
 569                                 (Uint8*)Header + (Header->DataOffset>>4)*4,
 570                                 dataLen
 571                                 );
 572                         if(rv != 0) {
 573                                 Log_Notice("TCP", "TCP_INT_AppendRecieved rv %i", rv);
 574                                 break;
 575                         }
 576                         LOG("0x%08x += %i", Connection->NextSequenceRcv, dataLen);
 577                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 578                                 Connection->HighestSequenceRcvd += dataLen;
 579                         Connection->NextSequenceRcv += dataLen;
 580
 581                         // TODO: This should be moved out of the watcher thread,
 582                         // so that a single lost packet on one connection doesn't cause
 583                         // all connections on the interface to lag.
 584                         // - Meh, no real issue, as the cache shouldn't be that large
 585                         TCP_INT_UpdateRecievedFromFuture(Connection);
 586
 587                         #if 1
 588                         // - Only send an ACK if we've had a burst
 589                         Uint32  bytes_since_last_ack = Connection->NextSequenceRcv - Connection->LastACKSequence;
 590                         LOG("bytes_since_last_ack = 0x%x", bytes_since_last_ack);
 591                         if( bytes_since_last_ack > TCP_DACK_THRESHOLD )
 592                         {
 593                                 TCP_INT_SendACK(Connection, "DACK Burst");
 594                                 // - Extend TCP deferred ACK timer
 595                                 Time_RemoveTimer(Connection->DeferredACKTimer);
 596                         }
 597                         // - Schedule the deferred ACK timer (if already scheduled, this is a NOP)
 598                         Time_ScheduleTimer(Connection->DeferredACKTimer, TCP_DACK_TIMEOUT);
 599                         #else
 600                         TCP_INT_SendACK(Connection, "RX");
 601                         #endif
 602                 }
 603                 // Check if the packet is in window
 604                 else if( sequence_num - Connection->NextSequenceRcv < TCP_WINDOW_SIZE )
 605                 {
 606                         Uint8   *dataptr = (Uint8*)Header + (Header->DataOffset>>4)*4;
 607                         Uint32  index = sequence_num % TCP_WINDOW_SIZE;
 608                         Uint32  max = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 609                         if( !(Connection->FuturePacketValidBytes[index/8] & (1 << (index%8))) )
 610                                 TCP_INT_SendACK(Connection, "Lost packet");
 611                         for( int i = 0; i < dataLen; i ++ )
 612                         {
 613                                 Connection->FuturePacketValidBytes[index/8] |= 1 << (index%8);
 614                                 Connection->FuturePacketData[index] = dataptr[i];
 615                                 // Do a wrap increment
 616                                 index ++;
 617                                 if(index == TCP_WINDOW_SIZE)    index = 0;
 618                                 if(index == max)        break;
 619                         }
 620                         Uint32  rel_highest = Connection->HighestSequenceRcvd - Connection->NextSequenceRcv;
 621                         Uint32  rel_this = index - Connection->NextSequenceRcv;
 622                         LOG("Updating highest this(0x%x) > highest(%x)", rel_this, rel_highest);
 623                         if( rel_this > rel_highest )
 624                         {
 625                                 Connection->HighestSequenceRcvd = index;
 626                         }
 627                 }
 628                 // Badly out of sequence packet
 629                 else
 630                 {
 631                         Log_Log("TCP", "Fully out of sequence packet (0x%08x not between 0x%08x and 0x%08x), dropped",
 632                                 sequence_num, Connection->NextSequenceRcv, Connection->NextSequenceRcv+TCP_WINDOW_SIZE);
 633                         // Spec says we should send an empty ACK with the current state
 634                         TCP_INT_SendACK(Connection, "Bad Seq");
 635                 }
 636                 break;
 637
 638         // --- Remote close states
 639         case TCP_ST_CLOSE_WAIT:
 640
 641                 // Ignore everything, CLOSE_WAIT is terminated by the client
 642                 Log_Debug("TCP", "CLOSE WAIT - Ignoring packets");
 643
 644                 break;
 645
 646         // LAST-ACK - Waiting for the ACK of FIN (from CLOSE WAIT)
 647         case TCP_ST_LAST_ACK:
 648                 if( Header->Flags & TCP_FLAG_ACK )
 649                 {
 650                         Connection->State = TCP_ST_FINISHED;    // Connection completed
 651                         Log_Log("TCP", "LAST-ACK to CLOSED - Connection remote closed");
 652                         TCP_int_FreeTCB(Connection);
 653                 }
 654                 break;
 655
 656         // --- Local close States
 657         case TCP_ST_FIN_WAIT1:
 658                 if( Header->Flags & TCP_FLAG_FIN )
 659                 {
 660                         Connection->State = TCP_ST_CLOSING;
 661                         Log_Debug("TCP", "Conn %p closed, sent FIN and recieved FIN", Connection);
 662                         VFS_MarkError(&Connection->Node, 1);
 663
 664                         TCP_INT_SendACK(Connection, "FINWAIT-1 FIN");
 665                         break ;
 666                 }
 667
 668                 // TODO: Make sure that the packet is actually ACKing the FIN
 669                 if( Header->Flags & TCP_FLAG_ACK )
 670                 {
 671                         Connection->State = TCP_ST_FIN_WAIT2;
 672                         Log_Debug("TCP", "Conn %p closed, sent FIN ACKed", Connection);
 673                         VFS_MarkError(&Connection->Node, 1);
 674                         return 0;
 675                 }
 676                 break;
 677
 678         case TCP_ST_FIN_WAIT2:
 679                 if( Header->Flags & TCP_FLAG_FIN )
 680                 {
 681                         Connection->State = TCP_ST_TIME_WAIT;
 682                         Log_Debug("TCP", "Conn %p FINWAIT-2 -> TIME WAIT", Connection);
 683                         TCP_INT_SendACK(Connection, "FINWAIT-2 FIN");
 684                 }
 685                 break;
 686
 687         case TCP_ST_CLOSING:
 688                 // TODO: Make sure that the packet is actually ACKing the FIN
 689                 if( Header->Flags & TCP_FLAG_ACK )
 690                 {
 691                         Connection->State = TCP_ST_TIME_WAIT;
 692                         Log_Debug("TCP", "Conn %p CLOSING -> TIME WAIT", Connection);
 693                         VFS_MarkError(&Connection->Node, 1);
 694                         return 0;
 695                 }
 696                 break;
 697
 698         // --- Closed (or near closed) states) ---
 699         case TCP_ST_TIME_WAIT:
 700                 Log_Log("TCP", "Packets on Time-Wait, ignored");
 701                 break;
 702
 703         case TCP_ST_FINISHED:
 704                 Log_Log("TCP", "Packets when CLOSED, ignoring");
 705                 break;
 706         case TCP_ST_FORCE_CLOSE:
 707                 Log_Log("TCP", "Packets when force CLOSED, ignoring");
 708                 return 1;
 709
 710         //default:
 711         //      Log_Warning("TCP", "Unhandled TCP state %i", Connection->State);
 712         //      break;
 713         }
 714
 715         return 0;
 716
 717 }
 718
 719 /**
 720  * \brief Appends a packet to the recieved list
 721  * \param Connection    Connection structure
 722  * \param Data  Packet contents
 723  * \param Length        Length of \a Data
 724  */
 725 int TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length)
 726 {
 727         Mutex_Acquire( &Connection->lRecievedPackets );
 728
 729         if(Connection->RecievedBuffer->Length + Length > Connection->RecievedBuffer->Space )
 730         {
 731                 VFS_MarkAvaliable(&Connection->Node, 1);
 732                 Log_Error("TCP", "Buffer filled, packet dropped (:%i) - %i + %i > %i",
 733                         Connection->LocalPort, Connection->RecievedBuffer->Length, Length,
 734                         Connection->RecievedBuffer->Space
 735                         );
 736                 Mutex_Release( &Connection->lRecievedPackets );
 737                 return 1;
 738         }
 739
 740         RingBuffer_Write( Connection->RecievedBuffer, Data, Length );
 741
 742         VFS_MarkAvaliable(&Connection->Node, 1);
 743
 744         Mutex_Release( &Connection->lRecievedPackets );
 745         return 0;
 746 }
 747
 748 /**
 749  * \brief Updates the connections recieved list from the future list
 750  * \param Connection    Connection structure
 751  *
 752  * Updates the recieved packets list with packets from the future (out
 753  * of order) packets list that are now able to be added in direct
 754  * sequence.
 755  */
 756 void TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection)
 757 {
 758         // Calculate length of contiguous bytes
 759         const size_t    length = Connection->HighestSequenceRcvd - Connection->NextSequenceRcv;
 760         Uint32  index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 761         size_t  runlength = length;
 762         LOG("HSR=0x%x,NSR=0x%x", Connection->HighestSequenceRcvd, Connection->NextSequenceRcv);
 763         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 764         {
 765                 return ;
 766         }
 767         LOG("length=%u, index=0x%x", length, index);
 768         for( int i = 0; i < length; i ++ )
 769         {
 770                  int    bit = index % 8;
 771                 Uint8   bitfield_byte = Connection->FuturePacketValidBytes[index / 8];
 772                 if( (bitfield_byte & (1 << bit)) == 0 ) {
 773                         runlength = i;
 774                         LOG("Hit missing, break");
 775                         break;
 776                 }
 777
 778                 if( bitfield_byte == 0xFF ) {
 779                          int    inc = 8 - bit;
 780                         i += inc - 1;
 781                         index += inc;
 782                 }
 783                 else {
 784                         index ++;
 785                 }
 786                 if(index > TCP_WINDOW_SIZE)
 787                         index -= TCP_WINDOW_SIZE;
 788         }
 789
 790         index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 791         Connection->NextSequenceRcv += runlength;
 792
 793         // Write data to to the ring buffer
 794         if( TCP_WINDOW_SIZE - index > runlength )
 795         {
 796                 // Simple case
 797                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, runlength );
 798         }
 799         else
 800         {
 801                  int    endLen = TCP_WINDOW_SIZE - index;
 802                 // 2-part case
 803                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, endLen );
 804                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData, endLen - runlength );
 805         }
 806
 807         // Mark (now saved) bytes as invalid
 808         // - Align index
 809         while(index % 8 && runlength > 0)
 810         {
 811                 Connection->FuturePacketData[index] = 0;
 812                 Connection->FuturePacketValidBytes[index/8] &= ~(1 << (index%8));
 813                 index ++;
 814                 if(index > TCP_WINDOW_SIZE)
 815                         index -= TCP_WINDOW_SIZE;
 816                 runlength --;
 817         }
 818         while( runlength > 7 )
 819         {
 820                 Connection->FuturePacketData[index] = 0;
 821                 Connection->FuturePacketValidBytes[index/8] = 0;
 822                 runlength -= 8;
 823                 index += 8;
 824                 if(index > TCP_WINDOW_SIZE)
 825                         index -= TCP_WINDOW_SIZE;
 826         }
 827         while( runlength > 0)
 828         {
 829                 Connection->FuturePacketData[index] = 0;
 830                 Connection->FuturePacketData[index/8] &= ~(1 << (index%8));
 831                 index ++;
 832                 if(index > TCP_WINDOW_SIZE)
 833                         index -= TCP_WINDOW_SIZE;
 834                 runlength --;
 835         }
 836 }
 837
 838 void TCP_int_SendDelayedACK(void *ConnPtr)
 839 {
 840         TCP_INT_SendACK(ConnPtr, "DACK Timeout");
 841 }
 842
 843 void TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason)
 844 {
 845         tTCPHeader      hdr;
 846         // ACK Packet
 847         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
 848         hdr.DestPort = htons(Connection->RemotePort);
 849         hdr.SourcePort = htons(Connection->LocalPort);
 850         hdr.AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 851         hdr.SequenceNumber = htonl(Connection->NextSequenceSend);
 852         hdr.WindowSize = htons(TCP_WINDOW_SIZE);
 853         hdr.Flags = TCP_FLAG_ACK;       // TODO: Determine if SYN is wanted too
 854         hdr.Checksum = 0;       // TODO: Checksum
 855         hdr.UrgentPointer = 0;
 856         Log_Debug("TCP", "Sending ACK for 0x%08x (%s)", Connection->NextSequenceRcv, Reason);
 857         TCP_SendPacket( Connection, &hdr, 0, NULL );
 858         //Connection->NextSequenceSend ++;
 859         Connection->LastACKSequence = Connection->NextSequenceRcv;
 860 }
 861
 862 /**
 863  * \fn Uint16 TCP_GetUnusedPort()
 864  * \brief Gets an unused port and allocates it
 865  */
 866 Uint16 TCP_GetUnusedPort()
 867 {
 868         Uint16  ret;
 869
 870         // Get Next outbound port
 871         ret = giTCP_NextOutPort++;
 872         while( gaTCP_PortBitmap[ret/32] & (1UL << (ret%32)) )
 873         {
 874                 ret ++;
 875                 giTCP_NextOutPort++;
 876                 if(giTCP_NextOutPort == 0x10000) {
 877                         ret = giTCP_NextOutPort = TCP_MIN_DYNPORT;
 878                 }
 879         }
 880
 881         // Mark the new port as used
 882         gaTCP_PortBitmap[ret/32] |= 1 << (ret%32);
 883
 884         return ret;
 885 }
 886
 887 /**
 888  * \fn int TCP_AllocatePort(Uint16 Port)
 889  * \brief Marks a port as used
 890  */
 891 int TCP_AllocatePort(Uint16 Port)
 892 {
 893         // Check if the port has already been allocated
 894         if( gaTCP_PortBitmap[Port/32] & (1 << (Port%32)) )
 895                 return 0;
 896
 897         // Allocate
 898         gaTCP_PortBitmap[Port/32] |= 1 << (Port%32);
 899
 900         return 1;
 901 }
 902
 903 /**
 904  * \fn int TCP_DeallocatePort(Uint16 Port)
 905  * \brief Marks a port as unused
 906  */
 907 int TCP_DeallocatePort(Uint16 Port)
 908 {
 909         // Check if the port has already been allocated
 910         if( !(gaTCP_PortBitmap[Port/32] & (1 << (Port%32))) )
 911                 return 0;
 912
 913         // Allocate
 914         gaTCP_PortBitmap[Port/32] &= ~(1 << (Port%32));
 915
 916         return 1;
 917 }
 918
 919 tTCPConnection *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State)
 920 {
 921         tTCPConnection  *conn = calloc( sizeof(tTCPConnection) + TCP_WINDOW_SIZE + TCP_WINDOW_SIZE/8, 1 );
 922
 923         conn->State = State;
 924         conn->Interface = Interface;
 925         conn->LocalPort = -1;
 926         conn->RemotePort = -1;
 927
 928         conn->Node.Size = -1;
 929         conn->Node.ReferenceCount = 1;
 930         conn->Node.ImplPtr = conn;
 931         conn->Node.NumACLs = 1;
 932         conn->Node.ACLs = &gVFS_ACL_EveryoneRW;
 933         conn->Node.Type = &gTCP_ClientNodeType;
 934         conn->Node.BufferFull = 1;      // Cleared when connection opens
 935
 936         conn->RecievedBuffer = RingBuffer_Create( TCP_RECIEVE_BUFFER_SIZE );
 937         #if 0
 938         conn->SentBuffer = RingBuffer_Create( TCP_SEND_BUFFER_SIZE );
 939         Semaphore_Init(conn->SentBufferSpace, 0, TCP_SEND_BUFFER_SIZE, "TCP SentBuffer", conn->Name);
 940         #endif
 941
 942         conn->HighestSequenceRcvd = 0;
 943         #if CACHE_FUTURE_PACKETS_IN_BYTES
 944         // Future recieved data (ahead of the expected sequence number)
 945         conn->FuturePacketData = (Uint8*)conn + sizeof(tTCPConnection);
 946         conn->FuturePacketValidBytes = conn->FuturePacketData + TCP_WINDOW_SIZE;
 947         #endif
 948
 949         conn->DeferredACKTimer = Time_AllocateTimer( TCP_int_SendDelayedACK, conn);
 950         return conn;
 951 }
 952
 953 void TCP_int_FreeTCB(tTCPConnection *Connection)
 954 {
 955         ASSERTC(Connection->State, ==, TCP_ST_FINISHED);
 956         ASSERTC(Connection->Node.ReferenceCount, ==, 0);
 957
 958         if( Connection->Server )
 959         {
 960                 tTCPListener    *srv = Connection->Server;
 961                 SHORTLOCK(&srv->lConnections);
 962                 if(Connection->Prev)
 963                         Connection->Prev->Next = Connection->Next;
 964                 else
 965                         srv->Connections = Connection->Next;
 966                 if(Connection->Next)
 967                         Connection->Next->Prev = Connection->Prev;
 968                 else {
 969                         ASSERT(srv->ConnectionsTail == Connection);
 970                         srv->ConnectionsTail = Connection->Prev;
 971                 }
 972                 SHORTREL(&srv->lConnections);
 973         }
 974         else
 975         {
 976                 SHORTLOCK(&glTCP_OutbountCons);
 977                 if(Connection->Prev)
 978                         Connection->Prev->Next = Connection->Next;
 979                 else
 980                         gTCP_OutbountCons = Connection->Next;
 981                 if(Connection->Next)
 982                         Connection->Next->Prev = Connection->Prev;
 983                 else
 984                         ;
 985                 SHORTREL(&glTCP_OutbountCons);
 986         }
 987
 988         RingBuffer_Free(Connection->RecievedBuffer);
 989         Time_FreeTimer(Connection->DeferredACKTimer);
 990         // TODO: Force VFS to close handles? (they should all be closed);
 991         free(Connection);
 992 }
 993
 994 // --- Server
 995 tVFS_Node *TCP_Server_Init(tInterface *Interface)
 996 {
 997         tTCPListener    *srv;
 998
 999         srv = calloc( 1, sizeof(tTCPListener) );
1000
1001         if( srv == NULL ) {
1002                 Log_Warning("TCP", "malloc failed for listener (%i) bytes", sizeof(tTCPListener));
1003                 return NULL;
1004         }
1005
1006         srv->Interface = Interface;
1007         srv->Port = 0;
1008         srv->NextID = 0;
1009         srv->Connections = NULL;
1010         srv->ConnectionsTail = NULL;
1011         srv->NewConnections = NULL;
1012         srv->Next = NULL;
1013         srv->Node.Flags = VFS_FFLAG_DIRECTORY;
1014         srv->Node.Size = -1;
1015         srv->Node.ImplPtr = srv;
1016         srv->Node.NumACLs = 1;
1017         srv->Node.ACLs = &gVFS_ACL_EveryoneRW;
1018         srv->Node.Type = &gTCP_ServerNodeType;
1019
1020         SHORTLOCK(&glTCP_Listeners);
1021         srv->Next = gTCP_Listeners;
1022         gTCP_Listeners = srv;
1023         SHORTREL(&glTCP_Listeners);
1024
1025         return &srv->Node;
1026 }
1027
1028 /**
1029  * \brief Wait for a new connection and return the connection ID
1030  * \note Blocks until a new connection is made
1031  * \param Node  Server node
1032  * \param Pos   Position (ignored)
1033  */
1034 int TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Dest[FILENAME_MAX])
1035 {
1036         tTCPListener    *srv = Node->ImplPtr;
1037         tTCPConnection  *conn;
1038
1039         ENTER("pNode iPos", Node, Pos);
1040
1041         Log_Log("TCP", "Thread %i waiting for a connection", Threads_GetTID());
1042         Semaphore_Wait( &srv->WaitingConnections, 1 );
1043
1044         SHORTLOCK(&srv->lConnections);
1045         // Increment the new list (the current connection is still on the
1046         // normal list)
1047         conn = srv->NewConnections;
1048         srv->NewConnections = conn->Next;
1049
1050         if( srv->NewConnections == NULL )
1051                 VFS_MarkAvaliable( Node, 0 );
1052
1053         SHORTREL( &srv->lConnections );
1054
1055         LOG("conn = %p", conn);
1056         LOG("srv->Connections = %p", srv->Connections);
1057         LOG("srv->NewConnections = %p", srv->NewConnections);
1058         LOG("srv->ConnectionsTail = %p", srv->ConnectionsTail);
1059
1060         itoa(Dest, conn->Node.ImplInt, 16, 8, '0');
1061         Log_Log("TCP", "Thread %i got connection '%s'", Threads_GetTID(), Dest);
1062         LEAVE('i', 0);
1063         return 0;
1064 }
1065
1066 /**
1067  * \brief Gets a client connection node
1068  * \param Node  Server node
1069  * \param Name  Hexadecimal ID of the node
1070  */
1071 tVFS_Node *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags)
1072 {
1073         tTCPConnection  *conn;
1074         tTCPListener    *srv = Node->ImplPtr;
1075         char    tmp[9];
1076          int    id = atoi(Name);
1077
1078         ENTER("pNode sName", Node, Name);
1079
1080         // Check for a non-empty name
1081         if( Name[0] )
1082         {
1083                 // Sanity Check
1084                 itoa(tmp, id, 16, 8, '0');
1085                 if(strcmp(tmp, Name) != 0) {
1086                         LOG("'%s' != '%s' (%08x)", Name, tmp, id);
1087                         LEAVE('n');
1088                         return NULL;
1089                 }
1090
1091                 Log_Debug("TCP", "srv->Connections = %p", srv->Connections);
1092                 Log_Debug("TCP", "srv->NewConnections = %p", srv->NewConnections);
1093                 Log_Debug("TCP", "srv->ConnectionsTail = %p", srv->ConnectionsTail);
1094
1095                 // Search
1096                 SHORTLOCK( &srv->lConnections );
1097                 for(conn = srv->Connections;
1098                         conn;
1099                         conn = conn->Next)
1100                 {
1101                         LOG("conn->Node.ImplInt = %i", conn->Node.ImplInt);
1102                         if(conn->Node.ImplInt == id)    break;
1103                 }
1104                 SHORTREL( &srv->lConnections );
1105
1106                 // If not found, ret NULL
1107                 if(!conn) {
1108                         LOG("Connection %i not found", id);
1109                         LEAVE('n');
1110                         return NULL;
1111                 }
1112         }
1113         // Empty Name - Check for a new connection and if it's there, open it
1114         else
1115         {
1116                 SHORTLOCK( &srv->lConnections );
1117                 conn = srv->NewConnections;
1118                 if( conn != NULL )
1119                         srv->NewConnections = conn->Next;
1120                 VFS_MarkAvaliable( Node, srv->NewConnections != NULL );
1121                 SHORTREL( &srv->lConnections );
1122                 if( !conn ) {
1123                         LOG("No new connections");
1124                         LEAVE('n');
1125                         return NULL;
1126                 }
1127         }
1128
1129         // Return node
1130         LEAVE('p', &conn->Node);
1131         return &conn->Node;
1132 }
1133
1134 /**
1135  * \brief Handle IOCtl calls
1136  */
1137 int TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data)
1138 {
1139         tTCPListener    *srv = Node->ImplPtr;
1140
1141         switch(ID)
1142         {
1143         case 4: // Get/Set Port
1144                 if(!Data)       // Get Port
1145                         return srv->Port;
1146
1147                 if(srv->Port)   // Wait, you can't CHANGE the port
1148                         return -1;
1149
1150                 if(!CheckMem(Data, sizeof(Uint16)))     // Sanity check
1151                         return -1;
1152
1153                 // Permissions check
1154                 if(Threads_GetUID() != 0
1155                 && *(Uint16*)Data != 0
1156                 && *(Uint16*)Data < 1024)
1157                         return -1;
1158
1159                 // TODO: Check if a port is in use
1160
1161                 // Set Port
1162                 srv->Port = *(Uint16*)Data;
1163                 if(srv->Port == 0)      // Allocate a random port
1164                         srv->Port = TCP_GetUnusedPort();
1165                 else    // Else, mark this as used
1166                         TCP_AllocatePort(srv->Port);
1167
1168                 Log_Log("TCP", "Server %p listening on port %i", srv, srv->Port);
1169
1170                 return srv->Port;
1171         }
1172         return 0;
1173 }
1174
1175 void TCP_Server_Close(tVFS_Node *Node)
1176 {
1177         free(Node->ImplPtr);
1178 }
1179
1180 // --- Client
1181 /**
1182  * \brief Create a client node
1183  */
1184 tVFS_Node *TCP_Client_Init(tInterface *Interface)
1185 {
1186         tTCPConnection  *conn = TCP_int_CreateConnection(Interface, TCP_ST_CLOSED);
1187
1188         SHORTLOCK(&glTCP_OutbountCons);
1189         conn->Server = NULL;
1190         conn->Prev = NULL;
1191         conn->Next = gTCP_OutbountCons;
1192         if(gTCP_OutbountCons)
1193                 gTCP_OutbountCons->Prev = conn;
1194         gTCP_OutbountCons = conn;
1195         SHORTREL(&glTCP_OutbountCons);
1196
1197         return &conn->Node;
1198 }
1199
1200 /**
1201  * \brief Wait for a packet and return it
1202  * \note If \a Length is smaller than the size of the packet, the rest
1203  *       of the packet's data will be discarded.
1204  */
1205 size_t TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags)
1206 {
1207         tTCPConnection  *conn = Node->ImplPtr;
1208         size_t  len;
1209
1210         ENTER("pNode XOffset xLength pBuffer", Node, Offset, Length, Buffer);
1211         LOG("conn = %p {State:%i}", conn, conn->State);
1212
1213         // If the connection has been closed (state > ST_OPEN) then clear
1214         // any stale data in the buffer (until it is empty (until it is empty))
1215         if( conn->State > TCP_ST_ESTABLISHED )
1216         {
1217                 LOG("Connection closed");
1218                 Mutex_Acquire( &conn->lRecievedPackets );
1219                 len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1220                 Mutex_Release( &conn->lRecievedPackets );
1221
1222                 if( len == 0 ) {
1223                         VFS_MarkAvaliable(Node, 0);
1224                         errno = 0;
1225                         LEAVE('i', -1);
1226                         return -1;
1227                 }
1228
1229                 LEAVE('i', len);
1230                 return len;
1231         }
1232
1233         // Wait
1234         {
1235                 tTime   *timeout = NULL;
1236                 tTime   timeout_zero = 0;
1237                 if( Flags & VFS_IOFLAG_NOBLOCK )
1238                         timeout = &timeout_zero;
1239                 if( !VFS_SelectNode(Node, VFS_SELECT_READ|VFS_SELECT_ERROR, timeout, "TCP_Client_Read") ) {
1240                         errno = EWOULDBLOCK;
1241                         LEAVE('i', -1);
1242                         return -1;
1243                 }
1244         }
1245
1246         // Lock list and read as much as possible (up to `Length`)
1247         Mutex_Acquire( &conn->lRecievedPackets );
1248         len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1249
1250         if( len == 0 || conn->RecievedBuffer->Length == 0 ) {
1251                 LOG("Marking as none avaliable (len = %i)", len);
1252                 VFS_MarkAvaliable(Node, 0);
1253         }
1254
1255         // Release the lock (we don't need it any more)
1256         Mutex_Release( &conn->lRecievedPackets );
1257
1258         LEAVE('i', len);
1259         return len;
1260 }
1261
1262 /**
1263  * \brief Send a data packet on a connection
1264  */
1265 void TCP_INT_SendDataPacket(tTCPConnection *Connection, size_t Length, const void *Data)
1266 {
1267         char    buf[sizeof(tTCPHeader)+Length];
1268         tTCPHeader      *packet = (void*)buf;
1269
1270         // - Stop Delayed ACK timer (as this data packet ACKs)
1271         Time_RemoveTimer(Connection->DeferredACKTimer);
1272
1273         // TODO: Don't exceed window size
1274
1275         packet->SourcePort = htons(Connection->LocalPort);
1276         packet->DestPort = htons(Connection->RemotePort);
1277         packet->DataOffset = (sizeof(tTCPHeader)/4)*16;
1278         packet->WindowSize = htons(TCP_WINDOW_SIZE);
1279
1280         packet->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
1281         packet->SequenceNumber = htonl(Connection->NextSequenceSend);
1282         packet->Flags = TCP_FLAG_PSH|TCP_FLAG_ACK;      // Hey, ACK if you can!
1283         packet->UrgentPointer = 0;
1284
1285         memcpy(packet->Options, Data, Length);
1286
1287         Log_Debug("TCP", "Send sequence 0x%08x", Connection->NextSequenceSend);
1288 #if HEXDUMP_OUTGOING
1289         Debug_HexDump("TCP_INT_SendDataPacket: Data = ", Data, Length);
1290 #endif
1291
1292         TCP_SendPacket( Connection, packet, Length, Data );
1293
1294         // TODO: Start a retransmit time (if data is not ACKed in x seconds, send again)
1295
1296         Connection->NextSequenceSend += Length;
1297 }
1298
1299 /**
1300  * \brief Send some bytes on a connection
1301  */
1302 size_t TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags)
1303 {
1304         tTCPConnection  *conn = Node->ImplPtr;
1305         size_t  rem = Length;
1306
1307         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1308
1309 //      #if DEBUG
1310 //      Debug_HexDump("TCP_Client_Write: Buffer = ",
1311 //              Buffer, Length);
1312 //      #endif
1313
1314         // Don't allow a write to a closed connection
1315         if( conn->State > TCP_ST_ESTABLISHED ) {
1316                 VFS_MarkError(Node, 1);
1317                 errno = 0;
1318                 LEAVE('i', -1);
1319                 return -1;
1320         }
1321
1322         // Wait
1323         {
1324                 tTime   *timeout = NULL;
1325                 tTime   timeout_zero = 0;
1326                 if( Flags & VFS_IOFLAG_NOBLOCK )
1327                         timeout = &timeout_zero;
1328                 if( !VFS_SelectNode(Node, VFS_SELECT_WRITE|VFS_SELECT_ERROR, timeout, "TCP_Client_Write") ) {
1329                         errno = EWOULDBLOCK;
1330                         LEAVE('i', -1);
1331                         return -1;
1332                 }
1333         }
1334
1335         do
1336         {
1337                  int    len = (rem < TCP_MAX_PACKET_SIZE) ? rem : TCP_MAX_PACKET_SIZE;
1338
1339                 #if 0
1340                 // Wait for space in the buffer
1341                 Semaphore_Signal( &Connection->SentBufferSpace, len );
1342
1343                 // Save data to buffer (and update the length read by the ammount written)
1344                 len = RingBuffer_Write( &Connection->SentBuffer, Buffer, len);
1345                 #endif
1346
1347                 // Send packet
1348                 TCP_INT_SendDataPacket(conn, len, Buffer);
1349
1350                 Buffer += len;
1351                 rem -= len;
1352         } while( rem > 0 );
1353
1354         LEAVE('i', Length);
1355         return Length;
1356 }
1357
1358 /**
1359  * \brief Open a connection to another host using TCP
1360  * \param Conn  Connection structure
1361  */
1362 void TCP_StartConnection(tTCPConnection *Conn)
1363 {
1364         tTCPHeader      hdr = {0};
1365
1366         Conn->State = TCP_ST_SYN_SENT;
1367
1368         hdr.SourcePort = htons(Conn->LocalPort);
1369         hdr.DestPort = htons(Conn->RemotePort);
1370         Conn->NextSequenceSend = rand();
1371         hdr.SequenceNumber = htonl(Conn->NextSequenceSend);
1372         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
1373         hdr.Flags = TCP_FLAG_SYN;
1374         hdr.WindowSize = htons(TCP_WINDOW_SIZE);        // Max
1375         hdr.Checksum = 0;       // TODO
1376
1377         TCP_SendPacket( Conn, &hdr, 0, NULL );
1378
1379         Conn->NextSequenceSend ++;
1380         Conn->State = TCP_ST_SYN_SENT;
1381
1382         return ;
1383 }
1384
1385 /**
1386  * \brief Control a client socket
1387  */
1388 int TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data)
1389 {
1390         tTCPConnection  *conn = Node->ImplPtr;
1391
1392         ENTER("pNode iID pData", Node, ID, Data);
1393
1394         switch(ID)
1395         {
1396         case 4: // Get/Set local port
1397                 if(!Data)
1398                         LEAVE_RET('i', conn->LocalPort);
1399                 if(conn->State != TCP_ST_CLOSED)
1400                         LEAVE_RET('i', -1);
1401                 if(!CheckMem(Data, sizeof(Uint16)))
1402                         LEAVE_RET('i', -1);
1403
1404                 if(Threads_GetUID() != 0 && *(Uint16*)Data < 1024)
1405                         LEAVE_RET('i', -1);
1406
1407                 conn->LocalPort = *(Uint16*)Data;
1408                 LEAVE_RET('i', conn->LocalPort);
1409
1410         case 5: // Get/Set remote port
1411                 if(!Data)       LEAVE_RET('i', conn->RemotePort);
1412                 if(conn->State != TCP_ST_CLOSED)        LEAVE_RET('i', -1);
1413                 if(!CheckMem(Data, sizeof(Uint16)))     LEAVE_RET('i', -1);
1414                 conn->RemotePort = *(Uint16*)Data;
1415                 LEAVE_RET('i', conn->RemotePort);
1416
1417         case 6: // Set Remote IP
1418                 if( conn->State != TCP_ST_CLOSED )
1419                         LEAVE_RET('i', -1);
1420                 if( conn->Interface->Type == 4 )
1421                 {
1422                         if(!CheckMem(Data, sizeof(tIPv4)))      LEAVE_RET('i', -1);
1423                         conn->RemoteIP.v4 = *(tIPv4*)Data;
1424                 }
1425                 else if( conn->Interface->Type == 6 )
1426                 {
1427                         if(!CheckMem(Data, sizeof(tIPv6)))      LEAVE_RET('i', -1);
1428                         conn->RemoteIP.v6 = *(tIPv6*)Data;
1429                 }
1430                 LEAVE_RET('i', 0);
1431
1432         case 7: // Connect
1433                 if(conn->LocalPort == 0xFFFF)
1434                         conn->LocalPort = TCP_GetUnusedPort();
1435                 if(conn->RemotePort == -1)
1436                         LEAVE_RET('i', 0);
1437
1438                 {
1439                         tTime   timeout = conn->Interface->TimeoutDelay;
1440
1441                         TCP_StartConnection(conn);
1442                         VFS_SelectNode(&conn->Node, VFS_SELECT_WRITE, &timeout, "TCP Connection");
1443                         if( conn->State == TCP_ST_SYN_SENT )
1444                                 LEAVE_RET('i', 0);
1445                 }
1446
1447                 LEAVE_RET('i', 1);
1448
1449         // Get recieve buffer length
1450         case 8:
1451                 LEAVE_RET('i', conn->RecievedBuffer->Length);
1452         }
1453
1454         return 0;
1455 }
1456
1457 void TCP_Client_Close(tVFS_Node *Node)
1458 {
1459         tTCPConnection  *conn = Node->ImplPtr;
1460         tTCPHeader      packet;
1461
1462         ENTER("pNode", Node);
1463
1464         ASSERT(Node->ReferenceCount != 0);
1465
1466         if( Node->ReferenceCount > 1 ) {
1467                 Node->ReferenceCount --;
1468                 LOG("Dereference only");
1469                 LEAVE('-');
1470                 return ;
1471         }
1472         Node->ReferenceCount --;
1473
1474         if( conn->State == TCP_ST_CLOSE_WAIT || conn->State == TCP_ST_ESTABLISHED )
1475         {
1476                 packet.SourcePort = htons(conn->LocalPort);
1477                 packet.DestPort = htons(conn->RemotePort);
1478                 packet.DataOffset = (sizeof(tTCPHeader)/4)*16;
1479                 packet.WindowSize = TCP_WINDOW_SIZE;
1480
1481                 packet.AcknowlegementNumber = 0;
1482                 packet.SequenceNumber = htonl(conn->NextSequenceSend);
1483                 packet.Flags = TCP_FLAG_FIN;
1484
1485                 TCP_SendPacket( conn, &packet, 0, NULL );
1486         }
1487
1488         Time_RemoveTimer(conn->DeferredACKTimer);
1489
1490         switch( conn->State )
1491         {
1492         case TCP_ST_CLOSED:
1493                 Log_Warning("TCP", "Closing connection that was never opened");
1494                 TCP_int_FreeTCB(conn);
1495                 break;
1496         case TCP_ST_FORCE_CLOSE:
1497                 conn->State = TCP_ST_FINISHED;
1498                 TCP_int_FreeTCB(conn);
1499                 break;
1500         case TCP_ST_CLOSE_WAIT:
1501                 conn->State = TCP_ST_LAST_ACK;
1502                 break;
1503         case TCP_ST_ESTABLISHED:
1504                 conn->State = TCP_ST_FIN_WAIT1;
1505                 while( conn->State == TCP_ST_FIN_WAIT1 )
1506                         Threads_Yield();
1507                 // No free, freed after TIME_WAIT
1508                 break;
1509         default:
1510                 Log_Warning("TCP", "Unhandled connection state %i in TCP_Client_Close",
1511                         conn->State);
1512                 break;
1513         }
1514
1515         LEAVE('-');
1516 }
1517
1518 /**
1519  * \brief Checks if a value is between two others (after taking into account wrapping)
1520  */
1521 int WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue)
1522 {
1523         if( MaxValue < 0xFFFFFFFF )
1524         {
1525                 Lower %= MaxValue + 1;
1526                 Value %= MaxValue + 1;
1527                 Higher %= MaxValue + 1;
1528         }
1529
1530         // Simple Case, no wrap ?
1531         //       Lower Value Higher
1532         // | ... + ... + ... + ... |
1533
1534         if( Lower < Higher ) {
1535                 return Lower < Value && Value < Higher;
1536         }
1537         // Higher has wrapped below lower
1538
1539         // Value > Lower ?
1540         //       Higher Lower Value
1541         // | ... +  ... + ... + ... |
1542         if( Value > Lower ) {
1543                 return 1;
1544         }
1545
1546         // Value < Higher ?
1547         //       Value Higher Lower
1548         // | ... + ... +  ... + ... |
1549         if( Value < Higher ) {
1550                 return 1;
1551         }
1552
1553         return 0;
1554 }
1555 Uint32 GetRelative(Uint32 Base, Uint32 Value)
1556 {
1557         if( Value < Base )
1558                 return Value - Base + 0xFFFFFFFF;
1559         else
1560                 return Value - Base;
1561 }