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