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