KernelLand/Modules/IPStack/tcp.c

   1 /*
   2  * Acess2 IP Stack
   3  * - TCP Handling
   4  */
   5 #define DEBUG   0
   6 #include "ipstack.h"
   7 #include "ipv4.h"
   8 #include "ipv6.h"
   9 #include "tcp.h"
  10
  11 #define HEXDUMP_INCOMING        0
  12 #define HEXDUMP_OUTGOING        0
  13
  14 #define TCP_MIN_DYNPORT 0xC000
  15 #define TCP_MAX_HALFOPEN        1024    // Should be enough
  16
  17 #define TCP_MAX_PACKET_SIZE     1024
  18 #define TCP_WINDOW_SIZE 0x2000
  19 #define TCP_RECIEVE_BUFFER_SIZE 0x8000
  20 #define TCP_DACK_THRESHOLD      4096
  21 #define TCP_DACK_TIMEOUT        500
  22
  23 #define TCP_DEBUG       0       // Set to non-0 to enable TCP packet logging
  24
  25 // === PROTOTYPES ===
  26 void    TCP_Initialise(void);
  27 void    TCP_StartConnection(tTCPConnection *Conn);
  28 void    TCP_SendPacket(tTCPConnection *Conn, tTCPHeader *Header, size_t DataLen, const void *Data);
  29 void    TCP_int_SendPacket(tInterface *Interface, const void *Dest, tTCPHeader *Header, size_t Length, const void *Data);
  30 void    TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer);
  31 void    TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length);
  32 int     TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length);
  33 void    TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection);
  34 void    TCP_int_SendDelayedACK(void *ConnPtr);
  35 void    TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason);
  36 Uint16  TCP_GetUnusedPort();
  37  int    TCP_AllocatePort(Uint16 Port);
  38  int    TCP_DeallocatePort(Uint16 Port);
  39 tTCPConnection  *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State);
  40 // --- Server
  41 tVFS_Node       *TCP_Server_Init(tInterface *Interface);
  42  int    TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Name[FILENAME_MAX]);
  43 tVFS_Node       *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags);
  44  int    TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data);
  45 void    TCP_Server_Close(tVFS_Node *Node);
  46 // --- Client
  47 tVFS_Node       *TCP_Client_Init(tInterface *Interface);
  48 size_t  TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags);
  49 size_t  TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags);
  50  int    TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data);
  51 void    TCP_Client_Close(tVFS_Node *Node);
  52 // --- Helpers
  53  int    WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue);
  54
  55 // === TEMPLATES ===
  56 tSocketFile     gTCP_ServerFile = {NULL, "tcps", TCP_Server_Init};
  57 tSocketFile     gTCP_ClientFile = {NULL, "tcpc", TCP_Client_Init};
  58 tVFS_NodeType   gTCP_ServerNodeType = {
  59         .TypeName = "TCP Server",
  60         .ReadDir = TCP_Server_ReadDir,
  61         .FindDir = TCP_Server_FindDir,
  62         .IOCtl   = TCP_Server_IOCtl,
  63         .Close   = TCP_Server_Close
  64         };
  65 tVFS_NodeType   gTCP_ClientNodeType = {
  66         .TypeName = "TCP Client/Connection",
  67         .Read  = TCP_Client_Read,
  68         .Write = TCP_Client_Write,
  69         .IOCtl = TCP_Client_IOCtl,
  70         .Close = TCP_Client_Close
  71         };
  72
  73 // === GLOBALS ===
  74  int    giTCP_NumHalfopen = 0;
  75 tShortSpinlock  glTCP_Listeners;
  76 tTCPListener    *gTCP_Listeners;
  77 tShortSpinlock  glTCP_OutbountCons;
  78 tTCPConnection  *gTCP_OutbountCons;
  79 Uint32  gaTCP_PortBitmap[0x800];
  80  int    giTCP_NextOutPort = TCP_MIN_DYNPORT;
  81
  82 // === CODE ===
  83 /**
  84  * \brief Initialise the TCP Layer
  85  *
  86  * Registers the client and server files and the GetPacket callback
  87  */
  88 void TCP_Initialise(void)
  89 {
  90         giTCP_NextOutPort += rand()%128;
  91         IPStack_AddFile(&gTCP_ServerFile);
  92         IPStack_AddFile(&gTCP_ClientFile);
  93         IPv4_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  94         IPv6_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  95 }
  96
  97 /**
  98  * \brief Sends a packet from the specified connection, calculating the checksums
  99  * \param Conn  Connection
 100  * \param Length        Length of data
 101  * \param Data  Packet data (cast as a TCP Header)
 102  */
 103 void TCP_SendPacket( tTCPConnection *Conn, tTCPHeader *Header, size_t Length, const void *Data )
 104 {
 105         TCP_int_SendPacket(Conn->Interface, &Conn->RemoteIP, Header, Length, Data);
 106 }
 107
 108 void TCP_int_SendPacket(tInterface *Interface, const void *Dest, tTCPHeader *Header, size_t Length, const void *Data )
 109 {
 110         tIPStackBuffer  *buffer;
 111         Uint16  checksum[3];
 112          int    packlen = sizeof(*Header) + Length;
 113
 114         buffer = IPStack_Buffer_CreateBuffer(2 + IPV4_BUFFERS);
 115         if( Data && Length )
 116                 IPStack_Buffer_AppendSubBuffer(buffer, Length, 0, Data, NULL, NULL);
 117         IPStack_Buffer_AppendSubBuffer(buffer, sizeof(*Header), 0, Header, NULL, NULL);
 118
 119         LOG("Sending %i+%i to %s:%i", sizeof(*Header), Length,
 120                 IPStack_PrintAddress(Interface->Type, Dest),
 121                 ntohs(Header->RemotePort)
 122                 );
 123
 124         Header->Checksum = 0;
 125         checksum[1] = htons( ~IPv4_Checksum(Header, sizeof(tTCPHeader)) );
 126         checksum[2] = htons( ~IPv4_Checksum(Data, Length) );
 127
 128         // TODO: Fragment packet
 129
 130         switch( Interface->Type )
 131         {
 132         case 4:
 133                 // Get IPv4 pseudo-header checksum
 134                 {
 135                         Uint32  buf[3];
 136                         buf[0] = ((tIPv4*)Interface->Address)->L;
 137                         buf[1] = ((tIPv4*)Dest)->L;
 138                         buf[2] = (htons(packlen)<<16) | (6<<8) | 0;
 139                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 140                 }
 141                 // - Combine checksums
 142                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );
 143                 IPv4_SendPacket(Interface, *(tIPv4*)Dest, IP4PROT_TCP, 0, buffer);
 144                 break;
 145
 146         case 6:
 147                 // Append IPv6 Pseudo Header
 148                 {
 149                         Uint32  buf[4+4+1+1];
 150                         memcpy(buf, Interface->Address, 16);
 151                         memcpy(&buf[4], Dest, 16);
 152                         buf[8] = htonl(packlen);
 153                         buf[9] = htonl(6);
 154                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 155                 }
 156                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );  // Combine the two
 157                 IPv6_SendPacket(Interface, *(tIPv6*)Dest, IP4PROT_TCP, buffer);
 158                 break;
 159         }
 160 }
 161
 162 /**
 163  * \brief Handles a packet from the IP Layer
 164  * \param Interface     Interface the packet arrived from
 165  * \param Address       Pointer to the addres structure
 166  * \param Length        Size of packet in bytes
 167  * \param Buffer        Packet data
 168  */
 169 void TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer)
 170 {
 171         tTCPHeader      *hdr = Buffer;
 172
 173         #if TCP_DEBUG
 174         Log_Log("TCP", "TCP_GetPacket: <Local>:%i from [%s]:%i, Flags = %s%s%s%s%s%s%s%s",
 175                 ntohs(hdr->DestPort),
 176                 IPStack_PrintAddress(Interface->Type, Address),
 177                 ntohs(hdr->SourcePort),
 178                 (hdr->Flags & TCP_FLAG_CWR) ? "CWR " : "",
 179                 (hdr->Flags & TCP_FLAG_ECE) ? "ECE " : "",
 180                 (hdr->Flags & TCP_FLAG_URG) ? "URG " : "",
 181                 (hdr->Flags & TCP_FLAG_ACK) ? "ACK " : "",
 182                 (hdr->Flags & TCP_FLAG_PSH) ? "PSH " : "",
 183                 (hdr->Flags & TCP_FLAG_RST) ? "RST " : "",
 184                 (hdr->Flags & TCP_FLAG_SYN) ? "SYN " : "",
 185                 (hdr->Flags & TCP_FLAG_FIN) ? "FIN " : ""
 186                 );
 187         #endif
 188
 189         if( Length > (hdr->DataOffset >> 4)*4 )
 190         {
 191                 LOG("SequenceNumber = 0x%x", ntohl(hdr->SequenceNumber));
 192 #if HEXDUMP_INCOMING
 193                 Debug_HexDump(
 194                         "TCP_GetPacket: Packet Data = ",
 195                         (Uint8*)hdr + (hdr->DataOffset >> 4)*4,
 196                         Length - (hdr->DataOffset >> 4)*4
 197                         );
 198 #endif
 199         }
 200
 201         // Check Servers
 202         for( tTCPListener *srv = gTCP_Listeners; srv; srv = srv->Next )
 203         {
 204                 // Check if the server is active
 205                 if(srv->Port == 0)      continue;
 206                 // Check the interface
 207                 if(srv->Interface && srv->Interface != Interface)       continue;
 208                 // Check the destination port
 209                 if(srv->Port != htons(hdr->DestPort))   continue;
 210
 211                 Log_Log("TCP", "TCP_GetPacket: Matches server %p", srv);
 212                 // Is this in an established connection?
 213                 for( tTCPConnection *conn = srv->Connections; conn; conn = conn->Next )
 214                 {
 215                         // Check that it is coming in on the same interface
 216                         if(conn->Interface != Interface)        continue;
 217
 218                         // Check Source Port
 219                         Log_Log("TCP", "TCP_GetPacket: conn->RemotePort(%i) == hdr->SourcePort(%i)",
 220                                 conn->RemotePort, ntohs(hdr->SourcePort));
 221                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 222
 223                         // Check Source IP
 224                         Log_Debug("TCP", "TCP_GetPacket: conn->RemoteIP(%s)",
 225                                 IPStack_PrintAddress(conn->Interface->Type, &conn->RemoteIP));
 226                         Log_Debug("TCP", "                == Address(%s)",
 227                                 IPStack_PrintAddress(conn->Interface->Type, Address));
 228                         if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 229                                 continue ;
 230
 231                         Log_Log("TCP", "TCP_GetPacket: Matches connection %p", conn);
 232                         // We have a response!
 233                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 234
 235                         return;
 236                 }
 237
 238                 Log_Log("TCP", "TCP_GetPacket: Opening Connection");
 239                 // Open a new connection (well, check that it's a SYN)
 240                 if(hdr->Flags != TCP_FLAG_SYN) {
 241                         Log_Log("TCP", "TCP_GetPacket: Packet is not a SYN");
 242                         return ;
 243                 }
 244
 245                 // TODO: Check for halfopen max
 246
 247                 tTCPConnection *conn = TCP_int_CreateConnection(Interface, TCP_ST_SYN_RCVD);
 248                 conn->LocalPort = srv->Port;
 249                 conn->RemotePort = ntohs(hdr->SourcePort);
 250
 251                 switch(Interface->Type)
 252                 {
 253                 case 4: conn->RemoteIP.v4 = *(tIPv4*)Address;   break;
 254                 case 6: conn->RemoteIP.v6 = *(tIPv6*)Address;   break;
 255                 }
 256
 257                 conn->NextSequenceRcv = ntohl( hdr->SequenceNumber ) + 1;
 258                 conn->HighestSequenceRcvd = conn->NextSequenceRcv;
 259                 conn->NextSequenceSend = rand();
 260
 261                 conn->Node.ImplInt = srv->NextID ++;
 262
 263                 // Hmm... Theoretically, this lock will never have to wait,
 264                 // as the interface is locked to the watching thread, and this
 265                 // runs in the watching thread. But, it's a good idea to have
 266                 // it, just in case
 267                 // Oh, wait, there is a case where a wildcard can be used
 268                 // (srv->Interface == NULL) so having the lock is a good idea
 269                 SHORTLOCK(&srv->lConnections);
 270                 if( !srv->Connections )
 271                         srv->Connections = conn;
 272                 else
 273                         srv->ConnectionsTail->Next = conn;
 274                 srv->ConnectionsTail = conn;
 275                 if(!srv->NewConnections)
 276                         srv->NewConnections = conn;
 277                 VFS_MarkAvaliable( &srv->Node, 1 );
 278                 SHORTREL(&srv->lConnections);
 279                 Semaphore_Signal(&srv->WaitingConnections, 1);
 280
 281                 // Send the SYN ACK
 282                 hdr->Flags |= TCP_FLAG_ACK;
 283                 hdr->AcknowlegementNumber = htonl(conn->NextSequenceRcv);
 284                 hdr->SequenceNumber = htonl(conn->NextSequenceSend);
 285                 hdr->DestPort = hdr->SourcePort;
 286                 hdr->SourcePort = htons(srv->Port);
 287                 hdr->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 288                 TCP_SendPacket( conn, hdr, 0, NULL );
 289                 conn->NextSequenceSend ++;
 290                 return ;
 291         }
 292
 293         // Check Open Connections
 294         {
 295                 for( tTCPConnection *conn = gTCP_OutbountCons; conn; conn = conn->Next )
 296                 {
 297                         // Check that it is coming in on the same interface
 298                         if(conn->Interface != Interface)        continue;
 299
 300                         // Check Source Port
 301                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 302
 303                         // Check Source IP
 304                         if(conn->Interface->Type == 6 && !IP6_EQU(conn->RemoteIP.v6, *(tIPv6*)Address))
 305                                 continue;
 306                         if(conn->Interface->Type == 4 && !IP4_EQU(conn->RemoteIP.v4, *(tIPv4*)Address))
 307                                 continue;
 308
 309                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 310                         return ;
 311                 }
 312         }
 313
 314         Log_Log("TCP", "TCP_GetPacket: No Match");
 315         // If not a RST, send a RST
 316         if( !(hdr->Flags & TCP_FLAG_RST) )
 317         {
 318                 tTCPHeader      out_hdr = {0};
 319
 320                 out_hdr.DataOffset = (sizeof(out_hdr)/4) << 4;
 321                 out_hdr.DestPort = hdr->SourcePort;
 322                 out_hdr.SourcePort = hdr->DestPort;
 323
 324                 size_t  data_len = Length - (hdr->DataOffset>>4)*4;
 325                 out_hdr.AcknowlegementNumber = htonl( ntohl(hdr->SequenceNumber) + data_len );
 326                 if( hdr->Flags & TCP_FLAG_ACK ) {
 327                         out_hdr.Flags = TCP_FLAG_RST;
 328                         out_hdr.SequenceNumber = hdr->AcknowlegementNumber;
 329                 }
 330                 else {
 331                         out_hdr.Flags = TCP_FLAG_RST|TCP_FLAG_ACK;
 332                         out_hdr.SequenceNumber = 0;
 333                 }
 334                 TCP_int_SendPacket(Interface, Address, &out_hdr, 0, NULL);
 335         }
 336 }
 337
 338 /**
 339  * \brief Handles a packet sent to a specific connection
 340  * \param Connection    TCP Connection pointer
 341  * \param Header        TCP Packet pointer
 342  * \param Length        Length of the packet
 343  */
 344 void TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length)
 345 {
 346          int    dataLen;
 347         Uint32  sequence_num;
 348
 349         // Silently drop once finished
 350         // TODO: Check if this needs to be here
 351         if( Connection->State == TCP_ST_FINISHED ) {
 352                 Log_Log("TCP", "Packet ignored - connection finnished");
 353                 return ;
 354         }
 355
 356         // Syncronise sequence values
 357         if(Header->Flags & TCP_FLAG_SYN) {
 358                 // TODO: What if the packet also has data?
 359                 if( Connection->LastACKSequence != Connection->NextSequenceRcv )
 360                         TCP_INT_SendACK(Connection, "SYN");
 361                 Connection->NextSequenceRcv = ntohl(Header->SequenceNumber);
 362                 // TODO: Process HighestSequenceRcvd
 363                 // HACK!
 364                 if( Connection->HighestSequenceRcvd == 0 )
 365                         Connection->HighestSequenceRcvd = Connection->NextSequenceRcv;
 366                 Connection->LastACKSequence = Connection->NextSequenceRcv;
 367         }
 368
 369         // Ackowledge a sent packet
 370         if(Header->Flags & TCP_FLAG_ACK) {
 371                 // TODO: Process an ACKed Packet
 372                 LOG("Conn %p, Sent packet 0x%x ACKed", Connection, Header->AcknowlegementNumber);
 373         }
 374
 375         // Get length of data
 376         dataLen = Length - (Header->DataOffset>>4)*4;
 377         LOG("dataLen = %i", dataLen);
 378         #if TCP_DEBUG
 379         Log_Debug("TCP", "State %i, dataLen = %x", Connection->State, dataLen);
 380         #endif
 381
 382         //
 383         // State Machine
 384         //
 385         switch( Connection->State )
 386         {
 387         // Pre-init connection?
 388         case TCP_ST_CLOSED:
 389                 Log_Log("TCP", "Packets to a closed connection?!");
 390                 break;
 391
 392         // --- Init States ---
 393         // SYN sent, expecting SYN-ACK Connection Opening
 394         case TCP_ST_SYN_SENT:
 395                 if( Header->Flags & TCP_FLAG_SYN )
 396                 {
 397                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 398                                 Connection->HighestSequenceRcvd ++;
 399                         Connection->NextSequenceRcv ++;
 400
 401                         if( Header->Flags & TCP_FLAG_ACK )
 402                         {
 403                                 Log_Log("TCP", "ACKing SYN-ACK");
 404                                 Connection->State = TCP_ST_OPEN;
 405                                 VFS_MarkFull(&Connection->Node, 0);
 406                         }
 407                         else
 408                         {
 409                                 Log_Log("TCP", "ACKing SYN");
 410                                 Connection->State = TCP_ST_SYN_RCVD;
 411                         }
 412                         Header->DestPort = Header->SourcePort;
 413                         Header->SourcePort = htons(Connection->LocalPort);
 414                         Header->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 415                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 416                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 417                         Header->Flags = TCP_FLAG_ACK;
 418                         Header->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 419                         TCP_SendPacket( Connection, Header, 0, NULL );
 420                 }
 421                 break;
 422
 423         // SYN-ACK sent, expecting ACK
 424         case TCP_ST_SYN_RCVD:
 425                 if( Header->Flags & TCP_FLAG_ACK )
 426                 {
 427                         // TODO: Handle max half-open limit
 428                         Log_Log("TCP", "Connection fully opened");
 429                         Connection->State = TCP_ST_OPEN;
 430                         VFS_MarkFull(&Connection->Node, 0);
 431                 }
 432                 break;
 433
 434         // --- Established State ---
 435         case TCP_ST_OPEN:
 436                 // - Handle State changes
 437                 //
 438                 if( Header->Flags & TCP_FLAG_FIN ) {
 439                         Log_Log("TCP", "Conn %p closed, recieved FIN", Connection);
 440                         VFS_MarkError(&Connection->Node, 1);
 441                         Connection->State = TCP_ST_CLOSE_WAIT;
 442 //                      Header->Flags &= ~TCP_FLAG_FIN;
 443                         // CLOSE WAIT requires the client to close (or does it?)
 444                         #if 0
 445
 446                         #endif
 447                 }
 448
 449                 // Check for an empty packet
 450                 if(dataLen == 0) {
 451                         if( Header->Flags == TCP_FLAG_ACK )
 452                         {
 453                                 Log_Log("TCP", "ACK only packet");
 454                                 return ;
 455                         }
 456                         // TODO: Is this right? (empty packet counts as one byte)
 457                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 458                                 Connection->HighestSequenceRcvd ++;
 459                         Connection->NextSequenceRcv ++;
 460                         Log_Log("TCP", "Empty Packet, inc and ACK the current sequence number");
 461                         TCP_INT_SendACK(Connection, "Empty");
 462                         return ;
 463                 }
 464
 465                 // NOTES:
 466                 // Flags
 467                 //    PSH - Has Data?
 468                 // /NOTES
 469
 470                 sequence_num = ntohl(Header->SequenceNumber);
 471
 472                 LOG("0x%08x <= 0x%08x < 0x%08x",
 473                         Connection->NextSequenceRcv,
 474                         ntohl(Header->SequenceNumber),
 475                         Connection->NextSequenceRcv + TCP_WINDOW_SIZE
 476                         );
 477
 478                 // Is this packet the next expected packet?
 479                 if( sequence_num == Connection->NextSequenceRcv )
 480                 {
 481                          int    rv;
 482                         // Ooh, Goodie! Add it to the recieved list
 483                         rv = TCP_INT_AppendRecieved(Connection,
 484                                 (Uint8*)Header + (Header->DataOffset>>4)*4,
 485                                 dataLen
 486                                 );
 487                         if(rv != 0) {
 488                                 Log_Notice("TCP", "TCP_INT_AppendRecieved rv %i", rv);
 489                                 break;
 490                         }
 491                         LOG("0x%08x += %i", Connection->NextSequenceRcv, dataLen);
 492                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 493                                 Connection->HighestSequenceRcvd += dataLen;
 494                         Connection->NextSequenceRcv += dataLen;
 495
 496                         // TODO: This should be moved out of the watcher thread,
 497                         // so that a single lost packet on one connection doesn't cause
 498                         // all connections on the interface to lag.
 499                         // - Meh, no real issue, as the cache shouldn't be that large
 500                         TCP_INT_UpdateRecievedFromFuture(Connection);
 501
 502                         #if 1
 503                         // - Only send an ACK if we've had a burst
 504                         if( Connection->NextSequenceRcv > (Uint32)(TCP_DACK_THRESHOLD + Connection->LastACKSequence) )
 505                         {
 506                                 TCP_INT_SendACK(Connection, "DACK Burst");
 507                                 // - Extend TCP deferred ACK timer
 508                                 Time_RemoveTimer(Connection->DeferredACKTimer);
 509                         }
 510                         // - Schedule the deferred ACK timer (if already scheduled, this is a NOP)
 511                         Time_ScheduleTimer(Connection->DeferredACKTimer, TCP_DACK_TIMEOUT);
 512                         #else
 513                         TCP_INT_SendACK(Connection, "RX");
 514                         #endif
 515                 }
 516                 // Check if the packet is in window
 517                 else if( WrapBetween(Connection->NextSequenceRcv, sequence_num,
 518                                 Connection->NextSequenceRcv+TCP_WINDOW_SIZE, 0xFFFFFFFF) )
 519                 {
 520                         Uint8   *dataptr = (Uint8*)Header + (Header->DataOffset>>4)*4;
 521                         #if CACHE_FUTURE_PACKETS_IN_BYTES
 522                         Uint32  index;
 523
 524                         index = sequence_num % TCP_WINDOW_SIZE;
 525                         for( int i = 0; i < dataLen; i ++ )
 526                         {
 527                                 Connection->FuturePacketValidBytes[index/8] |= 1 << (index%8);
 528                                 Connection->FuturePacketData[index] = dataptr[i];
 529                                 // Do a wrap increment
 530                                 index ++;
 531                                 if(index == TCP_WINDOW_SIZE)    index = 0;
 532                         }
 533                         #else
 534                         tTCPStoredPacket        *pkt, *tmp, *prev = NULL;
 535
 536                         // Allocate and fill cached packet
 537                         pkt = malloc( sizeof(tTCPStoredPacket) + dataLen );
 538                         pkt->Next = NULL;
 539                         pkt->Sequence = ntohl(Header->SequenceNumber);
 540                         pkt->Length = dataLen;
 541                         memcpy(pkt->Data, dataptr, dataLen);
 542
 543                         Log_Log("TCP", "We missed a packet, caching",
 544                                 pkt->Sequence, Connection->NextSequenceRcv);
 545
 546                         // No? Well, let's cache it and look at it later
 547                         SHORTLOCK( &Connection->lFuturePackets );
 548                         for(tmp = Connection->FuturePackets;
 549                                 tmp;
 550                                 prev = tmp, tmp = tmp->Next)
 551                         {
 552                                 if(tmp->Sequence >= pkt->Sequence)      break;
 553                         }
 554
 555                         // Add if before first, or sequences don't match
 556                         if( !tmp || tmp->Sequence != pkt->Sequence )
 557                         {
 558                                 if(prev)
 559                                         prev->Next = pkt;
 560                                 else
 561                                         Connection->FuturePackets = pkt;
 562                                 pkt->Next = tmp;
 563                         }
 564                         // Replace if larger
 565                         else if(pkt->Length > tmp->Length)
 566                         {
 567                                 if(prev)
 568                                         prev->Next = pkt;
 569                                 pkt->Next = tmp->Next;
 570                                 free(tmp);
 571                         }
 572                         else
 573                         {
 574                                 free(pkt);      // TODO: Find some way to remove this
 575                         }
 576                         SHORTREL( &Connection->lFuturePackets );
 577                         #endif
 578                 }
 579                 // Badly out of sequence packet
 580                 else
 581                 {
 582                         Log_Log("TCP", "Fully out of sequence packet (0x%08x not between 0x%08x and 0x%08x), dropped",
 583                                 sequence_num, Connection->NextSequenceRcv, Connection->NextSequenceRcv+TCP_WINDOW_SIZE);
 584                         // Spec says we should send an empty ACK with the current state
 585                         TCP_INT_SendACK(Connection, "Bad Seq");
 586                 }
 587                 break;
 588
 589         // --- Remote close states
 590         case TCP_ST_CLOSE_WAIT:
 591
 592                 // Ignore everything, CLOSE_WAIT is terminated by the client
 593                 Log_Debug("TCP", "CLOSE WAIT - Ignoring packets");
 594
 595                 break;
 596
 597         // LAST-ACK - Waiting for the ACK of FIN (from CLOSE WAIT)
 598         case TCP_ST_LAST_ACK:
 599                 if( Header->Flags & TCP_FLAG_ACK )
 600                 {
 601                         Connection->State = TCP_ST_FINISHED;    // Connection completed
 602                         Log_Log("TCP", "LAST-ACK to CLOSED - Connection remote closed");
 603                         // TODO: Destrory the TCB
 604                 }
 605                 break;
 606
 607         // --- Local close States
 608         case TCP_ST_FIN_WAIT1:
 609                 if( Header->Flags & TCP_FLAG_FIN )
 610                 {
 611                         Connection->State = TCP_ST_CLOSING;
 612                         Log_Debug("TCP", "Conn %p closed, sent FIN and recieved FIN", Connection);
 613                         VFS_MarkError(&Connection->Node, 1);
 614
 615                         // ACK Packet
 616                         Header->DestPort = Header->SourcePort;
 617                         Header->SourcePort = htons(Connection->LocalPort);
 618                         Header->AcknowlegementNumber = Header->SequenceNumber;
 619                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 620                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 621                         Header->Flags = TCP_FLAG_ACK;
 622                         TCP_SendPacket( Connection, Header, 0, NULL );
 623                         break ;
 624                 }
 625
 626                 // TODO: Make sure that the packet is actually ACKing the FIN
 627                 if( Header->Flags & TCP_FLAG_ACK )
 628                 {
 629                         Connection->State = TCP_ST_FIN_WAIT2;
 630                         Log_Debug("TCP", "Conn %p closed, sent FIN ACKed", Connection);
 631                         VFS_MarkError(&Connection->Node, 1);
 632                         return ;
 633                 }
 634                 break;
 635
 636         case TCP_ST_FIN_WAIT2:
 637                 if( Header->Flags & TCP_FLAG_FIN )
 638                 {
 639                         Connection->State = TCP_ST_TIME_WAIT;
 640                         Log_Debug("TCP", "FIN sent and recieved, ACKing and going into TIME WAIT %p FINWAIT-2 -> TIME WAIT", Connection);
 641                         // Send ACK
 642                         Header->DestPort = Header->SourcePort;
 643                         Header->SourcePort = htons(Connection->LocalPort);
 644                         Header->AcknowlegementNumber = Header->SequenceNumber;
 645                         Header->SequenceNumber = htonl(Connection->NextSequenceSend);
 646                         Header->WindowSize = htons(TCP_WINDOW_SIZE);
 647                         Header->Flags = TCP_FLAG_ACK;
 648                         TCP_SendPacket( Connection, Header, 0, NULL );
 649                 }
 650                 break;
 651
 652         case TCP_ST_CLOSING:
 653                 // TODO: Make sure that the packet is actually ACKing the FIN
 654                 if( Header->Flags & TCP_FLAG_ACK )
 655                 {
 656                         Connection->State = TCP_ST_TIME_WAIT;
 657                         Log_Debug("TCP", "Conn %p CLOSING -> TIME WAIT", Connection);
 658                         VFS_MarkError(&Connection->Node, 1);
 659                         return ;
 660                 }
 661                 break;
 662
 663         // --- Closed (or near closed) states) ---
 664         case TCP_ST_TIME_WAIT:
 665                 Log_Log("TCP", "Packets on Time-Wait, ignored");
 666                 break;
 667
 668         case TCP_ST_FINISHED:
 669                 Log_Log("TCP", "Packets when CLOSED, ignoring");
 670                 break;
 671
 672         //default:
 673         //      Log_Warning("TCP", "Unhandled TCP state %i", Connection->State);
 674         //      break;
 675         }
 676
 677 }
 678
 679 /**
 680  * \brief Appends a packet to the recieved list
 681  * \param Connection    Connection structure
 682  * \param Data  Packet contents
 683  * \param Length        Length of \a Data
 684  */
 685 int TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length)
 686 {
 687         Mutex_Acquire( &Connection->lRecievedPackets );
 688
 689         if(Connection->RecievedBuffer->Length + Length > Connection->RecievedBuffer->Space )
 690         {
 691                 VFS_MarkAvaliable(&Connection->Node, 1);
 692                 Log_Error("TCP", "Buffer filled, packet dropped (:%i) - %i + %i > %i",
 693                         Connection->LocalPort, Connection->RecievedBuffer->Length, Length,
 694                         Connection->RecievedBuffer->Space
 695                         );
 696                 Mutex_Release( &Connection->lRecievedPackets );
 697                 return 1;
 698         }
 699
 700         RingBuffer_Write( Connection->RecievedBuffer, Data, Length );
 701
 702         VFS_MarkAvaliable(&Connection->Node, 1);
 703
 704         Mutex_Release( &Connection->lRecievedPackets );
 705         return 0;
 706 }
 707
 708 /**
 709  * \brief Updates the connections recieved list from the future list
 710  * \param Connection    Connection structure
 711  *
 712  * Updates the recieved packets list with packets from the future (out
 713  * of order) packets list that are now able to be added in direct
 714  * sequence.
 715  */
 716 void TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection)
 717 {
 718         #if CACHE_FUTURE_PACKETS_IN_BYTES
 719         // Calculate length of contiguous bytes
 720          int    length = Connection->HighestSequenceRcvd - Connection->NextSequenceRcv;
 721         Uint32  index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 722         LOG("length=%i, index=%i", length, index);
 723         for( int i = 0; i < length; i ++ )
 724         {
 725                  int    bit = index % 8;
 726                 Uint8   bitfield_byte = Connection->FuturePacketValidBytes[index / 8];
 727                 if( (bitfield_byte & (1 << bit)) == 0 ) {
 728                         length = i;
 729                         break;
 730                 }
 731
 732                 if( bitfield_byte == 0xFF ) {
 733                          int    inc = 8 - bit;
 734                         i += inc - 1;
 735                         index += inc;
 736                 }
 737                 else {
 738                         index ++;
 739                 }
 740                 if(index > TCP_WINDOW_SIZE)
 741                         index -= TCP_WINDOW_SIZE;
 742         }
 743
 744         index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 745
 746         // Write data to to the ring buffer
 747         if( TCP_WINDOW_SIZE - index > length )
 748         {
 749                 // Simple case
 750                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, length );
 751         }
 752         else
 753         {
 754                  int    endLen = TCP_WINDOW_SIZE - index;
 755                 // 2-part case
 756                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, endLen );
 757                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData, endLen - length );
 758         }
 759
 760         // Mark (now saved) bytes as invalid
 761         // - Align index
 762         while(index % 8 && length > 0)
 763         {
 764                 Connection->FuturePacketData[index] = 0;
 765                 Connection->FuturePacketValidBytes[index/8] &= ~(1 << (index%8));
 766                 index ++;
 767                 if(index > TCP_WINDOW_SIZE)
 768                         index -= TCP_WINDOW_SIZE;
 769                 length --;
 770         }
 771         while( length > 7 )
 772         {
 773                 Connection->FuturePacketData[index] = 0;
 774                 Connection->FuturePacketValidBytes[index/8] = 0;
 775                 length -= 8;
 776                 index += 8;
 777                 if(index > TCP_WINDOW_SIZE)
 778                         index -= TCP_WINDOW_SIZE;
 779         }
 780         while(length)
 781         {
 782                 Connection->FuturePacketData[index] = 0;
 783                 Connection->FuturePacketData[index/8] &= ~(1 << (index%8));
 784                 index ++;
 785                 if(index > TCP_WINDOW_SIZE)
 786                         index -= TCP_WINDOW_SIZE;
 787                 length --;
 788         }
 789
 790         #else
 791         tTCPStoredPacket        *pkt;
 792         for(;;)
 793         {
 794                 SHORTLOCK( &Connection->lFuturePackets );
 795
 796                 // Clear out duplicates from cache
 797                 // - If a packet has just been recieved, and it is expected, then
 798                 //   (since NextSequenceRcv = rcvd->Sequence + rcvd->Length) all
 799                 //   packets in cache that are smaller than the next expected
 800                 //   are now defunct.
 801                 pkt = Connection->FuturePackets;
 802                 while(pkt && pkt->Sequence < Connection->NextSequenceRcv)
 803                 {
 804                         tTCPStoredPacket        *next = pkt->Next;
 805                         free(pkt);
 806                         pkt = next;
 807                 }
 808
 809                 // If there's no packets left in cache, stop looking
 810                 if(!pkt || pkt->Sequence > Connection->NextSequenceRcv) {
 811                         SHORTREL( &Connection->lFuturePackets );
 812                         return;
 813                 }
 814
 815                 // Delete packet from future list
 816                 Connection->FuturePackets = pkt->Next;
 817
 818                 // Release list
 819                 SHORTREL( &Connection->lFuturePackets );
 820
 821                 // Looks like we found one
 822                 TCP_INT_AppendRecieved(Connection, pkt->Data, pkt->Length);
 823                 if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 824                         Connection->HighestSequenceRcvd += pkt->Length;
 825                 Connection->NextSequenceRcv += pkt->Length;
 826                 free(pkt);
 827         }
 828         #endif
 829 }
 830
 831 void TCP_int_SendDelayedACK(void *ConnPtr)
 832 {
 833         TCP_INT_SendACK(ConnPtr, "DACK Timeout");
 834 }
 835
 836 void TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason)
 837 {
 838         tTCPHeader      hdr;
 839         // ACK Packet
 840         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
 841         hdr.DestPort = htons(Connection->RemotePort);
 842         hdr.SourcePort = htons(Connection->LocalPort);
 843         hdr.AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 844         hdr.SequenceNumber = htonl(Connection->NextSequenceSend);
 845         hdr.WindowSize = htons(TCP_WINDOW_SIZE);
 846         hdr.Flags = TCP_FLAG_ACK;       // TODO: Determine if SYN is wanted too
 847         hdr.Checksum = 0;       // TODO: Checksum
 848         hdr.UrgentPointer = 0;
 849         Log_Debug("TCP", "Sending ACK for 0x%08x (%s)", Connection->NextSequenceRcv, Reason);
 850         TCP_SendPacket( Connection, &hdr, 0, NULL );
 851         //Connection->NextSequenceSend ++;
 852         Connection->LastACKSequence = Connection->NextSequenceRcv;
 853 }
 854
 855 /**
 856  * \fn Uint16 TCP_GetUnusedPort()
 857  * \brief Gets an unused port and allocates it
 858  */
 859 Uint16 TCP_GetUnusedPort()
 860 {
 861         Uint16  ret;
 862
 863         // Get Next outbound port
 864         ret = giTCP_NextOutPort++;
 865         while( gaTCP_PortBitmap[ret/32] & (1UL << (ret%32)) )
 866         {
 867                 ret ++;
 868                 giTCP_NextOutPort++;
 869                 if(giTCP_NextOutPort == 0x10000) {
 870                         ret = giTCP_NextOutPort = TCP_MIN_DYNPORT;
 871                 }
 872         }
 873
 874         // Mark the new port as used
 875         gaTCP_PortBitmap[ret/32] |= 1 << (ret%32);
 876
 877         return ret;
 878 }
 879
 880 /**
 881  * \fn int TCP_AllocatePort(Uint16 Port)
 882  * \brief Marks a port as used
 883  */
 884 int TCP_AllocatePort(Uint16 Port)
 885 {
 886         // Check if the port has already been allocated
 887         if( gaTCP_PortBitmap[Port/32] & (1 << (Port%32)) )
 888                 return 0;
 889
 890         // Allocate
 891         gaTCP_PortBitmap[Port/32] |= 1 << (Port%32);
 892
 893         return 1;
 894 }
 895
 896 /**
 897  * \fn int TCP_DeallocatePort(Uint16 Port)
 898  * \brief Marks a port as unused
 899  */
 900 int TCP_DeallocatePort(Uint16 Port)
 901 {
 902         // Check if the port has already been allocated
 903         if( !(gaTCP_PortBitmap[Port/32] & (1 << (Port%32))) )
 904                 return 0;
 905
 906         // Allocate
 907         gaTCP_PortBitmap[Port/32] &= ~(1 << (Port%32));
 908
 909         return 1;
 910 }
 911
 912 tTCPConnection *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State)
 913 {
 914         tTCPConnection  *conn = calloc( sizeof(tTCPConnection) + TCP_WINDOW_SIZE + TCP_WINDOW_SIZE/8, 1 );
 915
 916         conn->State = State;
 917         conn->Interface = Interface;
 918         conn->LocalPort = -1;
 919         conn->RemotePort = -1;
 920
 921         conn->Node.ReferenceCount = 1;
 922         conn->Node.ImplPtr = conn;
 923         conn->Node.NumACLs = 1;
 924         conn->Node.ACLs = &gVFS_ACL_EveryoneRW;
 925         conn->Node.Type = &gTCP_ClientNodeType;
 926         conn->Node.BufferFull = 1;      // Cleared when connection opens
 927
 928         conn->RecievedBuffer = RingBuffer_Create( TCP_RECIEVE_BUFFER_SIZE );
 929         #if 0
 930         conn->SentBuffer = RingBuffer_Create( TCP_SEND_BUFFER_SIZE );
 931         Semaphore_Init(conn->SentBufferSpace, 0, TCP_SEND_BUFFER_SIZE, "TCP SentBuffer", conn->Name);
 932         #endif
 933
 934         #if CACHE_FUTURE_PACKETS_IN_BYTES
 935         // Future recieved data (ahead of the expected sequence number)
 936         conn->FuturePacketData = (Uint8*)conn + sizeof(tTCPConnection);
 937         conn->FuturePacketValidBytes = conn->FuturePacketData + TCP_WINDOW_SIZE;
 938         #endif
 939
 940         conn->DeferredACKTimer = Time_AllocateTimer( TCP_int_SendDelayedACK, conn);
 941         return conn;
 942 }
 943
 944 // --- Server
 945 tVFS_Node *TCP_Server_Init(tInterface *Interface)
 946 {
 947         tTCPListener    *srv;
 948
 949         srv = calloc( 1, sizeof(tTCPListener) );
 950
 951         if( srv == NULL ) {
 952                 Log_Warning("TCP", "malloc failed for listener (%i) bytes", sizeof(tTCPListener));
 953                 return NULL;
 954         }
 955
 956         srv->Interface = Interface;
 957         srv->Port = 0;
 958         srv->NextID = 0;
 959         srv->Connections = NULL;
 960         srv->ConnectionsTail = NULL;
 961         srv->NewConnections = NULL;
 962         srv->Next = NULL;
 963         srv->Node.Flags = VFS_FFLAG_DIRECTORY;
 964         srv->Node.Size = -1;
 965         srv->Node.ImplPtr = srv;
 966         srv->Node.NumACLs = 1;
 967         srv->Node.ACLs = &gVFS_ACL_EveryoneRW;
 968         srv->Node.Type = &gTCP_ServerNodeType;
 969
 970         SHORTLOCK(&glTCP_Listeners);
 971         srv->Next = gTCP_Listeners;
 972         gTCP_Listeners = srv;
 973         SHORTREL(&glTCP_Listeners);
 974
 975         return &srv->Node;
 976 }
 977
 978 /**
 979  * \brief Wait for a new connection and return the connection ID
 980  * \note Blocks until a new connection is made
 981  * \param Node  Server node
 982  * \param Pos   Position (ignored)
 983  */
 984 int TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Dest[FILENAME_MAX])
 985 {
 986         tTCPListener    *srv = Node->ImplPtr;
 987         tTCPConnection  *conn;
 988
 989         ENTER("pNode iPos", Node, Pos);
 990
 991         Log_Log("TCP", "Thread %i waiting for a connection", Threads_GetTID());
 992         Semaphore_Wait( &srv->WaitingConnections, 1 );
 993
 994         SHORTLOCK(&srv->lConnections);
 995         // Increment the new list (the current connection is still on the
 996         // normal list)
 997         conn = srv->NewConnections;
 998         srv->NewConnections = conn->Next;
 999
1000         if( srv->NewConnections == NULL )
1001                 VFS_MarkAvaliable( Node, 0 );
1002
1003         SHORTREL( &srv->lConnections );
1004
1005         LOG("conn = %p", conn);
1006         LOG("srv->Connections = %p", srv->Connections);
1007         LOG("srv->NewConnections = %p", srv->NewConnections);
1008         LOG("srv->ConnectionsTail = %p", srv->ConnectionsTail);
1009
1010         itoa(Dest, conn->Node.ImplInt, 16, 8, '0');
1011         Log_Log("TCP", "Thread %i got connection '%s'", Threads_GetTID(), Dest);
1012         LEAVE('i', 0);
1013         return 0;
1014 }
1015
1016 /**
1017  * \brief Gets a client connection node
1018  * \param Node  Server node
1019  * \param Name  Hexadecimal ID of the node
1020  */
1021 tVFS_Node *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags)
1022 {
1023         tTCPConnection  *conn;
1024         tTCPListener    *srv = Node->ImplPtr;
1025         char    tmp[9];
1026          int    id = atoi(Name);
1027
1028         ENTER("pNode sName", Node, Name);
1029
1030         // Check for a non-empty name
1031         if( Name[0] )
1032         {
1033                 // Sanity Check
1034                 itoa(tmp, id, 16, 8, '0');
1035                 if(strcmp(tmp, Name) != 0) {
1036                         LOG("'%s' != '%s' (%08x)", Name, tmp, id);
1037                         LEAVE('n');
1038                         return NULL;
1039                 }
1040
1041                 Log_Debug("TCP", "srv->Connections = %p", srv->Connections);
1042                 Log_Debug("TCP", "srv->NewConnections = %p", srv->NewConnections);
1043                 Log_Debug("TCP", "srv->ConnectionsTail = %p", srv->ConnectionsTail);
1044
1045                 // Search
1046                 SHORTLOCK( &srv->lConnections );
1047                 for(conn = srv->Connections;
1048                         conn;
1049                         conn = conn->Next)
1050                 {
1051                         LOG("conn->Node.ImplInt = %i", conn->Node.ImplInt);
1052                         if(conn->Node.ImplInt == id)    break;
1053                 }
1054                 SHORTREL( &srv->lConnections );
1055
1056                 // If not found, ret NULL
1057                 if(!conn) {
1058                         LOG("Connection %i not found", id);
1059                         LEAVE('n');
1060                         return NULL;
1061                 }
1062         }
1063         // Empty Name - Check for a new connection and if it's there, open it
1064         else
1065         {
1066                 SHORTLOCK( &srv->lConnections );
1067                 conn = srv->NewConnections;
1068                 if( conn != NULL )
1069                         srv->NewConnections = conn->Next;
1070                 VFS_MarkAvaliable( Node, srv->NewConnections != NULL );
1071                 SHORTREL( &srv->lConnections );
1072                 if( !conn ) {
1073                         LOG("No new connections");
1074                         LEAVE('n');
1075                         return NULL;
1076                 }
1077         }
1078
1079         // Return node
1080         LEAVE('p', &conn->Node);
1081         return &conn->Node;
1082 }
1083
1084 /**
1085  * \brief Handle IOCtl calls
1086  */
1087 int TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data)
1088 {
1089         tTCPListener    *srv = Node->ImplPtr;
1090
1091         switch(ID)
1092         {
1093         case 4: // Get/Set Port
1094                 if(!Data)       // Get Port
1095                         return srv->Port;
1096
1097                 if(srv->Port)   // Wait, you can't CHANGE the port
1098                         return -1;
1099
1100                 if(!CheckMem(Data, sizeof(Uint16)))     // Sanity check
1101                         return -1;
1102
1103                 // Permissions check
1104                 if(Threads_GetUID() != 0
1105                 && *(Uint16*)Data != 0
1106                 && *(Uint16*)Data < 1024)
1107                         return -1;
1108
1109                 // TODO: Check if a port is in use
1110
1111                 // Set Port
1112                 srv->Port = *(Uint16*)Data;
1113                 if(srv->Port == 0)      // Allocate a random port
1114                         srv->Port = TCP_GetUnusedPort();
1115                 else    // Else, mark this as used
1116                         TCP_AllocatePort(srv->Port);
1117
1118                 Log_Log("TCP", "Server %p listening on port %i", srv, srv->Port);
1119
1120                 return srv->Port;
1121         }
1122         return 0;
1123 }
1124
1125 void TCP_Server_Close(tVFS_Node *Node)
1126 {
1127         free(Node->ImplPtr);
1128 }
1129
1130 // --- Client
1131 /**
1132  * \brief Create a client node
1133  */
1134 tVFS_Node *TCP_Client_Init(tInterface *Interface)
1135 {
1136         tTCPConnection  *conn = TCP_int_CreateConnection(Interface, TCP_ST_CLOSED);
1137
1138         SHORTLOCK(&glTCP_OutbountCons);
1139         conn->Next = gTCP_OutbountCons;
1140         gTCP_OutbountCons = conn;
1141         SHORTREL(&glTCP_OutbountCons);
1142
1143         return &conn->Node;
1144 }
1145
1146 /**
1147  * \brief Wait for a packet and return it
1148  * \note If \a Length is smaller than the size of the packet, the rest
1149  *       of the packet's data will be discarded.
1150  */
1151 size_t TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags)
1152 {
1153         tTCPConnection  *conn = Node->ImplPtr;
1154         size_t  len;
1155
1156         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1157         LOG("conn = %p {State:%i}", conn, conn->State);
1158
1159         // If the connection has been closed (state > ST_OPEN) then clear
1160         // any stale data in the buffer (until it is empty (until it is empty))
1161         if( conn->State > TCP_ST_OPEN )
1162         {
1163                 Mutex_Acquire( &conn->lRecievedPackets );
1164                 len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1165                 Mutex_Release( &conn->lRecievedPackets );
1166
1167                 if( len == 0 ) {
1168                         VFS_MarkAvaliable(Node, 0);
1169                         errno = 0;
1170                         LEAVE('i', -1);
1171                         return -1;
1172                 }
1173
1174                 LEAVE('i', len);
1175                 return len;
1176         }
1177
1178         // Wait
1179         {
1180                 tTime   *timeout = NULL;
1181                 tTime   timeout_zero = 0;
1182                 if( Flags & VFS_IOFLAG_NOBLOCK )
1183                         timeout = &timeout_zero;
1184                 if( !VFS_SelectNode(Node, VFS_SELECT_READ|VFS_SELECT_ERROR, timeout, "TCP_Client_Read") ) {
1185                         errno = EWOULDBLOCK;
1186                         LEAVE('i', -1);
1187                         return -1;
1188                 }
1189         }
1190
1191         // Lock list and read as much as possible (up to `Length`)
1192         Mutex_Acquire( &conn->lRecievedPackets );
1193         len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1194
1195         if( len == 0 || conn->RecievedBuffer->Length == 0 ) {
1196                 LOG("Marking as none avaliable (len = %i)", len);
1197                 VFS_MarkAvaliable(Node, 0);
1198         }
1199
1200         // Release the lock (we don't need it any more)
1201         Mutex_Release( &conn->lRecievedPackets );
1202
1203         LEAVE('i', len);
1204         return len;
1205 }
1206
1207 /**
1208  * \brief Send a data packet on a connection
1209  */
1210 void TCP_INT_SendDataPacket(tTCPConnection *Connection, size_t Length, const void *Data)
1211 {
1212         char    buf[sizeof(tTCPHeader)+Length];
1213         tTCPHeader      *packet = (void*)buf;
1214
1215         // - Stop Delayed ACK timer (as this data packet ACKs)
1216         Time_RemoveTimer(Connection->DeferredACKTimer);
1217
1218         packet->SourcePort = htons(Connection->LocalPort);
1219         packet->DestPort = htons(Connection->RemotePort);
1220         packet->DataOffset = (sizeof(tTCPHeader)/4)*16;
1221         packet->WindowSize = htons(TCP_WINDOW_SIZE);
1222
1223         packet->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
1224         packet->SequenceNumber = htonl(Connection->NextSequenceSend);
1225         packet->Flags = TCP_FLAG_PSH|TCP_FLAG_ACK;      // Hey, ACK if you can!
1226         packet->UrgentPointer = 0;
1227
1228         memcpy(packet->Options, Data, Length);
1229
1230         Log_Debug("TCP", "Send sequence 0x%08x", Connection->NextSequenceSend);
1231 #if HEXDUMP_OUTGOING
1232         Debug_HexDump("TCP_INT_SendDataPacket: Data = ", Data, Length);
1233 #endif
1234
1235         TCP_SendPacket( Connection, packet, Length, Data );
1236
1237         Connection->NextSequenceSend += Length;
1238 }
1239
1240 /**
1241  * \brief Send some bytes on a connection
1242  */
1243 size_t TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags)
1244 {
1245         tTCPConnection  *conn = Node->ImplPtr;
1246         size_t  rem = Length;
1247
1248         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1249
1250 //      #if DEBUG
1251 //      Debug_HexDump("TCP_Client_Write: Buffer = ",
1252 //              Buffer, Length);
1253 //      #endif
1254
1255         // Don't allow a write to a closed connection
1256         if( conn->State > TCP_ST_OPEN ) {
1257                 VFS_MarkError(Node, 1);
1258                 errno = 0;
1259                 LEAVE('i', -1);
1260                 return -1;
1261         }
1262
1263         // Wait
1264         {
1265                 tTime   *timeout = NULL;
1266                 tTime   timeout_zero = 0;
1267                 if( Flags & VFS_IOFLAG_NOBLOCK )
1268                         timeout = &timeout_zero;
1269                 if( !VFS_SelectNode(Node, VFS_SELECT_WRITE|VFS_SELECT_ERROR, timeout, "TCP_Client_Write") ) {
1270                         errno = EWOULDBLOCK;
1271                         LEAVE('i', -1);
1272                         return -1;
1273                 }
1274         }
1275
1276         do
1277         {
1278                  int    len = (rem < TCP_MAX_PACKET_SIZE) ? rem : TCP_MAX_PACKET_SIZE;
1279
1280                 #if 0
1281                 // Wait for space in the buffer
1282                 Semaphore_Signal( &Connection->SentBufferSpace, len );
1283
1284                 // Save data to buffer (and update the length read by the ammount written)
1285                 len = RingBuffer_Write( &Connection->SentBuffer, Buffer, len);
1286                 #endif
1287
1288                 // Send packet
1289                 TCP_INT_SendDataPacket(conn, len, Buffer);
1290
1291                 Buffer += len;
1292                 rem -= len;
1293         } while( rem > 0 );
1294
1295         LEAVE('i', Length);
1296         return Length;
1297 }
1298
1299 /**
1300  * \brief Open a connection to another host using TCP
1301  * \param Conn  Connection structure
1302  */
1303 void TCP_StartConnection(tTCPConnection *Conn)
1304 {
1305         tTCPHeader      hdr = {0};
1306
1307         Conn->State = TCP_ST_SYN_SENT;
1308
1309         hdr.SourcePort = htons(Conn->LocalPort);
1310         hdr.DestPort = htons(Conn->RemotePort);
1311         Conn->NextSequenceSend = rand();
1312         hdr.SequenceNumber = htonl(Conn->NextSequenceSend);
1313         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
1314         hdr.Flags = TCP_FLAG_SYN;
1315         hdr.WindowSize = htons(TCP_WINDOW_SIZE);        // Max
1316         hdr.Checksum = 0;       // TODO
1317
1318         TCP_SendPacket( Conn, &hdr, 0, NULL );
1319
1320         Conn->NextSequenceSend ++;
1321         Conn->State = TCP_ST_SYN_SENT;
1322
1323         return ;
1324 }
1325
1326 /**
1327  * \brief Control a client socket
1328  */
1329 int TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data)
1330 {
1331         tTCPConnection  *conn = Node->ImplPtr;
1332
1333         ENTER("pNode iID pData", Node, ID, Data);
1334
1335         switch(ID)
1336         {
1337         case 4: // Get/Set local port
1338                 if(!Data)
1339                         LEAVE_RET('i', conn->LocalPort);
1340                 if(conn->State != TCP_ST_CLOSED)
1341                         LEAVE_RET('i', -1);
1342                 if(!CheckMem(Data, sizeof(Uint16)))
1343                         LEAVE_RET('i', -1);
1344
1345                 if(Threads_GetUID() != 0 && *(Uint16*)Data < 1024)
1346                         LEAVE_RET('i', -1);
1347
1348                 conn->LocalPort = *(Uint16*)Data;
1349                 LEAVE_RET('i', conn->LocalPort);
1350
1351         case 5: // Get/Set remote port
1352                 if(!Data)       LEAVE_RET('i', conn->RemotePort);
1353                 if(conn->State != TCP_ST_CLOSED)        LEAVE_RET('i', -1);
1354                 if(!CheckMem(Data, sizeof(Uint16)))     LEAVE_RET('i', -1);
1355                 conn->RemotePort = *(Uint16*)Data;
1356                 LEAVE_RET('i', conn->RemotePort);
1357
1358         case 6: // Set Remote IP
1359                 if( conn->State != TCP_ST_CLOSED )
1360                         LEAVE_RET('i', -1);
1361                 if( conn->Interface->Type == 4 )
1362                 {
1363                         if(!CheckMem(Data, sizeof(tIPv4)))      LEAVE_RET('i', -1);
1364                         conn->RemoteIP.v4 = *(tIPv4*)Data;
1365                 }
1366                 else if( conn->Interface->Type == 6 )
1367                 {
1368                         if(!CheckMem(Data, sizeof(tIPv6)))      LEAVE_RET('i', -1);
1369                         conn->RemoteIP.v6 = *(tIPv6*)Data;
1370                 }
1371                 LEAVE_RET('i', 0);
1372
1373         case 7: // Connect
1374                 if(conn->LocalPort == 0xFFFF)
1375                         conn->LocalPort = TCP_GetUnusedPort();
1376                 if(conn->RemotePort == -1)
1377                         LEAVE_RET('i', 0);
1378
1379                 {
1380                         tTime   timeout = conn->Interface->TimeoutDelay;
1381
1382                         TCP_StartConnection(conn);
1383                         VFS_SelectNode(&conn->Node, VFS_SELECT_WRITE, &timeout, "TCP Connection");
1384                         if( conn->State == TCP_ST_SYN_SENT )
1385                                 LEAVE_RET('i', 0);
1386                 }
1387
1388                 LEAVE_RET('i', 1);
1389
1390         // Get recieve buffer length
1391         case 8:
1392                 LEAVE_RET('i', conn->RecievedBuffer->Length);
1393         }
1394
1395         return 0;
1396 }
1397
1398 void TCP_Client_Close(tVFS_Node *Node)
1399 {
1400         tTCPConnection  *conn = Node->ImplPtr;
1401         tTCPHeader      packet;
1402
1403         ENTER("pNode", Node);
1404
1405         ASSERT(Node->ReferenceCount != 0);
1406
1407         if( Node->ReferenceCount > 1 ) {
1408                 Node->ReferenceCount --;
1409                 LOG("Dereference only");
1410                 LEAVE('-');
1411                 return ;
1412         }
1413
1414         if( conn->State == TCP_ST_CLOSE_WAIT || conn->State == TCP_ST_OPEN )
1415         {
1416                 packet.SourcePort = htons(conn->LocalPort);
1417                 packet.DestPort = htons(conn->RemotePort);
1418                 packet.DataOffset = (sizeof(tTCPHeader)/4)*16;
1419                 packet.WindowSize = TCP_WINDOW_SIZE;
1420
1421                 packet.AcknowlegementNumber = 0;
1422                 packet.SequenceNumber = htonl(conn->NextSequenceSend);
1423                 packet.Flags = TCP_FLAG_FIN;
1424
1425                 TCP_SendPacket( conn, &packet, 0, NULL );
1426         }
1427
1428         switch( conn->State )
1429         {
1430         case TCP_ST_CLOSED:
1431                 Log_Warning("TCP", "Closing connection that was never opened");
1432                 break;
1433         case TCP_ST_CLOSE_WAIT:
1434                 conn->State = TCP_ST_LAST_ACK;
1435                 break;
1436         case TCP_ST_OPEN:
1437                 conn->State = TCP_ST_FIN_WAIT1;
1438                 while( conn->State == TCP_ST_FIN_WAIT1 )        Threads_Yield();
1439                 break;
1440         default:
1441                 Log_Warning("TCP", "Unhandled connection state %i in TCP_Client_Close",
1442                         conn->State);
1443                 break;
1444         }
1445
1446         Time_RemoveTimer(conn->DeferredACKTimer);
1447         Time_FreeTimer(conn->DeferredACKTimer);
1448         free(conn);
1449
1450         LEAVE('-');
1451 }
1452
1453 /**
1454  * \brief Checks if a value is between two others (after taking into account wrapping)
1455  */
1456 int WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue)
1457 {
1458         if( MaxValue < 0xFFFFFFFF )
1459         {
1460                 Lower %= MaxValue + 1;
1461                 Value %= MaxValue + 1;
1462                 Higher %= MaxValue + 1;
1463         }
1464
1465         // Simple Case, no wrap ?
1466         //       Lower Value Higher
1467         // | ... + ... + ... + ... |
1468
1469         if( Lower < Higher ) {
1470                 return Lower < Value && Value < Higher;
1471         }
1472         // Higher has wrapped below lower
1473
1474         // Value > Lower ?
1475         //       Higher Lower Value
1476         // | ... +  ... + ... + ... |
1477         if( Value > Lower ) {
1478                 return 1;
1479         }
1480
1481         // Value < Higher ?
1482         //       Value Higher Lower
1483         // | ... + ... +  ... + ... |
1484         if( Value < Higher ) {
1485                 return 1;
1486         }
1487
1488         return 0;
1489 }