KernelLand/Modules/IPStack/tcp.c

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