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 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
  56 // === TEMPLATES ===
  57 tSocketFile     gTCP_ServerFile = {NULL, "tcps", TCP_Server_Init};
  58 tSocketFile     gTCP_ClientFile = {NULL, "tcpc", TCP_Client_Init};
  59 tVFS_NodeType   gTCP_ServerNodeType = {
  60         .TypeName = "TCP Server",
  61         .ReadDir = TCP_Server_ReadDir,
  62         .FindDir = TCP_Server_FindDir,
  63         .IOCtl   = TCP_Server_IOCtl,
  64         .Close   = TCP_Server_Close
  65         };
  66 tVFS_NodeType   gTCP_ClientNodeType = {
  67         .TypeName = "TCP Client/Connection",
  68         .Read  = TCP_Client_Read,
  69         .Write = TCP_Client_Write,
  70         .IOCtl = TCP_Client_IOCtl,
  71         .Close = TCP_Client_Close
  72         };
  73
  74 // === GLOBALS ===
  75  int    giTCP_NumHalfopen = 0;
  76 tShortSpinlock  glTCP_Listeners;
  77 tTCPListener    *gTCP_Listeners;
  78 tShortSpinlock  glTCP_OutbountCons;
  79 tTCPConnection  *gTCP_OutbountCons;
  80 Uint32  gaTCP_PortBitmap[0x800];
  81  int    giTCP_NextOutPort = TCP_MIN_DYNPORT;
  82
  83 // === CODE ===
  84 /**
  85  * \brief Initialise the TCP Layer
  86  *
  87  * Registers the client and server files and the GetPacket callback
  88  */
  89 void TCP_Initialise(void)
  90 {
  91         giTCP_NextOutPort += rand()%128;
  92         IPStack_AddFile(&gTCP_ServerFile);
  93         IPStack_AddFile(&gTCP_ClientFile);
  94         IPv4_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  95         IPv6_RegisterCallback(IP4PROT_TCP, TCP_GetPacket);
  96 }
  97
  98 /**
  99  * \brief Sends a packet from the specified connection, calculating the checksums
 100  * \param Conn  Connection
 101  * \param Length        Length of data
 102  * \param Data  Packet data (cast as a TCP Header)
 103  */
 104 void TCP_SendPacket( tTCPConnection *Conn, tTCPHeader *Header, size_t Length, const void *Data )
 105 {
 106         TCP_int_SendPacket(Conn->Interface, &Conn->RemoteIP, Header, Length, Data);
 107 }
 108
 109 void TCP_int_SendPacket(tInterface *Interface, const void *Dest, tTCPHeader *Header, size_t Length, const void *Data )
 110 {
 111         tIPStackBuffer  *buffer;
 112         Uint16  checksum[3];
 113          int    packlen = sizeof(*Header) + Length;
 114
 115         buffer = IPStack_Buffer_CreateBuffer(2 + IPV4_BUFFERS);
 116         if( Data && Length )
 117                 IPStack_Buffer_AppendSubBuffer(buffer, Length, 0, Data, NULL, NULL);
 118         IPStack_Buffer_AppendSubBuffer(buffer, sizeof(*Header), 0, Header, NULL, NULL);
 119
 120         LOG("Sending %i+%i to %s:%i", sizeof(*Header), Length,
 121                 IPStack_PrintAddress(Interface->Type, Dest),
 122                 ntohs(Header->RemotePort)
 123                 );
 124
 125         Header->Checksum = 0;
 126         checksum[1] = htons( ~IPv4_Checksum(Header, sizeof(tTCPHeader)) );
 127         checksum[2] = htons( ~IPv4_Checksum(Data, Length) );
 128
 129         // TODO: Fragment packet
 130
 131         switch( Interface->Type )
 132         {
 133         case 4:
 134                 // Get IPv4 pseudo-header checksum
 135                 {
 136                         Uint32  buf[3];
 137                         buf[0] = ((tIPv4*)Interface->Address)->L;
 138                         buf[1] = ((tIPv4*)Dest)->L;
 139                         buf[2] = htonl( (packlen) | (IP4PROT_TCP<<16) | (0<<24) );
 140                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 141                 }
 142                 // - Combine checksums
 143                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );
 144                 IPv4_SendPacket(Interface, *(tIPv4*)Dest, IP4PROT_TCP, 0, buffer);
 145                 break;
 146
 147         case 6:
 148                 // Append IPv6 Pseudo Header
 149                 {
 150                         Uint32  buf[4+4+1+1];
 151                         memcpy(buf, Interface->Address, 16);
 152                         memcpy(&buf[4], Dest, 16);
 153                         buf[8] = htonl(packlen);
 154                         buf[9] = htonl(IP4PROT_TCP);
 155                         checksum[0] = htons( ~IPv4_Checksum(buf, sizeof(buf)) );        // Partial checksum
 156                 }
 157                 Header->Checksum = htons( IPv4_Checksum(checksum, sizeof(checksum)) );  // Combine the two
 158                 IPv6_SendPacket(Interface, *(tIPv6*)Dest, IP4PROT_TCP, buffer);
 159                 break;
 160         }
 161 }
 162
 163 void TCP_int_SendRSTTo(tInterface *Interface, void *Address, size_t Length, const tTCPHeader *Header)
 164 {
 165         tTCPHeader      out_hdr = {0};
 166
 167         out_hdr.DataOffset = (sizeof(out_hdr)/4) << 4;
 168         out_hdr.DestPort = Header->SourcePort;
 169         out_hdr.SourcePort = Header->DestPort;
 170
 171         size_t  data_len = Length - (Header->DataOffset>>4)*4;
 172         out_hdr.AcknowlegementNumber = htonl( ntohl(Header->SequenceNumber) + data_len );
 173         if( Header->Flags & TCP_FLAG_ACK ) {
 174                 out_hdr.Flags = TCP_FLAG_RST;
 175                 out_hdr.SequenceNumber = Header->AcknowlegementNumber;
 176         }
 177         else {
 178                 out_hdr.Flags = TCP_FLAG_RST|TCP_FLAG_ACK;
 179                 out_hdr.SequenceNumber = 0;
 180         }
 181         TCP_int_SendPacket(Interface, Address, &out_hdr, 0, NULL);
 182 }
 183
 184 /**
 185  * \brief Handles a packet from the IP Layer
 186  * \param Interface     Interface the packet arrived from
 187  * \param Address       Pointer to the addres structure
 188  * \param Length        Size of packet in bytes
 189  * \param Buffer        Packet data
 190  */
 191 void TCP_GetPacket(tInterface *Interface, void *Address, int Length, void *Buffer)
 192 {
 193         tTCPHeader      *hdr = Buffer;
 194
 195         #if TCP_DEBUG
 196         Log_Log("TCP", "TCP_GetPacket: <Local>:%i from [%s]:%i, Flags = %s%s%s%s%s%s%s%s",
 197                 ntohs(hdr->DestPort),
 198                 IPStack_PrintAddress(Interface->Type, Address),
 199                 ntohs(hdr->SourcePort),
 200                 (hdr->Flags & TCP_FLAG_CWR) ? "CWR " : "",
 201                 (hdr->Flags & TCP_FLAG_ECE) ? "ECE " : "",
 202                 (hdr->Flags & TCP_FLAG_URG) ? "URG " : "",
 203                 (hdr->Flags & TCP_FLAG_ACK) ? "ACK " : "",
 204                 (hdr->Flags & TCP_FLAG_PSH) ? "PSH " : "",
 205                 (hdr->Flags & TCP_FLAG_RST) ? "RST " : "",
 206                 (hdr->Flags & TCP_FLAG_SYN) ? "SYN " : "",
 207                 (hdr->Flags & TCP_FLAG_FIN) ? "FIN " : ""
 208                 );
 209         #endif
 210
 211         if( Length > (hdr->DataOffset >> 4)*4 )
 212         {
 213                 LOG("SequenceNumber = 0x%x", ntohl(hdr->SequenceNumber));
 214 #if HEXDUMP_INCOMING
 215                 Debug_HexDump(
 216                         "TCP_GetPacket: Packet Data = ",
 217                         (Uint8*)hdr + (hdr->DataOffset >> 4)*4,
 218                         Length - (hdr->DataOffset >> 4)*4
 219                         );
 220 #endif
 221         }
 222
 223         // Check Servers
 224         for( tTCPListener *srv = gTCP_Listeners; srv; srv = srv->Next )
 225         {
 226                 // Check if the server is active
 227                 if(srv->Port == 0)      continue;
 228                 // Check the interface
 229                 if(srv->Interface && srv->Interface != Interface)       continue;
 230                 // Check the destination port
 231                 if(srv->Port != htons(hdr->DestPort))   continue;
 232
 233                 Log_Log("TCP", "TCP_GetPacket: Matches server %p", srv);
 234                 // Is this in an established connection?
 235                 for( tTCPConnection *conn = srv->Connections; conn; conn = conn->Next )
 236                 {
 237                         // Check that it is coming in on the same interface
 238                         if(conn->Interface != Interface)        continue;
 239
 240                         // Check Source Port
 241                         Log_Log("TCP", "TCP_GetPacket: conn->RemotePort(%i) == hdr->SourcePort(%i)",
 242                                 conn->RemotePort, ntohs(hdr->SourcePort));
 243                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 244
 245                         // Check Source IP
 246                         Log_Debug("TCP", "TCP_GetPacket: conn->RemoteIP(%s)",
 247                                 IPStack_PrintAddress(conn->Interface->Type, &conn->RemoteIP));
 248                         Log_Debug("TCP", "                == Address(%s)",
 249                                 IPStack_PrintAddress(conn->Interface->Type, Address));
 250                         if( IPStack_CompareAddress(conn->Interface->Type, &conn->RemoteIP, Address, -1) == 0 )
 251                                 continue ;
 252
 253                         Log_Log("TCP", "TCP_GetPacket: Matches connection %p", conn);
 254                         // We have a response!
 255                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 256
 257                         return;
 258                 }
 259
 260
 261                 if( hdr->Flags & TCP_FLAG_RST ) {
 262                         LOG("RST, ignore");
 263                         return ;
 264                 }
 265                 else if( hdr->Flags & TCP_FLAG_ACK ) {
 266                         LOG("ACK, send RST");
 267                         TCP_int_SendRSTTo(Interface, Address, Length, hdr);
 268                         return ;
 269                 }
 270                 else if( !(hdr->Flags & TCP_FLAG_SYN) ) {
 271                         LOG("Other, ignore");
 272                         return ;
 273                 }
 274                 Log_Log("TCP", "TCP_GetPacket: Opening Connection");
 275
 276                 // TODO: Check for halfopen max
 277
 278                 tTCPConnection *conn = TCP_int_CreateConnection(Interface, TCP_ST_SYN_RCVD);
 279                 conn->LocalPort = srv->Port;
 280                 conn->RemotePort = ntohs(hdr->SourcePort);
 281
 282                 switch(Interface->Type)
 283                 {
 284                 case 4: conn->RemoteIP.v4 = *(tIPv4*)Address;   break;
 285                 case 6: conn->RemoteIP.v6 = *(tIPv6*)Address;   break;
 286                 default:        ASSERTC(Interface->Type,==,4);  return;
 287                 }
 288
 289                 conn->NextSequenceRcv = ntohl( hdr->SequenceNumber ) + 1;
 290                 conn->HighestSequenceRcvd = conn->NextSequenceRcv;
 291                 conn->NextSequenceSend = rand();
 292
 293                 conn->Node.ImplInt = srv->NextID ++;
 294
 295                 // Hmm... Theoretically, this lock will never have to wait,
 296                 // as the interface is locked to the watching thread, and this
 297                 // runs in the watching thread. But, it's a good idea to have
 298                 // it, just in case
 299                 // Oh, wait, there is a case where a wildcard can be used
 300                 // (srv->Interface == NULL) so having the lock is a good idea
 301                 SHORTLOCK(&srv->lConnections);
 302                 conn->Server = srv;
 303                 conn->Prev = srv->ConnectionsTail;
 304                 if(srv->Connections) {
 305                         ASSERT(srv->ConnectionsTail);
 306                         srv->ConnectionsTail->Next = conn;
 307                 }
 308                 else {
 309                         ASSERT(!srv->ConnectionsTail);
 310                         srv->Connections = conn;
 311                 }
 312                 srv->ConnectionsTail = conn;
 313                 if(!srv->NewConnections)
 314                         srv->NewConnections = conn;
 315                 VFS_MarkAvaliable( &srv->Node, 1 );
 316                 SHORTREL(&srv->lConnections);
 317                 Semaphore_Signal(&srv->WaitingConnections, 1);
 318
 319                 // Send the SYN ACK
 320                 hdr->Flags |= TCP_FLAG_ACK;
 321                 hdr->AcknowlegementNumber = htonl(conn->NextSequenceRcv);
 322                 hdr->SequenceNumber = htonl(conn->NextSequenceSend);
 323                 hdr->DestPort = hdr->SourcePort;
 324                 hdr->SourcePort = htons(srv->Port);
 325                 hdr->DataOffset = (sizeof(tTCPHeader)/4) << 4;
 326                 TCP_SendPacket( conn, hdr, 0, NULL );
 327                 conn->NextSequenceSend ++;
 328                 return ;
 329         }
 330
 331         // Check Open Connections
 332         {
 333                 for( tTCPConnection *conn = gTCP_OutbountCons; conn; conn = conn->Next )
 334                 {
 335                         // Check that it is coming in on the same interface
 336                         if(conn->Interface != Interface)        continue;
 337
 338                         // Check Source Port
 339                         if(conn->RemotePort != ntohs(hdr->SourcePort))  continue;
 340
 341                         // Check Source IP
 342                         if(conn->Interface->Type == 6 && !IP6_EQU(conn->RemoteIP.v6, *(tIPv6*)Address))
 343                                 continue;
 344                         if(conn->Interface->Type == 4 && !IP4_EQU(conn->RemoteIP.v4, *(tIPv4*)Address))
 345                                 continue;
 346
 347                         TCP_INT_HandleConnectionPacket(conn, hdr, Length);
 348                         return ;
 349                 }
 350         }
 351
 352         Log_Log("TCP", "TCP_GetPacket: No Match");
 353         // If not a RST, send a RST
 354         if( !(hdr->Flags & TCP_FLAG_RST) )
 355         {
 356                 TCP_int_SendRSTTo(Interface, Address, Length, hdr);
 357         }
 358 }
 359
 360 /**
 361  * \brief Handles a packet sent to a specific connection
 362  * \param Connection    TCP Connection pointer
 363  * \param Header        TCP Packet pointer
 364  * \param Length        Length of the packet
 365  */
 366 void TCP_INT_HandleConnectionPacket(tTCPConnection *Connection, tTCPHeader *Header, int Length)
 367 {
 368          int    dataLen;
 369         Uint32  sequence_num;
 370
 371         // Silently drop once finished
 372         // TODO: Check if this needs to be here
 373         if( Connection->State == TCP_ST_FINISHED ) {
 374                 Log_Log("TCP", "Packet ignored - connection finnished");
 375                 return ;
 376         }
 377
 378         // Syncronise sequence values
 379         if(Header->Flags & TCP_FLAG_SYN) {
 380                 // TODO: What if the packet also has data?
 381                 if( Connection->LastACKSequence != Connection->NextSequenceRcv )
 382                         TCP_INT_SendACK(Connection, "SYN");
 383                 Connection->NextSequenceRcv = ntohl(Header->SequenceNumber);
 384                 // TODO: Process HighestSequenceRcvd
 385                 // HACK!
 386                 if( Connection->HighestSequenceRcvd == 0 )
 387                         Connection->HighestSequenceRcvd = Connection->NextSequenceRcv;
 388                 Connection->LastACKSequence = Connection->NextSequenceRcv;
 389         }
 390
 391         // Ackowledge a sent packet
 392         if(Header->Flags & TCP_FLAG_ACK) {
 393                 // TODO: Process an ACKed Packet
 394                 LOG("Conn %p, Sent packet 0x%x ACKed", Connection, Header->AcknowlegementNumber);
 395         }
 396
 397         // Get length of data
 398         dataLen = Length - (Header->DataOffset>>4)*4;
 399         LOG("dataLen = %i", dataLen);
 400         #if TCP_DEBUG
 401         Log_Debug("TCP", "State %i, dataLen = %x", Connection->State, dataLen);
 402         #endif
 403
 404         //
 405         // State Machine
 406         //
 407         switch( Connection->State )
 408         {
 409         // Pre-init connection?
 410         case TCP_ST_CLOSED:
 411                 Log_Log("TCP", "Packets to a closed connection?!");
 412                 break;
 413
 414         // --- Init States ---
 415         // SYN sent, expecting SYN-ACK Connection Opening
 416         case TCP_ST_SYN_SENT:
 417                 if( Header->Flags & TCP_FLAG_SYN )
 418                 {
 419                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 420                                 Connection->HighestSequenceRcvd ++;
 421                         Connection->NextSequenceRcv ++;
 422
 423                         if( Header->Flags & TCP_FLAG_ACK )
 424                         {
 425                                 Log_Log("TCP", "ACKing SYN-ACK");
 426                                 Connection->State = TCP_ST_OPEN;
 427                                 VFS_MarkFull(&Connection->Node, 0);
 428                                 TCP_INT_SendACK(Connection, "SYN-ACK");
 429                         }
 430                         else
 431                         {
 432                                 Log_Log("TCP", "ACKing SYN");
 433                                 Connection->State = TCP_ST_SYN_RCVD;
 434                                 TCP_INT_SendACK(Connection, "SYN");
 435                         }
 436                 }
 437                 break;
 438
 439         // SYN-ACK sent, expecting ACK
 440         case TCP_ST_SYN_RCVD:
 441                 if( Header->Flags & TCP_FLAG_ACK )
 442                 {
 443                         // TODO: Handle max half-open limit
 444                         Log_Log("TCP", "Connection fully opened");
 445                         Connection->State = TCP_ST_OPEN;
 446                         VFS_MarkFull(&Connection->Node, 0);
 447                 }
 448                 break;
 449
 450         // --- Established State ---
 451         case TCP_ST_OPEN:
 452                 // - Handle State changes
 453                 //
 454                 if( Header->Flags & TCP_FLAG_FIN ) {
 455                         Log_Log("TCP", "Conn %p closed, recieved FIN", Connection);
 456                         VFS_MarkError(&Connection->Node, 1);
 457                         Connection->NextSequenceRcv ++;
 458                         TCP_INT_SendACK(Connection, "FIN Received");
 459                         Connection->State = TCP_ST_CLOSE_WAIT;
 460                         // CLOSE WAIT requires the client to close
 461                         return ;
 462                 }
 463
 464                 // Check for an empty packet
 465                 if(dataLen == 0) {
 466                         if( Header->Flags == TCP_FLAG_ACK )
 467                         {
 468                                 Log_Log("TCP", "ACK only packet");
 469                                 return ;
 470                         }
 471                         // TODO: Is this right? (empty packet counts as one byte)
 472                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 473                                 Connection->HighestSequenceRcvd ++;
 474                         Connection->NextSequenceRcv ++;
 475                         Log_Log("TCP", "Empty Packet, inc and ACK the current sequence number");
 476                         TCP_INT_SendACK(Connection, "Empty");
 477                         return ;
 478                 }
 479
 480                 // NOTES:
 481                 // Flags
 482                 //    PSH - Has Data?
 483                 // /NOTES
 484
 485                 sequence_num = ntohl(Header->SequenceNumber);
 486
 487                 LOG("0x%08x <= 0x%08x < 0x%08x",
 488                         Connection->NextSequenceRcv,
 489                         ntohl(Header->SequenceNumber),
 490                         Connection->NextSequenceRcv + TCP_WINDOW_SIZE
 491                         );
 492
 493                 // Is this packet the next expected packet?
 494                 if( sequence_num == Connection->NextSequenceRcv )
 495                 {
 496                          int    rv;
 497                         // Ooh, Goodie! Add it to the recieved list
 498                         rv = TCP_INT_AppendRecieved(Connection,
 499                                 (Uint8*)Header + (Header->DataOffset>>4)*4,
 500                                 dataLen
 501                                 );
 502                         if(rv != 0) {
 503                                 Log_Notice("TCP", "TCP_INT_AppendRecieved rv %i", rv);
 504                                 break;
 505                         }
 506                         LOG("0x%08x += %i", Connection->NextSequenceRcv, dataLen);
 507                         if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 508                                 Connection->HighestSequenceRcvd += dataLen;
 509                         Connection->NextSequenceRcv += dataLen;
 510
 511                         // TODO: This should be moved out of the watcher thread,
 512                         // so that a single lost packet on one connection doesn't cause
 513                         // all connections on the interface to lag.
 514                         // - Meh, no real issue, as the cache shouldn't be that large
 515                         TCP_INT_UpdateRecievedFromFuture(Connection);
 516
 517                         #if 1
 518                         // - Only send an ACK if we've had a burst
 519                         if( Connection->NextSequenceRcv > (Uint32)(TCP_DACK_THRESHOLD + Connection->LastACKSequence) )
 520                         {
 521                                 TCP_INT_SendACK(Connection, "DACK Burst");
 522                                 // - Extend TCP deferred ACK timer
 523                                 Time_RemoveTimer(Connection->DeferredACKTimer);
 524                         }
 525                         // - Schedule the deferred ACK timer (if already scheduled, this is a NOP)
 526                         Time_ScheduleTimer(Connection->DeferredACKTimer, TCP_DACK_TIMEOUT);
 527                         #else
 528                         TCP_INT_SendACK(Connection, "RX");
 529                         #endif
 530                 }
 531                 // Check if the packet is in window
 532                 else if( WrapBetween(Connection->NextSequenceRcv, sequence_num,
 533                                 Connection->NextSequenceRcv+TCP_WINDOW_SIZE, 0xFFFFFFFF) )
 534                 {
 535                         Uint8   *dataptr = (Uint8*)Header + (Header->DataOffset>>4)*4;
 536                         #if CACHE_FUTURE_PACKETS_IN_BYTES
 537                         Uint32  index;
 538
 539                         index = sequence_num % TCP_WINDOW_SIZE;
 540                         for( int i = 0; i < dataLen; i ++ )
 541                         {
 542                                 Connection->FuturePacketValidBytes[index/8] |= 1 << (index%8);
 543                                 Connection->FuturePacketData[index] = dataptr[i];
 544                                 // Do a wrap increment
 545                                 index ++;
 546                                 if(index == TCP_WINDOW_SIZE)    index = 0;
 547                         }
 548                         #else
 549                         tTCPStoredPacket        *pkt, *tmp, *prev = NULL;
 550
 551                         // Allocate and fill cached packet
 552                         pkt = malloc( sizeof(tTCPStoredPacket) + dataLen );
 553                         pkt->Next = NULL;
 554                         pkt->Sequence = ntohl(Header->SequenceNumber);
 555                         pkt->Length = dataLen;
 556                         memcpy(pkt->Data, dataptr, dataLen);
 557
 558                         Log_Log("TCP", "We missed a packet, caching",
 559                                 pkt->Sequence, Connection->NextSequenceRcv);
 560
 561                         // No? Well, let's cache it and look at it later
 562                         SHORTLOCK( &Connection->lFuturePackets );
 563                         for(tmp = Connection->FuturePackets;
 564                                 tmp;
 565                                 prev = tmp, tmp = tmp->Next)
 566                         {
 567                                 if(tmp->Sequence >= pkt->Sequence)      break;
 568                         }
 569
 570                         // Add if before first, or sequences don't match
 571                         if( !tmp || tmp->Sequence != pkt->Sequence )
 572                         {
 573                                 if(prev)
 574                                         prev->Next = pkt;
 575                                 else
 576                                         Connection->FuturePackets = pkt;
 577                                 pkt->Next = tmp;
 578                         }
 579                         // Replace if larger
 580                         else if(pkt->Length > tmp->Length)
 581                         {
 582                                 if(prev)
 583                                         prev->Next = pkt;
 584                                 pkt->Next = tmp->Next;
 585                                 free(tmp);
 586                         }
 587                         else
 588                         {
 589                                 free(pkt);      // TODO: Find some way to remove this
 590                         }
 591                         SHORTREL( &Connection->lFuturePackets );
 592                         #endif
 593                 }
 594                 // Badly out of sequence packet
 595                 else
 596                 {
 597                         Log_Log("TCP", "Fully out of sequence packet (0x%08x not between 0x%08x and 0x%08x), dropped",
 598                                 sequence_num, Connection->NextSequenceRcv, Connection->NextSequenceRcv+TCP_WINDOW_SIZE);
 599                         // Spec says we should send an empty ACK with the current state
 600                         TCP_INT_SendACK(Connection, "Bad Seq");
 601                 }
 602                 break;
 603
 604         // --- Remote close states
 605         case TCP_ST_CLOSE_WAIT:
 606
 607                 // Ignore everything, CLOSE_WAIT is terminated by the client
 608                 Log_Debug("TCP", "CLOSE WAIT - Ignoring packets");
 609
 610                 break;
 611
 612         // LAST-ACK - Waiting for the ACK of FIN (from CLOSE WAIT)
 613         case TCP_ST_LAST_ACK:
 614                 if( Header->Flags & TCP_FLAG_ACK )
 615                 {
 616                         Connection->State = TCP_ST_FINISHED;    // Connection completed
 617                         Log_Log("TCP", "LAST-ACK to CLOSED - Connection remote closed");
 618                         TCP_int_FreeTCB(Connection);
 619                 }
 620                 break;
 621
 622         // --- Local close States
 623         case TCP_ST_FIN_WAIT1:
 624                 if( Header->Flags & TCP_FLAG_FIN )
 625                 {
 626                         Connection->State = TCP_ST_CLOSING;
 627                         Log_Debug("TCP", "Conn %p closed, sent FIN and recieved FIN", Connection);
 628                         VFS_MarkError(&Connection->Node, 1);
 629
 630                         TCP_INT_SendACK(Connection, "FINWAIT-1 FIN");
 631                         break ;
 632                 }
 633
 634                 // TODO: Make sure that the packet is actually ACKing the FIN
 635                 if( Header->Flags & TCP_FLAG_ACK )
 636                 {
 637                         Connection->State = TCP_ST_FIN_WAIT2;
 638                         Log_Debug("TCP", "Conn %p closed, sent FIN ACKed", Connection);
 639                         VFS_MarkError(&Connection->Node, 1);
 640                         return ;
 641                 }
 642                 break;
 643
 644         case TCP_ST_FIN_WAIT2:
 645                 if( Header->Flags & TCP_FLAG_FIN )
 646                 {
 647                         Connection->State = TCP_ST_TIME_WAIT;
 648                         Log_Debug("TCP", "Conn %p FINWAIT-2 -> TIME WAIT", Connection);
 649                         TCP_INT_SendACK(Connection, "FINWAIT-2 FIN");
 650                 }
 651                 break;
 652
 653         case TCP_ST_CLOSING:
 654                 // TODO: Make sure that the packet is actually ACKing the FIN
 655                 if( Header->Flags & TCP_FLAG_ACK )
 656                 {
 657                         Connection->State = TCP_ST_TIME_WAIT;
 658                         Log_Debug("TCP", "Conn %p CLOSING -> TIME WAIT", Connection);
 659                         VFS_MarkError(&Connection->Node, 1);
 660                         return ;
 661                 }
 662                 break;
 663
 664         // --- Closed (or near closed) states) ---
 665         case TCP_ST_TIME_WAIT:
 666                 Log_Log("TCP", "Packets on Time-Wait, ignored");
 667                 break;
 668
 669         case TCP_ST_FINISHED:
 670                 Log_Log("TCP", "Packets when CLOSED, ignoring");
 671                 break;
 672
 673         //default:
 674         //      Log_Warning("TCP", "Unhandled TCP state %i", Connection->State);
 675         //      break;
 676         }
 677
 678 }
 679
 680 /**
 681  * \brief Appends a packet to the recieved list
 682  * \param Connection    Connection structure
 683  * \param Data  Packet contents
 684  * \param Length        Length of \a Data
 685  */
 686 int TCP_INT_AppendRecieved(tTCPConnection *Connection, const void *Data, size_t Length)
 687 {
 688         Mutex_Acquire( &Connection->lRecievedPackets );
 689
 690         if(Connection->RecievedBuffer->Length + Length > Connection->RecievedBuffer->Space )
 691         {
 692                 VFS_MarkAvaliable(&Connection->Node, 1);
 693                 Log_Error("TCP", "Buffer filled, packet dropped (:%i) - %i + %i > %i",
 694                         Connection->LocalPort, Connection->RecievedBuffer->Length, Length,
 695                         Connection->RecievedBuffer->Space
 696                         );
 697                 Mutex_Release( &Connection->lRecievedPackets );
 698                 return 1;
 699         }
 700
 701         RingBuffer_Write( Connection->RecievedBuffer, Data, Length );
 702
 703         VFS_MarkAvaliable(&Connection->Node, 1);
 704
 705         Mutex_Release( &Connection->lRecievedPackets );
 706         return 0;
 707 }
 708
 709 /**
 710  * \brief Updates the connections recieved list from the future list
 711  * \param Connection    Connection structure
 712  *
 713  * Updates the recieved packets list with packets from the future (out
 714  * of order) packets list that are now able to be added in direct
 715  * sequence.
 716  */
 717 void TCP_INT_UpdateRecievedFromFuture(tTCPConnection *Connection)
 718 {
 719         #if CACHE_FUTURE_PACKETS_IN_BYTES
 720         // Calculate length of contiguous bytes
 721          int    length = Connection->HighestSequenceRcvd - Connection->NextSequenceRcv;
 722         Uint32  index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 723         LOG("length=%i, index=%i", length, index);
 724         for( int i = 0; i < length; i ++ )
 725         {
 726                  int    bit = index % 8;
 727                 Uint8   bitfield_byte = Connection->FuturePacketValidBytes[index / 8];
 728                 if( (bitfield_byte & (1 << bit)) == 0 ) {
 729                         length = i;
 730                         break;
 731                 }
 732
 733                 if( bitfield_byte == 0xFF ) {
 734                          int    inc = 8 - bit;
 735                         i += inc - 1;
 736                         index += inc;
 737                 }
 738                 else {
 739                         index ++;
 740                 }
 741                 if(index > TCP_WINDOW_SIZE)
 742                         index -= TCP_WINDOW_SIZE;
 743         }
 744
 745         index = Connection->NextSequenceRcv % TCP_WINDOW_SIZE;
 746
 747         // Write data to to the ring buffer
 748         if( TCP_WINDOW_SIZE - index > length )
 749         {
 750                 // Simple case
 751                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, length );
 752         }
 753         else
 754         {
 755                  int    endLen = TCP_WINDOW_SIZE - index;
 756                 // 2-part case
 757                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData + index, endLen );
 758                 RingBuffer_Write( Connection->RecievedBuffer, Connection->FuturePacketData, endLen - length );
 759         }
 760
 761         // Mark (now saved) bytes as invalid
 762         // - Align index
 763         while(index % 8 && length > 0)
 764         {
 765                 Connection->FuturePacketData[index] = 0;
 766                 Connection->FuturePacketValidBytes[index/8] &= ~(1 << (index%8));
 767                 index ++;
 768                 if(index > TCP_WINDOW_SIZE)
 769                         index -= TCP_WINDOW_SIZE;
 770                 length --;
 771         }
 772         while( length > 7 )
 773         {
 774                 Connection->FuturePacketData[index] = 0;
 775                 Connection->FuturePacketValidBytes[index/8] = 0;
 776                 length -= 8;
 777                 index += 8;
 778                 if(index > TCP_WINDOW_SIZE)
 779                         index -= TCP_WINDOW_SIZE;
 780         }
 781         while(length)
 782         {
 783                 Connection->FuturePacketData[index] = 0;
 784                 Connection->FuturePacketData[index/8] &= ~(1 << (index%8));
 785                 index ++;
 786                 if(index > TCP_WINDOW_SIZE)
 787                         index -= TCP_WINDOW_SIZE;
 788                 length --;
 789         }
 790
 791         #else
 792         tTCPStoredPacket        *pkt;
 793         for(;;)
 794         {
 795                 SHORTLOCK( &Connection->lFuturePackets );
 796
 797                 // Clear out duplicates from cache
 798                 // - If a packet has just been recieved, and it is expected, then
 799                 //   (since NextSequenceRcv = rcvd->Sequence + rcvd->Length) all
 800                 //   packets in cache that are smaller than the next expected
 801                 //   are now defunct.
 802                 pkt = Connection->FuturePackets;
 803                 while(pkt && pkt->Sequence < Connection->NextSequenceRcv)
 804                 {
 805                         tTCPStoredPacket        *next = pkt->Next;
 806                         free(pkt);
 807                         pkt = next;
 808                 }
 809
 810                 // If there's no packets left in cache, stop looking
 811                 if(!pkt || pkt->Sequence > Connection->NextSequenceRcv) {
 812                         SHORTREL( &Connection->lFuturePackets );
 813                         return;
 814                 }
 815
 816                 // Delete packet from future list
 817                 Connection->FuturePackets = pkt->Next;
 818
 819                 // Release list
 820                 SHORTREL( &Connection->lFuturePackets );
 821
 822                 // Looks like we found one
 823                 TCP_INT_AppendRecieved(Connection, pkt->Data, pkt->Length);
 824                 if( Connection->HighestSequenceRcvd == Connection->NextSequenceRcv )
 825                         Connection->HighestSequenceRcvd += pkt->Length;
 826                 Connection->NextSequenceRcv += pkt->Length;
 827                 free(pkt);
 828         }
 829         #endif
 830 }
 831
 832 void TCP_int_SendDelayedACK(void *ConnPtr)
 833 {
 834         TCP_INT_SendACK(ConnPtr, "DACK Timeout");
 835 }
 836
 837 void TCP_INT_SendACK(tTCPConnection *Connection, const char *Reason)
 838 {
 839         tTCPHeader      hdr;
 840         // ACK Packet
 841         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
 842         hdr.DestPort = htons(Connection->RemotePort);
 843         hdr.SourcePort = htons(Connection->LocalPort);
 844         hdr.AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
 845         hdr.SequenceNumber = htonl(Connection->NextSequenceSend);
 846         hdr.WindowSize = htons(TCP_WINDOW_SIZE);
 847         hdr.Flags = TCP_FLAG_ACK;       // TODO: Determine if SYN is wanted too
 848         hdr.Checksum = 0;       // TODO: Checksum
 849         hdr.UrgentPointer = 0;
 850         Log_Debug("TCP", "Sending ACK for 0x%08x (%s)", Connection->NextSequenceRcv, Reason);
 851         TCP_SendPacket( Connection, &hdr, 0, NULL );
 852         //Connection->NextSequenceSend ++;
 853         Connection->LastACKSequence = Connection->NextSequenceRcv;
 854 }
 855
 856 /**
 857  * \fn Uint16 TCP_GetUnusedPort()
 858  * \brief Gets an unused port and allocates it
 859  */
 860 Uint16 TCP_GetUnusedPort()
 861 {
 862         Uint16  ret;
 863
 864         // Get Next outbound port
 865         ret = giTCP_NextOutPort++;
 866         while( gaTCP_PortBitmap[ret/32] & (1UL << (ret%32)) )
 867         {
 868                 ret ++;
 869                 giTCP_NextOutPort++;
 870                 if(giTCP_NextOutPort == 0x10000) {
 871                         ret = giTCP_NextOutPort = TCP_MIN_DYNPORT;
 872                 }
 873         }
 874
 875         // Mark the new port as used
 876         gaTCP_PortBitmap[ret/32] |= 1 << (ret%32);
 877
 878         return ret;
 879 }
 880
 881 /**
 882  * \fn int TCP_AllocatePort(Uint16 Port)
 883  * \brief Marks a port as used
 884  */
 885 int TCP_AllocatePort(Uint16 Port)
 886 {
 887         // Check if the port has already been allocated
 888         if( gaTCP_PortBitmap[Port/32] & (1 << (Port%32)) )
 889                 return 0;
 890
 891         // Allocate
 892         gaTCP_PortBitmap[Port/32] |= 1 << (Port%32);
 893
 894         return 1;
 895 }
 896
 897 /**
 898  * \fn int TCP_DeallocatePort(Uint16 Port)
 899  * \brief Marks a port as unused
 900  */
 901 int TCP_DeallocatePort(Uint16 Port)
 902 {
 903         // Check if the port has already been allocated
 904         if( !(gaTCP_PortBitmap[Port/32] & (1 << (Port%32))) )
 905                 return 0;
 906
 907         // Allocate
 908         gaTCP_PortBitmap[Port/32] &= ~(1 << (Port%32));
 909
 910         return 1;
 911 }
 912
 913 tTCPConnection *TCP_int_CreateConnection(tInterface *Interface, enum eTCPConnectionState State)
 914 {
 915         tTCPConnection  *conn = calloc( sizeof(tTCPConnection) + TCP_WINDOW_SIZE + TCP_WINDOW_SIZE/8, 1 );
 916
 917         conn->State = State;
 918         conn->Interface = Interface;
 919         conn->LocalPort = -1;
 920         conn->RemotePort = -1;
 921
 922         conn->Node.ReferenceCount = 1;
 923         conn->Node.ImplPtr = conn;
 924         conn->Node.NumACLs = 1;
 925         conn->Node.ACLs = &gVFS_ACL_EveryoneRW;
 926         conn->Node.Type = &gTCP_ClientNodeType;
 927         conn->Node.BufferFull = 1;      // Cleared when connection opens
 928
 929         conn->RecievedBuffer = RingBuffer_Create( TCP_RECIEVE_BUFFER_SIZE );
 930         #if 0
 931         conn->SentBuffer = RingBuffer_Create( TCP_SEND_BUFFER_SIZE );
 932         Semaphore_Init(conn->SentBufferSpace, 0, TCP_SEND_BUFFER_SIZE, "TCP SentBuffer", conn->Name);
 933         #endif
 934
 935         #if CACHE_FUTURE_PACKETS_IN_BYTES
 936         // Future recieved data (ahead of the expected sequence number)
 937         conn->FuturePacketData = (Uint8*)conn + sizeof(tTCPConnection);
 938         conn->FuturePacketValidBytes = conn->FuturePacketData + TCP_WINDOW_SIZE;
 939         #endif
 940
 941         conn->DeferredACKTimer = Time_AllocateTimer( TCP_int_SendDelayedACK, conn);
 942         return conn;
 943 }
 944
 945 void TCP_int_FreeTCB(tTCPConnection *Connection)
 946 {
 947         ASSERTC(Connection->State, ==, TCP_ST_FINISHED);
 948         ASSERTC(Connection->Node.ReferenceCount, ==, 0);
 949
 950         if( Connection->Server )
 951         {
 952                 tTCPListener    *srv = Connection->Server;
 953                 SHORTLOCK(&srv->lConnections);
 954                 if(Connection->Prev)
 955                         Connection->Prev->Next = Connection->Next;
 956                 else
 957                         srv->Connections = Connection->Next;
 958                 if(Connection->Next)
 959                         Connection->Next->Prev = Connection->Prev;
 960                 else {
 961                         ASSERT(srv->ConnectionsTail == Connection);
 962                         srv->ConnectionsTail = Connection->Prev;
 963                 }
 964                 SHORTREL(&srv->lConnections);
 965         }
 966         else
 967         {
 968                 SHORTLOCK(&glTCP_OutbountCons);
 969                 if(Connection->Prev)
 970                         Connection->Prev->Next = Connection->Next;
 971                 else
 972                         gTCP_OutbountCons = Connection->Next;
 973                 if(Connection->Next)
 974                         Connection->Next->Prev = Connection->Prev;
 975                 else
 976                         ;
 977                 SHORTREL(&glTCP_OutbountCons);
 978         }
 979
 980         RingBuffer_Free(Connection->RecievedBuffer);
 981         Time_FreeTimer(Connection->DeferredACKTimer);
 982         // TODO: Force VFS to close handles? (they should all be closed);
 983         free(Connection);
 984 }
 985
 986 // --- Server
 987 tVFS_Node *TCP_Server_Init(tInterface *Interface)
 988 {
 989         tTCPListener    *srv;
 990
 991         srv = calloc( 1, sizeof(tTCPListener) );
 992
 993         if( srv == NULL ) {
 994                 Log_Warning("TCP", "malloc failed for listener (%i) bytes", sizeof(tTCPListener));
 995                 return NULL;
 996         }
 997
 998         srv->Interface = Interface;
 999         srv->Port = 0;
1000         srv->NextID = 0;
1001         srv->Connections = NULL;
1002         srv->ConnectionsTail = NULL;
1003         srv->NewConnections = NULL;
1004         srv->Next = NULL;
1005         srv->Node.Flags = VFS_FFLAG_DIRECTORY;
1006         srv->Node.Size = -1;
1007         srv->Node.ImplPtr = srv;
1008         srv->Node.NumACLs = 1;
1009         srv->Node.ACLs = &gVFS_ACL_EveryoneRW;
1010         srv->Node.Type = &gTCP_ServerNodeType;
1011
1012         SHORTLOCK(&glTCP_Listeners);
1013         srv->Next = gTCP_Listeners;
1014         gTCP_Listeners = srv;
1015         SHORTREL(&glTCP_Listeners);
1016
1017         return &srv->Node;
1018 }
1019
1020 /**
1021  * \brief Wait for a new connection and return the connection ID
1022  * \note Blocks until a new connection is made
1023  * \param Node  Server node
1024  * \param Pos   Position (ignored)
1025  */
1026 int TCP_Server_ReadDir(tVFS_Node *Node, int Pos, char Dest[FILENAME_MAX])
1027 {
1028         tTCPListener    *srv = Node->ImplPtr;
1029         tTCPConnection  *conn;
1030
1031         ENTER("pNode iPos", Node, Pos);
1032
1033         Log_Log("TCP", "Thread %i waiting for a connection", Threads_GetTID());
1034         Semaphore_Wait( &srv->WaitingConnections, 1 );
1035
1036         SHORTLOCK(&srv->lConnections);
1037         // Increment the new list (the current connection is still on the
1038         // normal list)
1039         conn = srv->NewConnections;
1040         srv->NewConnections = conn->Next;
1041
1042         if( srv->NewConnections == NULL )
1043                 VFS_MarkAvaliable( Node, 0 );
1044
1045         SHORTREL( &srv->lConnections );
1046
1047         LOG("conn = %p", conn);
1048         LOG("srv->Connections = %p", srv->Connections);
1049         LOG("srv->NewConnections = %p", srv->NewConnections);
1050         LOG("srv->ConnectionsTail = %p", srv->ConnectionsTail);
1051
1052         itoa(Dest, conn->Node.ImplInt, 16, 8, '0');
1053         Log_Log("TCP", "Thread %i got connection '%s'", Threads_GetTID(), Dest);
1054         LEAVE('i', 0);
1055         return 0;
1056 }
1057
1058 /**
1059  * \brief Gets a client connection node
1060  * \param Node  Server node
1061  * \param Name  Hexadecimal ID of the node
1062  */
1063 tVFS_Node *TCP_Server_FindDir(tVFS_Node *Node, const char *Name, Uint Flags)
1064 {
1065         tTCPConnection  *conn;
1066         tTCPListener    *srv = Node->ImplPtr;
1067         char    tmp[9];
1068          int    id = atoi(Name);
1069
1070         ENTER("pNode sName", Node, Name);
1071
1072         // Check for a non-empty name
1073         if( Name[0] )
1074         {
1075                 // Sanity Check
1076                 itoa(tmp, id, 16, 8, '0');
1077                 if(strcmp(tmp, Name) != 0) {
1078                         LOG("'%s' != '%s' (%08x)", Name, tmp, id);
1079                         LEAVE('n');
1080                         return NULL;
1081                 }
1082
1083                 Log_Debug("TCP", "srv->Connections = %p", srv->Connections);
1084                 Log_Debug("TCP", "srv->NewConnections = %p", srv->NewConnections);
1085                 Log_Debug("TCP", "srv->ConnectionsTail = %p", srv->ConnectionsTail);
1086
1087                 // Search
1088                 SHORTLOCK( &srv->lConnections );
1089                 for(conn = srv->Connections;
1090                         conn;
1091                         conn = conn->Next)
1092                 {
1093                         LOG("conn->Node.ImplInt = %i", conn->Node.ImplInt);
1094                         if(conn->Node.ImplInt == id)    break;
1095                 }
1096                 SHORTREL( &srv->lConnections );
1097
1098                 // If not found, ret NULL
1099                 if(!conn) {
1100                         LOG("Connection %i not found", id);
1101                         LEAVE('n');
1102                         return NULL;
1103                 }
1104         }
1105         // Empty Name - Check for a new connection and if it's there, open it
1106         else
1107         {
1108                 SHORTLOCK( &srv->lConnections );
1109                 conn = srv->NewConnections;
1110                 if( conn != NULL )
1111                         srv->NewConnections = conn->Next;
1112                 VFS_MarkAvaliable( Node, srv->NewConnections != NULL );
1113                 SHORTREL( &srv->lConnections );
1114                 if( !conn ) {
1115                         LOG("No new connections");
1116                         LEAVE('n');
1117                         return NULL;
1118                 }
1119         }
1120
1121         // Return node
1122         LEAVE('p', &conn->Node);
1123         return &conn->Node;
1124 }
1125
1126 /**
1127  * \brief Handle IOCtl calls
1128  */
1129 int TCP_Server_IOCtl(tVFS_Node *Node, int ID, void *Data)
1130 {
1131         tTCPListener    *srv = Node->ImplPtr;
1132
1133         switch(ID)
1134         {
1135         case 4: // Get/Set Port
1136                 if(!Data)       // Get Port
1137                         return srv->Port;
1138
1139                 if(srv->Port)   // Wait, you can't CHANGE the port
1140                         return -1;
1141
1142                 if(!CheckMem(Data, sizeof(Uint16)))     // Sanity check
1143                         return -1;
1144
1145                 // Permissions check
1146                 if(Threads_GetUID() != 0
1147                 && *(Uint16*)Data != 0
1148                 && *(Uint16*)Data < 1024)
1149                         return -1;
1150
1151                 // TODO: Check if a port is in use
1152
1153                 // Set Port
1154                 srv->Port = *(Uint16*)Data;
1155                 if(srv->Port == 0)      // Allocate a random port
1156                         srv->Port = TCP_GetUnusedPort();
1157                 else    // Else, mark this as used
1158                         TCP_AllocatePort(srv->Port);
1159
1160                 Log_Log("TCP", "Server %p listening on port %i", srv, srv->Port);
1161
1162                 return srv->Port;
1163         }
1164         return 0;
1165 }
1166
1167 void TCP_Server_Close(tVFS_Node *Node)
1168 {
1169         free(Node->ImplPtr);
1170 }
1171
1172 // --- Client
1173 /**
1174  * \brief Create a client node
1175  */
1176 tVFS_Node *TCP_Client_Init(tInterface *Interface)
1177 {
1178         tTCPConnection  *conn = TCP_int_CreateConnection(Interface, TCP_ST_CLOSED);
1179
1180         SHORTLOCK(&glTCP_OutbountCons);
1181         conn->Server = NULL;
1182         conn->Prev = NULL;
1183         conn->Next = gTCP_OutbountCons;
1184         gTCP_OutbountCons->Prev = conn;
1185         gTCP_OutbountCons = conn;
1186         SHORTREL(&glTCP_OutbountCons);
1187
1188         return &conn->Node;
1189 }
1190
1191 /**
1192  * \brief Wait for a packet and return it
1193  * \note If \a Length is smaller than the size of the packet, the rest
1194  *       of the packet's data will be discarded.
1195  */
1196 size_t TCP_Client_Read(tVFS_Node *Node, off_t Offset, size_t Length, void *Buffer, Uint Flags)
1197 {
1198         tTCPConnection  *conn = Node->ImplPtr;
1199         size_t  len;
1200
1201         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1202         LOG("conn = %p {State:%i}", conn, conn->State);
1203
1204         // If the connection has been closed (state > ST_OPEN) then clear
1205         // any stale data in the buffer (until it is empty (until it is empty))
1206         if( conn->State > TCP_ST_OPEN )
1207         {
1208                 Mutex_Acquire( &conn->lRecievedPackets );
1209                 len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1210                 Mutex_Release( &conn->lRecievedPackets );
1211
1212                 if( len == 0 ) {
1213                         VFS_MarkAvaliable(Node, 0);
1214                         errno = 0;
1215                         LEAVE('i', -1);
1216                         return -1;
1217                 }
1218
1219                 LEAVE('i', len);
1220                 return len;
1221         }
1222
1223         // Wait
1224         {
1225                 tTime   *timeout = NULL;
1226                 tTime   timeout_zero = 0;
1227                 if( Flags & VFS_IOFLAG_NOBLOCK )
1228                         timeout = &timeout_zero;
1229                 if( !VFS_SelectNode(Node, VFS_SELECT_READ|VFS_SELECT_ERROR, timeout, "TCP_Client_Read") ) {
1230                         errno = EWOULDBLOCK;
1231                         LEAVE('i', -1);
1232                         return -1;
1233                 }
1234         }
1235
1236         // Lock list and read as much as possible (up to `Length`)
1237         Mutex_Acquire( &conn->lRecievedPackets );
1238         len = RingBuffer_Read( Buffer, conn->RecievedBuffer, Length );
1239
1240         if( len == 0 || conn->RecievedBuffer->Length == 0 ) {
1241                 LOG("Marking as none avaliable (len = %i)", len);
1242                 VFS_MarkAvaliable(Node, 0);
1243         }
1244
1245         // Release the lock (we don't need it any more)
1246         Mutex_Release( &conn->lRecievedPackets );
1247
1248         LEAVE('i', len);
1249         return len;
1250 }
1251
1252 /**
1253  * \brief Send a data packet on a connection
1254  */
1255 void TCP_INT_SendDataPacket(tTCPConnection *Connection, size_t Length, const void *Data)
1256 {
1257         char    buf[sizeof(tTCPHeader)+Length];
1258         tTCPHeader      *packet = (void*)buf;
1259
1260         // - Stop Delayed ACK timer (as this data packet ACKs)
1261         Time_RemoveTimer(Connection->DeferredACKTimer);
1262
1263         packet->SourcePort = htons(Connection->LocalPort);
1264         packet->DestPort = htons(Connection->RemotePort);
1265         packet->DataOffset = (sizeof(tTCPHeader)/4)*16;
1266         packet->WindowSize = htons(TCP_WINDOW_SIZE);
1267
1268         packet->AcknowlegementNumber = htonl(Connection->NextSequenceRcv);
1269         packet->SequenceNumber = htonl(Connection->NextSequenceSend);
1270         packet->Flags = TCP_FLAG_PSH|TCP_FLAG_ACK;      // Hey, ACK if you can!
1271         packet->UrgentPointer = 0;
1272
1273         memcpy(packet->Options, Data, Length);
1274
1275         Log_Debug("TCP", "Send sequence 0x%08x", Connection->NextSequenceSend);
1276 #if HEXDUMP_OUTGOING
1277         Debug_HexDump("TCP_INT_SendDataPacket: Data = ", Data, Length);
1278 #endif
1279
1280         TCP_SendPacket( Connection, packet, Length, Data );
1281
1282         Connection->NextSequenceSend += Length;
1283 }
1284
1285 /**
1286  * \brief Send some bytes on a connection
1287  */
1288 size_t TCP_Client_Write(tVFS_Node *Node, off_t Offset, size_t Length, const void *Buffer, Uint Flags)
1289 {
1290         tTCPConnection  *conn = Node->ImplPtr;
1291         size_t  rem = Length;
1292
1293         ENTER("pNode XOffset XLength pBuffer", Node, Offset, Length, Buffer);
1294
1295 //      #if DEBUG
1296 //      Debug_HexDump("TCP_Client_Write: Buffer = ",
1297 //              Buffer, Length);
1298 //      #endif
1299
1300         // Don't allow a write to a closed connection
1301         if( conn->State > TCP_ST_OPEN ) {
1302                 VFS_MarkError(Node, 1);
1303                 errno = 0;
1304                 LEAVE('i', -1);
1305                 return -1;
1306         }
1307
1308         // Wait
1309         {
1310                 tTime   *timeout = NULL;
1311                 tTime   timeout_zero = 0;
1312                 if( Flags & VFS_IOFLAG_NOBLOCK )
1313                         timeout = &timeout_zero;
1314                 if( !VFS_SelectNode(Node, VFS_SELECT_WRITE|VFS_SELECT_ERROR, timeout, "TCP_Client_Write") ) {
1315                         errno = EWOULDBLOCK;
1316                         LEAVE('i', -1);
1317                         return -1;
1318                 }
1319         }
1320
1321         do
1322         {
1323                  int    len = (rem < TCP_MAX_PACKET_SIZE) ? rem : TCP_MAX_PACKET_SIZE;
1324
1325                 #if 0
1326                 // Wait for space in the buffer
1327                 Semaphore_Signal( &Connection->SentBufferSpace, len );
1328
1329                 // Save data to buffer (and update the length read by the ammount written)
1330                 len = RingBuffer_Write( &Connection->SentBuffer, Buffer, len);
1331                 #endif
1332
1333                 // Send packet
1334                 TCP_INT_SendDataPacket(conn, len, Buffer);
1335
1336                 Buffer += len;
1337                 rem -= len;
1338         } while( rem > 0 );
1339
1340         LEAVE('i', Length);
1341         return Length;
1342 }
1343
1344 /**
1345  * \brief Open a connection to another host using TCP
1346  * \param Conn  Connection structure
1347  */
1348 void TCP_StartConnection(tTCPConnection *Conn)
1349 {
1350         tTCPHeader      hdr = {0};
1351
1352         Conn->State = TCP_ST_SYN_SENT;
1353
1354         hdr.SourcePort = htons(Conn->LocalPort);
1355         hdr.DestPort = htons(Conn->RemotePort);
1356         Conn->NextSequenceSend = rand();
1357         hdr.SequenceNumber = htonl(Conn->NextSequenceSend);
1358         hdr.DataOffset = (sizeof(tTCPHeader)/4) << 4;
1359         hdr.Flags = TCP_FLAG_SYN;
1360         hdr.WindowSize = htons(TCP_WINDOW_SIZE);        // Max
1361         hdr.Checksum = 0;       // TODO
1362
1363         TCP_SendPacket( Conn, &hdr, 0, NULL );
1364
1365         Conn->NextSequenceSend ++;
1366         Conn->State = TCP_ST_SYN_SENT;
1367
1368         return ;
1369 }
1370
1371 /**
1372  * \brief Control a client socket
1373  */
1374 int TCP_Client_IOCtl(tVFS_Node *Node, int ID, void *Data)
1375 {
1376         tTCPConnection  *conn = Node->ImplPtr;
1377
1378         ENTER("pNode iID pData", Node, ID, Data);
1379
1380         switch(ID)
1381         {
1382         case 4: // Get/Set local port
1383                 if(!Data)
1384                         LEAVE_RET('i', conn->LocalPort);
1385                 if(conn->State != TCP_ST_CLOSED)
1386                         LEAVE_RET('i', -1);
1387                 if(!CheckMem(Data, sizeof(Uint16)))
1388                         LEAVE_RET('i', -1);
1389
1390                 if(Threads_GetUID() != 0 && *(Uint16*)Data < 1024)
1391                         LEAVE_RET('i', -1);
1392
1393                 conn->LocalPort = *(Uint16*)Data;
1394                 LEAVE_RET('i', conn->LocalPort);
1395
1396         case 5: // Get/Set remote port
1397                 if(!Data)       LEAVE_RET('i', conn->RemotePort);
1398                 if(conn->State != TCP_ST_CLOSED)        LEAVE_RET('i', -1);
1399                 if(!CheckMem(Data, sizeof(Uint16)))     LEAVE_RET('i', -1);
1400                 conn->RemotePort = *(Uint16*)Data;
1401                 LEAVE_RET('i', conn->RemotePort);
1402
1403         case 6: // Set Remote IP
1404                 if( conn->State != TCP_ST_CLOSED )
1405                         LEAVE_RET('i', -1);
1406                 if( conn->Interface->Type == 4 )
1407                 {
1408                         if(!CheckMem(Data, sizeof(tIPv4)))      LEAVE_RET('i', -1);
1409                         conn->RemoteIP.v4 = *(tIPv4*)Data;
1410                 }
1411                 else if( conn->Interface->Type == 6 )
1412                 {
1413                         if(!CheckMem(Data, sizeof(tIPv6)))      LEAVE_RET('i', -1);
1414                         conn->RemoteIP.v6 = *(tIPv6*)Data;
1415                 }
1416                 LEAVE_RET('i', 0);
1417
1418         case 7: // Connect
1419                 if(conn->LocalPort == 0xFFFF)
1420                         conn->LocalPort = TCP_GetUnusedPort();
1421                 if(conn->RemotePort == -1)
1422                         LEAVE_RET('i', 0);
1423
1424                 {
1425                         tTime   timeout = conn->Interface->TimeoutDelay;
1426
1427                         TCP_StartConnection(conn);
1428                         VFS_SelectNode(&conn->Node, VFS_SELECT_WRITE, &timeout, "TCP Connection");
1429                         if( conn->State == TCP_ST_SYN_SENT )
1430                                 LEAVE_RET('i', 0);
1431                 }
1432
1433                 LEAVE_RET('i', 1);
1434
1435         // Get recieve buffer length
1436         case 8:
1437                 LEAVE_RET('i', conn->RecievedBuffer->Length);
1438         }
1439
1440         return 0;
1441 }
1442
1443 void TCP_Client_Close(tVFS_Node *Node)
1444 {
1445         tTCPConnection  *conn = Node->ImplPtr;
1446         tTCPHeader      packet;
1447
1448         ENTER("pNode", Node);
1449
1450         ASSERT(Node->ReferenceCount != 0);
1451
1452         if( Node->ReferenceCount > 1 ) {
1453                 Node->ReferenceCount --;
1454                 LOG("Dereference only");
1455                 LEAVE('-');
1456                 return ;
1457         }
1458         Node->ReferenceCount --;
1459
1460         if( conn->State == TCP_ST_CLOSE_WAIT || conn->State == TCP_ST_OPEN )
1461         {
1462                 packet.SourcePort = htons(conn->LocalPort);
1463                 packet.DestPort = htons(conn->RemotePort);
1464                 packet.DataOffset = (sizeof(tTCPHeader)/4)*16;
1465                 packet.WindowSize = TCP_WINDOW_SIZE;
1466
1467                 packet.AcknowlegementNumber = 0;
1468                 packet.SequenceNumber = htonl(conn->NextSequenceSend);
1469                 packet.Flags = TCP_FLAG_FIN;
1470
1471                 TCP_SendPacket( conn, &packet, 0, NULL );
1472         }
1473
1474         Time_RemoveTimer(conn->DeferredACKTimer);
1475
1476         switch( conn->State )
1477         {
1478         case TCP_ST_CLOSED:
1479                 Log_Warning("TCP", "Closing connection that was never opened");
1480                 TCP_int_FreeTCB(conn);
1481                 break;
1482         case TCP_ST_CLOSE_WAIT:
1483                 conn->State = TCP_ST_LAST_ACK;
1484                 break;
1485         case TCP_ST_OPEN:
1486                 conn->State = TCP_ST_FIN_WAIT1;
1487                 while( conn->State == TCP_ST_FIN_WAIT1 )
1488                         Threads_Yield();
1489                 // No free, freed after TIME_WAIT
1490                 break;
1491         default:
1492                 Log_Warning("TCP", "Unhandled connection state %i in TCP_Client_Close",
1493                         conn->State);
1494                 break;
1495         }
1496
1497         LEAVE('-');
1498 }
1499
1500 /**
1501  * \brief Checks if a value is between two others (after taking into account wrapping)
1502  */
1503 int WrapBetween(Uint32 Lower, Uint32 Value, Uint32 Higher, Uint32 MaxValue)
1504 {
1505         if( MaxValue < 0xFFFFFFFF )
1506         {
1507                 Lower %= MaxValue + 1;
1508                 Value %= MaxValue + 1;
1509                 Higher %= MaxValue + 1;
1510         }
1511
1512         // Simple Case, no wrap ?
1513         //       Lower Value Higher
1514         // | ... + ... + ... + ... |
1515
1516         if( Lower < Higher ) {
1517                 return Lower < Value && Value < Higher;
1518         }
1519         // Higher has wrapped below lower
1520
1521         // Value > Lower ?
1522         //       Higher Lower Value
1523         // | ... +  ... + ... + ... |
1524         if( Value > Lower ) {
1525                 return 1;
1526         }
1527
1528         // Value < Higher ?
1529         //       Value Higher Lower
1530         // | ... + ... +  ... + ... |
1531         if( Value < Higher ) {
1532                 return 1;
1533         }
1534
1535         return 0;
1536 }