Ticket #5240: tcp-draft-02.diff

headers/private/kernel/boot/net/NetDefs.h

@@ -1 +1 @@
 /*
  * Copyright 2005, Ingo Weinhold <bonefish@cs.tu-berlin.de>.
+ * Copyright 2010, Andreas Faerber <andreas.faerber@web.de>
  * All rights reserved. Distributed under the terms of the MIT License.
  */
 
@@ -157 +158 @@
 #define IP_DEFAULT_TIME_TO_LIVE     64      /* default ttl, from RFC 1340 */
 
 // IP protocols
+#define IPPROTO_TCP                 6
 #define IPPROTO_UDP                 17
 
 
@@ -174 +176 @@
 } __attribute__ ((__packed__));
 
 
+// Transmission Control Protocol (TCP)
+
+// TCP header (RFC 793, RFC 3168)
+struct tcp_header
+{
+    uint16 source;          // source port
+    uint16 destination;     // destination port
+    uint32 seqNumber;       // sequence number
+    uint32 ackNumber;       // acknowledgment number
+#if __BYTE_ORDER == __BIG_ENDIAN
+    uint8 dataOffset:4;     // data offset
+    uint8 reserved:4;       // reserved
+#elif __BYTE_ORDER == __LITTLE_ENDIAN
+    uint8 reserved:4;       // reserved
+    uint8 dataOffset:4;     // data offset
+#endif
+#if __BYTE_ORDER == __BIG_ENDIAN
+    bool cwr:1;
+    bool ece:1;
+    bool urg:1;
+    bool ack:1;
+    bool psh:1;
+    bool rst:1;
+    bool syn:1;
+    bool fin:1;
+#elif __BYTE_ORDER == __LITTLE_ENDIAN
+    bool fin:1;
+    bool syn:1;
+    bool rst:1;
+    bool psh:1;
+    bool ack:1;
+    bool urg:1;
+    bool ece:1;
+    bool cwr:1;
+#endif
+    uint16 window;          // window size
+    uint16 checksum;        // checksum
+    uint16 urgentPointer;   // urgent pointer
+} __attribute__ ((__packed__));
+
 // #pragma mark -
 
 // NetService
@@ -183 +225 @@
 extern const char *const kARPServiceName;
 extern const char *const kIPServiceName;
 extern const char *const kUDPServiceName;
+extern const char *const kTCPServiceName;
 
 class NetService {
 public:

headers/private/kernel/boot/net/NetStack.h

@@ -13 +13 @@
 class ARPService;
 class IPService;
 class UDPService;
+class TCPService;
 
 
 class NetStack {
@@ -34 +35 @@
     ARPService *GetARPService() const           { return fARPService; }
     IPService *GetIPService() const             { return fIPService; }
     UDPService *GetUDPService() const           { return fUDPService; }
+    TCPService *GetTCPService() const           { return fTCPService; }
 
 private:
     static NetStack     *sNetStack;
@@ -43 +45 @@
     ARPService          *fARPService;
     IPService           *fIPService;
     UDPService          *fUDPService;
+    TCPService          *fTCPService;
 };
 
 

new file headers/private/kernel/boot/net/TCP.h

@@ +1 @@
+/*
+ * Copyright 2010 Andreas Faerber <andreas.faerber@web.de>
+ * All rights reserved. Distributed under the terms of the MIT License.
+ */
+
+#ifndef BOOT_NET_TCP_H
+#define BOOT_NET_TCP_H
+
+#include <boot/net/IP.h>
+
+
+class TCPPacket {
+public:
+    TCPPacket();
+    ~TCPPacket();
+
+    status_t SetTo(const void* data, size_t size, ip_addr_t sourceAddress,
+        uint16 sourcePort, ip_addr_t destinationAddress,
+        uint16 destinationPort, uint32 sequenceNumber,
+        uint32 acknowledgmentNumber, uint8 flags);
+
+    ip_addr_t SourceAddress() const;
+    ip_addr_t DestinationAddress() const;
+    uint16 SourcePort() const;
+    uint16 DestinationPort() const;
+    uint32 SequenceNumber() const;
+    uint32 AcknowledgmentNumber() const;
+    const void* Data() const { return fData; }
+    size_t DataSize() const { return fSize; }
+    uint8 Flags() const { return fFlags; }
+
+    bool ProvidesSequenceNumber(uint32 sequenceNo) const;
+
+    TCPPacket* Next() const;
+    void SetNext(TCPPacket* packet);
+
+private:
+    ip_addr_t   fSourceAddress;
+    ip_addr_t   fDestinationAddress;
+    uint16      fSourcePort;
+    uint16      fDestinationPort;
+    uint32      fSequenceNumber;
+    uint32      fAcknowledgmentNumber;
+    void*       fData;
+    size_t      fSize;
+    uint8       fFlags;
+    TCPPacket*  fNext;
+};
+
+#define TCP_SYN     1
+#define TCP_ACK     2
+#define TCP_FIN     4
+#define TCP_RST     8
+
+class TCPService;
+
+enum TCPSocketState {
+    TCP_SOCKET_STATE_INITIAL,
+    TCP_SOCKET_STATE_SYN_SENT,
+    TCP_SOCKET_STATE_SYN_RECEIVED,
+    TCP_SOCKET_STATE_OPEN,
+    TCP_SOCKET_STATE_FIN_SENT,
+    TCP_SOCKET_STATE_CLOSED
+};
+
+class TCPSocket {
+public:
+    TCPSocket();
+    ~TCPSocket();
+
+    ip_addr_t Address() const   { return fAddress; }
+    uint16 Port() const         { return fPort; }
+
+    status_t Connect(ip_addr_t address, uint16 port);
+    status_t Close();
+    status_t Read(void* buffer, size_t bufferSize, size_t* bytesRead, bigtime_t timeout = 0);
+    status_t Write(const void* buffer, size_t bufferSize);
+
+    void Acknowledge(uint32 number);
+    void ProcessPacket(TCPPacket* packet);
+
+private:
+    TCPPacket* _PeekPacket();
+    TCPPacket* _DequeuePacket();
+    status_t _Send(TCPPacket* packet, bool enqueue = true);
+    status_t _ResendQueue();
+    void _EnqueueOutgoingPacket(TCPPacket* packet);
+    status_t _WaitForState(TCPSocketState state, bigtime_t timeout = 0);
+    status_t _Ack();
+
+    TCPService* fTCPService;
+    ip_addr_t   fAddress;
+    uint16      fPort;
+    ip_addr_t   fRemoteAddress;
+    uint16      fRemotePort;
+    uint32      fSequenceNumber;
+    uint32      fAcknowledgeNumber;
+    uint32      fNextSequence;
+    TCPPacket*  fFirstPacket;
+    TCPPacket*  fLastPacket;
+    TCPPacket*  fFirstSentPacket;
+    TCPPacket*  fLastSentPacket;
+    TCPSocketState fState;
+    TCPSocketState fRemoteState;
+};
+
+class TCPService : public IPSubService {
+public:
+    TCPService(IPService* ipService);
+    virtual ~TCPService();
+
+    status_t Init();
+
+    virtual uint8 IPProtocol() const;
+
+    virtual void HandleIPPacket(IPService* ipService, ip_addr_t sourceIP,
+        ip_addr_t destinationIP, const void* data, size_t size);
+
+    status_t Send(uint16 sourcePort, ip_addr_t destinationAddress,
+        uint16 destinationPort, uint32 sequenceNumber,
+        uint32 acknowledgmentNumber, uint8 flags, ChainBuffer* buffer);
+
+    void ProcessIncomingPackets();
+
+    status_t BindSocket(TCPSocket* socket);
+    void UnbindSocket(TCPSocket* socket);
+
+private:
+    uint16 _ChecksumBuffer(ChainBuffer* buffer, ip_addr_t source,
+        ip_addr_t destination, uint16 length);
+    uint16 _ChecksumData(const void* data, uint16 length, ip_addr_t source,
+        ip_addr_t destination);
+
+    TCPSocket* _FindSocket(ip_addr_t address, uint16 port);
+
+    IPService*          fIPService;
+    Vector<TCPSocket*>  fSockets;
+};
+
+#endif

src/system/boot/loader/net/NetDefs.cpp

@@ -16 +16 @@
 const char *const kARPServiceName = "arp";
 const char *const kIPServiceName = "ip";
 const char *const kUDPServiceName = "udp";
+const char *const kTCPServiceName = "tcp";
 
 
 // constructor

src/system/boot/loader/net/NetStack.cpp

@@ -13 +13 @@
 #include <boot/net/Ethernet.h>
 #include <boot/net/IP.h>
 #include <boot/net/UDP.h>
+#include <boot/net/TCP.h>
 
 
 // sNetStack
@@ -24 +25 @@
         fEthernetService(NULL),
         fARPService(NULL),
         fIPService(NULL),
-        fUDPService(NULL)
+        fUDPService(NULL),
+        fTCPService(NULL)
 {
 }
 
 // destructor
 NetStack::~NetStack()
 {
+    delete fTCPService;
     delete fUDPService;
     delete fIPService;
     delete fARPService;
@@ -73 +76 @@
     if (error != B_OK)
         return error;
 
+    // TCP service
+    fTCPService = new(nothrow) TCPService(fIPService);
+    if (fTCPService == NULL)
+        return B_NO_MEMORY;
+    error = fTCPService->Init();
+    if (error != B_OK)
+        return error;
+
     return B_OK;
 }
 

new file src/system/boot/loader/net/TCP.cpp

@@ +1 @@
+/*
+ * Copyright 2010 Andreas Faerber <andreas.faerber@web.de>
+ * All rights reserved. Distributed under the terms of the MIT License.
+ */
+
+#include <boot/net/TCP.h>
+
+#include <stdio.h>
+#include <KernelExport.h>
+
+#include <boot/net/ChainBuffer.h>
+#include <boot/net/NetStack.h>
+
+
+//#define TRACE_TCP
+#ifdef TRACE_TCP
+#   define TRACE(x) dprintf x
+#else
+#   define TRACE(x) ;
+#endif
+
+
+static int
+_rand(void)
+{
+    static int next = 0;
+    if (next == 0)
+        next = system_time();
+
+    next = next * 1103515245 + 12345;
+    return next;
+}
+
+
+TCPPacket::TCPPacket()
+    :
+    fData(NULL),
+    fNext(NULL)
+{
+}
+
+
+TCPPacket::~TCPPacket()
+{
+    free(fData);
+}
+
+
+status_t
+TCPPacket::SetTo(const void* data, size_t size, ip_addr_t sourceAddress,
+    uint16 sourcePort, ip_addr_t destinationAddress, uint16 destinationPort,
+    uint32 sequenceNumber, uint32 acknowledgmentNumber, uint8 flags)
+{
+    if (data == NULL && size > 0)
+        return B_BAD_VALUE;
+
+    if (size > 0) {
+        fData = malloc(size);
+        if (fData == NULL)
+            return B_NO_MEMORY;
+        memcpy(fData, data, size);
+    } else {
+        fData = NULL;
+    }
+
+    fSize = size;
+    fSourceAddress = sourceAddress;
+    fSourcePort = sourcePort;
+    fDestinationAddress = destinationAddress;
+    fDestinationPort = destinationPort;
+    fSequenceNumber = sequenceNumber;
+    fAcknowledgmentNumber = acknowledgmentNumber;
+    fFlags = flags;
+
+    return B_OK;
+}
+
+
+ip_addr_t
+TCPPacket::SourceAddress() const
+{
+    return fSourceAddress;
+}
+
+
+ip_addr_t
+TCPPacket::DestinationAddress() const
+{
+    return fDestinationAddress;
+}
+
+
+uint16
+TCPPacket::SourcePort() const
+{
+    return fSourcePort;
+}
+
+
+uint16
+TCPPacket::DestinationPort() const
+{
+    return fDestinationPort;
+}
+
+
+uint32
+TCPPacket::SequenceNumber() const
+{
+    return fSequenceNumber;
+}
+
+
+uint32
+TCPPacket::AcknowledgmentNumber() const
+{
+    return fAcknowledgmentNumber;
+}
+
+
+bool
+TCPPacket::ProvidesSequenceNumber(uint32 sequenceNumber) const
+{
+    return fSequenceNumber <= sequenceNumber
+        && fSequenceNumber + fSize > sequenceNumber;
+}
+
+
+TCPPacket*
+TCPPacket::Next() const
+{
+    return fNext;
+}
+
+
+void
+TCPPacket::SetNext(TCPPacket* packet)
+{
+    fNext = packet;
+}
+
+
+
+
+TCPSocket::TCPSocket()
+    :
+    fTCPService(NetStack::Default()->GetTCPService()),
+    fAddress(INADDR_ANY),
+    fPort(0),
+    fSequenceNumber(0),
+    fFirstPacket(NULL),
+    fLastPacket(NULL),
+    fFirstSentPacket(NULL),
+    fLastSentPacket(NULL),
+    fState(TCP_SOCKET_STATE_INITIAL),
+    fRemoteState(TCP_SOCKET_STATE_INITIAL)
+{
+}
+
+
+TCPSocket::~TCPSocket()
+{
+    if (fTCPService != NULL && fPort != 0)
+        fTCPService->UnbindSocket(this);
+}
+
+
+status_t
+TCPSocket::Connect(ip_addr_t address, uint16 port)
+{
+    fRemoteAddress = address;
+    fRemotePort = port;
+    fSequenceNumber = _rand();
+    fPort = 0xC000 + ((_rand() >> 16) & ~0xc000);
+    dprintf("TCPSocket::Connect(): connecting from port %u\n", fPort);
+    fAcknowledgeNumber = 0;
+    fNextSequence = 0;
+
+    status_t error = fTCPService->BindSocket(this);
+    if (error != B_OK)
+        return error;
+
+    // send SYN
+    TCPPacket* packet = new(nothrow) TCPPacket();
+    if (packet == NULL)
+        return B_NO_MEMORY;
+    error = packet->SetTo(NULL, 0, fAddress, fPort, address, port,
+        fSequenceNumber, fAcknowledgeNumber, TCP_SYN);
+    if (error != B_OK)
+        return error;
+    error = _Send(packet);
+    if (error != B_OK)
+        return error;
+    fState = TCP_SOCKET_STATE_SYN_SENT;
+    fSequenceNumber++;
+    TRACE(("SYN sent\n"));
+
+    // receive SYN-ACK
+    error = _WaitForState(TCP_SOCKET_STATE_OPEN, 1000000LL);
+    if (error != B_OK) {
+        TRACE(("no SYN-ACK received\n"));
+        return error;
+    }
+    TRACE(("SYN-ACK received\n"));
+
+    return B_OK;
+}
+
+
+status_t
+TCPSocket::Close()
+{
+    // send FIN
+    TCPPacket* packet = new(nothrow) TCPPacket();
+    if (packet == NULL)
+        return B_NO_MEMORY;
+    status_t error = packet->SetTo(NULL, 0, fAddress, fPort, fRemoteAddress, fRemotePort,
+        fSequenceNumber, fAcknowledgeNumber, TCP_FIN | TCP_ACK);
+    if (error != B_OK)
+        return error;
+    error = _Send(packet);
+    if (error != B_OK)
+        return error;
+    fState = TCP_SOCKET_STATE_FIN_SENT;
+    TRACE(("FIN sent\n"));
+
+    error = _WaitForState(TCP_SOCKET_STATE_CLOSED, 1000000LL);
+    if (error != B_OK)
+        return error;
+
+    return B_OK;
+}
+
+
+status_t
+TCPSocket::Read(void* buffer, size_t bufferSize, size_t* bytesRead, bigtime_t timeout)
+{
+    TRACE(("TCPSocket::Read(): size = %lu\n", bufferSize));
+    if (bytesRead == NULL)
+        return B_BAD_VALUE;
+
+    *bytesRead = 0;
+    TCPPacket* packet = NULL;
+    
+    bigtime_t startTime = system_time();
+    do {
+        fTCPService->ProcessIncomingPackets();
+        //_ResendQueue();
+        packet = _PeekPacket();
+        if (packet != NULL && fRemoteState != TCP_SOCKET_STATE_OPEN)
+            return B_WOULD_BLOCK;
+    } while (packet == NULL && system_time() - startTime < timeout);
+    if (packet == NULL)
+        return (timeout == 0) ? B_WOULD_BLOCK : B_TIMED_OUT;
+    uint32 packetOffset = fNextSequence - packet->SequenceNumber();
+    size_t readBytes = packet->DataSize() - packetOffset;
+    if (readBytes > bufferSize)
+        readBytes = bufferSize;
+    if (buffer != NULL)
+        memcpy(buffer, (uint8*)packet->Data() + packetOffset, readBytes);
+    *bytesRead = readBytes;
+    if (!packet->ProvidesSequenceNumber(fNextSequence + readBytes)) {
+        _DequeuePacket();
+        delete packet;
+        packet = NULL;
+    }
+    fNextSequence += readBytes;
+
+    if (packet == NULL && *bytesRead < bufferSize) {
+        do {
+            if (buffer != NULL)
+                buffer = (uint8*)buffer + readBytes;
+            bufferSize -= readBytes;
+            fTCPService->ProcessIncomingPackets();
+            packet = _PeekPacket();
+            if (packet == NULL && fRemoteState != TCP_SOCKET_STATE_OPEN)
+                break;
+            readBytes = 0;
+            if (packet == NULL)
+                continue;
+            readBytes = packet->DataSize();
+            if (readBytes > bufferSize)
+                readBytes = bufferSize;
+            if (buffer != NULL)
+                memcpy(buffer, packet->Data(), readBytes);
+            *bytesRead += readBytes;
+            if (readBytes == packet->DataSize()) {
+                _DequeuePacket();
+                delete packet;
+            }
+            fNextSequence += readBytes;
+        } while (readBytes < bufferSize && system_time() - startTime < timeout);
+    }
+
+    return B_OK;
+}
+
+
+status_t
+TCPSocket::Write(const void* buffer, size_t bufferSize)
+{
+    if (buffer == NULL || bufferSize <= 0)
+        return B_BAD_VALUE;
+
+    TCPPacket* packet = new(nothrow) TCPPacket();
+    if (packet == NULL)
+        return B_NO_MEMORY;
+    status_t error = packet->SetTo(buffer, bufferSize, fAddress, fPort,
+        fRemoteAddress, fRemotePort, fSequenceNumber, fAcknowledgeNumber,
+        TCP_ACK);
+    if (error != B_OK)
+        return error;
+    return _Send(packet);
+}
+
+
+void
+TCPSocket::Acknowledge(uint32 number)
+{
+    TRACE(("TCPSocket::Acknowledge(): %lu\n", number));
+    // dequeue packets
+    for (TCPPacket* packet = fFirstSentPacket; packet != NULL;
+            packet = fFirstSentPacket) {
+        if (packet->SequenceNumber() >= number)
+            return;
+        fFirstSentPacket = packet->Next();
+        delete packet;
+    }
+    fLastSentPacket = NULL;
+}
+
+
+void
+TCPSocket::ProcessPacket(TCPPacket* packet)
+{
+    TRACE(("TCPSocket::ProcessPacket()\n"));
+
+    if ((packet->Flags() & TCP_FIN) != 0) {
+        fRemoteState = TCP_SOCKET_STATE_FIN_SENT;
+        TRACE(("FIN received\n"));
+        _Ack();
+    }
+
+    if (fState == TCP_SOCKET_STATE_SYN_SENT) {
+        if ((packet->Flags() & TCP_SYN) != 0
+                && (packet->Flags() & TCP_ACK) != 0) {
+            fNextSequence = fAcknowledgeNumber = packet->SequenceNumber() + 1;
+            fRemoteState = TCP_SOCKET_STATE_SYN_SENT;
+            delete packet;
+            _Ack();
+            fState = fRemoteState = TCP_SOCKET_STATE_OPEN;
+            return;
+        }
+    } else if (fState == TCP_SOCKET_STATE_OPEN) {
+    } else if (fState == TCP_SOCKET_STATE_FIN_SENT) {
+        if ((packet->Flags() & TCP_ACK) != 0) {
+            TRACE(("FIN-ACK received\n"));
+            if (fRemoteState == TCP_SOCKET_STATE_FIN_SENT)
+                fState = TCP_SOCKET_STATE_CLOSED;
+        }
+    }
+
+    if (packet->DataSize() == 0) {
+        TRACE(("TCPSocket::ProcessPacket(): not queuing\n"));
+        return;
+    }
+
+    if (fLastPacket == NULL) {
+        // no packets enqueued
+        TRACE(("TCPSocket::ProcessPacket(): first in queue\n"));
+        packet->SetNext(NULL);
+        fFirstPacket = fLastPacket = packet;
+    } else if (fLastPacket->SequenceNumber() < packet->SequenceNumber()) {
+        // enqueue in back
+        TRACE(("TCPSocket::ProcessPacket(): enqueue in back\n"));
+        fLastPacket->SetNext(packet);
+        fLastPacket = packet;
+    } else if (fFirstPacket->SequenceNumber() > packet->SequenceNumber()) {
+        // enqueue in front
+        TRACE(("TCPSocket::ProcessPacket(): enqueue in front\n"));
+        packet->SetNext(fFirstPacket);
+        fFirstPacket = packet;
+    } else {
+        // enqueue in middle
+        TRACE(("TCPSocket::ProcessPacket(): enqueue in middle\n"));
+        for (TCPPacket* queuedPacket = fFirstPacket; queuedPacket != NULL;
+                queuedPacket = queuedPacket->Next()) {
+            if (queuedPacket->SequenceNumber() == packet->SequenceNumber()) {
+                TRACE(("TCPSocket::EnqueuePacket(): TCP packet dropped\n"));
+                // we may be waiting for a previous packet
+                return;
+            }
+            if (queuedPacket->Next()->SequenceNumber() > packet->SequenceNumber()) {
+                packet->SetNext(queuedPacket->Next());
+                queuedPacket->SetNext(packet);
+                break;
+            }
+        }
+    }
+    if (packet->SequenceNumber() == fAcknowledgeNumber)
+        fAcknowledgeNumber = packet->SequenceNumber() + packet->DataSize();
+            //XXX might be greater than that
+}
+
+
+TCPPacket*
+TCPSocket::_PeekPacket()
+{
+    TRACE(("TCPSocket::_PeekPacket(): fNextSequence = %lu\n", fNextSequence));
+    for (TCPPacket* packet = fFirstPacket; packet != NULL;
+            packet = packet->Next()) {
+        if (packet->ProvidesSequenceNumber(fNextSequence))
+            return packet;
+    }
+    return NULL;
+}
+
+
+TCPPacket*
+TCPSocket::_DequeuePacket()
+{
+    //TRACE(("TCPSocket::DequeuePacket()\n"));
+    if (fFirstPacket == NULL)
+        return NULL;
+
+    if (fFirstPacket->ProvidesSequenceNumber(fNextSequence)) {
+        TCPPacket* packet = fFirstPacket;
+        fFirstPacket = packet->Next();
+        if (fFirstPacket == NULL)
+            fLastPacket = NULL;
+        packet->SetNext(NULL);
+        return packet;
+    }
+
+    for (TCPPacket* packet = fFirstPacket;
+            packet != NULL && packet->Next() != NULL;
+            packet = packet->Next()) {
+        if (packet->Next()->ProvidesSequenceNumber(fNextSequence)) {
+            TCPPacket* nextPacket = packet->Next();
+            packet->SetNext(nextPacket->Next());
+            if (fLastPacket == nextPacket)
+                fLastPacket = packet;
+            fNextSequence += nextPacket->DataSize();
+            return nextPacket;
+        }
+    }
+    return NULL;
+}
+
+
+status_t
+TCPSocket::_Send(TCPPacket* packet, bool enqueue)
+{
+    ChainBuffer buffer((void*)packet->Data(), packet->DataSize());
+    status_t error = fTCPService->Send(fPort, fRemoteAddress, fRemotePort,
+        packet->SequenceNumber(), fAcknowledgeNumber, packet->Flags(),
+        &buffer);
+    if (error != B_OK)
+        return error;
+    fSequenceNumber += packet->DataSize();
+    if (enqueue)
+        _EnqueueOutgoingPacket(packet);
+    return B_OK;
+}
+
+
+status_t
+TCPSocket::_ResendQueue()
+{
+    TRACE(("TCPSocket::_ResendQueue()\n"));
+    for (TCPPacket* packet = fFirstSentPacket; packet != NULL;
+            packet = packet->Next()) {
+        ChainBuffer buffer((void*)packet->Data(), packet->DataSize());
+        status_t error = fTCPService->Send(fPort, fRemoteAddress, fRemotePort,
+            packet->SequenceNumber(), fAcknowledgeNumber, packet->Flags(),
+            &buffer);
+        if (error != B_OK)
+            return error;
+    }
+    return B_OK;
+}
+
+
+void
+TCPSocket::_EnqueueOutgoingPacket(TCPPacket* packet)
+{
+    if (fLastSentPacket != NULL) {
+        fLastSentPacket->SetNext(packet);
+        fLastSentPacket = packet;
+    } else {
+        fFirstSentPacket = fLastSentPacket = packet;
+    }
+}
+
+
+status_t
+TCPSocket::_Ack()
+{
+    TCPPacket* packet = new(nothrow) TCPPacket();
+    if (packet == NULL)
+        return B_NO_MEMORY;
+    status_t error = packet->SetTo(NULL, 0, fAddress, fPort, fRemoteAddress,
+        fRemotePort, fSequenceNumber, fAcknowledgeNumber, TCP_ACK);
+    if (error != B_OK)
+        return error;
+    error = _Send(packet);
+    if (error != B_OK)
+        return error;
+    return B_OK;
+}
+
+
+status_t
+TCPSocket::_WaitForState(TCPSocketState state, bigtime_t timeout)
+{
+    if (fTCPService == NULL)
+        return B_NO_INIT;
+
+    bigtime_t startTime = system_time();
+    do {
+        fTCPService->ProcessIncomingPackets();
+        if (fState == state)
+            return B_OK;
+    } while (system_time() - startTime < timeout);
+    return timeout == 0 ? B_WOULD_BLOCK : B_TIMED_OUT;
+}
+
+
+
+
+TCPService::TCPService(IPService* ipService)
+    :
+    IPSubService(kTCPServiceName),
+    fIPService(ipService)
+{
+}
+
+
+TCPService::~TCPService()
+{
+    if (fIPService != NULL)
+        fIPService->UnregisterIPSubService(this);
+}
+
+
+status_t
+TCPService::Init()
+{
+    if (fIPService == NULL)
+        return B_BAD_VALUE;
+    if (!fIPService->RegisterIPSubService(this))
+        return B_NO_MEMORY;
+    return B_OK;
+}
+
+
+uint8
+TCPService::IPProtocol() const
+{
+    return IPPROTO_TCP;
+}
+
+
+void
+TCPService::HandleIPPacket(IPService* ipService, ip_addr_t sourceIP,
+    ip_addr_t destinationIP, const void* data, size_t size)
+{
+    TRACE(("TCPService::HandleIPPacket(): source = %08lx, destination = %08lx, "
+        "%lu - %lu bytes\n", sourceIP, destinationIP, size, sizeof(tcp_header)));
+
+    if (data == NULL || size < sizeof(tcp_header))
+        return;
+
+    const tcp_header* header = (const tcp_header*)data;
+    uint16 source = ntohs(header->source);
+    uint16 destination = ntohs(header->destination);
+    uint32 sequenceNumber = ntohl(header->seqNumber);
+    uint32 ackedNumber = ntohl(header->ackNumber);
+    uint8 flags = 0;
+    if (header->syn)
+        flags |= TCP_SYN;
+    if (header->ack)
+        flags |= TCP_ACK;
+    if (header->fin)
+        flags |= TCP_FIN;
+    if (header->rst)
+        flags |= TCP_RST;
+    TRACE(("\tsource = %u, dest = %u, seq = %lu, ack = %lu, dataOffset = %u, flags %s %s %s %s\n",
+        source, destination, sequenceNumber, ackedNumber, header->dataOffset,
+        header->ack ? "ACK" : "", header->syn ? "SYN" : "", header->fin ? "FIN" : "", header->rst ? "RST" : ""));
+    if (header->dataOffset > 5) {
+        uint8* option = (uint8*)data + sizeof(tcp_header);
+        while ((uint32*)option < (uint32*)data + header->dataOffset) {
+            uint8 optionKind = option[0];
+            uint8 optionLength = 1;
+            if (optionKind > 1)
+                optionLength = option[1];
+            TRACE(("\tTCP option kind %u, length %u\n", optionKind, optionLength));
+            option += optionLength;
+        }
+    }
+
+    uint16 chksum = _ChecksumData(data, size, sourceIP, destinationIP);
+    if (chksum != 0) {
+        TRACE(("TCPService::HandleIPPacket(): invalid checksum (%04x)\n", header->checksum));
+        return;
+    }
+
+    TCPSocket* socket = _FindSocket(destinationIP, destination);
+    if (socket == NULL) {
+        //TODO if SYN, answer with RST?
+        TRACE(("TCPService::HandleIPPacket(): no socket\n"));
+        return;
+    }
+
+    if (header->ack) {
+        socket->Acknowledge(ackedNumber);
+    }
+
+    TCPPacket* packet = new(nothrow) TCPPacket();
+    if (packet == NULL)
+        return;
+    status_t error = packet->SetTo((uint32*)data + header->dataOffset,
+        size - header->dataOffset * 4, sourceIP, source, destinationIP,
+        destination, sequenceNumber, ackedNumber, flags);
+    if (error == B_OK)
+        socket->ProcessPacket(packet);
+    else
+        delete packet;
+}
+
+
+status_t
+TCPService::Send(uint16 sourcePort, ip_addr_t destinationAddress,
+    uint16 destinationPort, uint32 sequenceNumber,
+    uint32 acknowledgmentNumber, uint8 flags, ChainBuffer* buffer)
+{
+    TRACE(("TCPService::Send(): seq = %lu, ack = %lu\n", sequenceNumber, acknowledgmentNumber));
+    if (fIPService == NULL)
+        return B_NO_INIT;
+    if (buffer == NULL)
+        return B_BAD_VALUE;
+
+    tcp_header header;
+    ChainBuffer headerBuffer(&header, sizeof(header), buffer);
+    memset(&header, 0, sizeof(header));
+    header.source = htons(sourcePort);
+    header.destination = htons(destinationPort);
+    header.seqNumber = htonl(sequenceNumber);
+    header.ackNumber = htonl(acknowledgmentNumber);
+    header.dataOffset = 5;
+    header.ack = (flags & TCP_ACK) != 0;
+    header.syn = (flags & TCP_SYN) != 0;
+    header.fin = (flags & TCP_FIN) != 0;
+    header.rst = (flags & TCP_RST) != 0;
+    header.window = 0xffff;
+
+    header.checksum = 0;
+    header.checksum = htons(_ChecksumBuffer(&headerBuffer,
+        fIPService->IPAddress(), destinationAddress,
+        headerBuffer.TotalSize()));
+
+    return fIPService->Send(destinationAddress, IPPROTO_TCP, &headerBuffer);
+}
+
+
+void
+TCPService::ProcessIncomingPackets()
+{
+    if (fIPService != NULL)
+        fIPService->ProcessIncomingPackets();
+}
+
+
+status_t
+TCPService::BindSocket(TCPSocket* socket)
+{
+    if (socket == NULL)
+        return B_BAD_VALUE;
+
+    if (_FindSocket(socket->Address(), socket->Port()))
+        return EADDRINUSE;
+
+    return fSockets.Add(socket);
+}
+
+
+void
+TCPService::UnbindSocket(TCPSocket* socket)
+{
+    fSockets.Remove(socket);
+}
+
+
+uint16
+TCPService::_ChecksumBuffer(ChainBuffer* buffer, ip_addr_t source,
+    ip_addr_t destination, uint16 length)
+{
+    struct pseudo_header {
+        ip_addr_t   source;
+        ip_addr_t   destination;
+        uint8       pad;
+        uint8       protocol;
+        uint16      length;
+    } __attribute__ ((__packed__));
+    pseudo_header header = {
+        htonl(source),
+        htonl(destination),
+        0,
+        IPPROTO_TCP,
+        htons(length)
+    };
+
+    ChainBuffer headerBuffer(&header, sizeof(header), buffer);
+    uint16 checksum = ip_checksum(&headerBuffer);
+    headerBuffer.DetachNext();
+    return checksum;
+}
+
+
+uint16
+TCPService::_ChecksumData(const void* data, uint16 length, ip_addr_t source,
+    ip_addr_t destination)
+{
+    ChainBuffer buffer((void*)data, length);
+    return _ChecksumBuffer(&buffer, source, destination, length);
+}
+
+
+TCPSocket*
+TCPService::_FindSocket(ip_addr_t address, uint16 port)
+{
+    for (int i = 0; i < fSockets.Count(); i++) {
+        TCPSocket* socket = fSockets.ElementAt(i);
+        if ((address == INADDR_ANY || socket->Address() == INADDR_ANY
+                    || socket->Address() == address)
+                && socket->Port() == port) {
+            return socket;
+        }
+    }
+    return NULL;
+}

src/system/boot/loader/net/Jamfile

@@ -13 +14 @@
     NetStack.cpp
     RemoteDisk.cpp
     UDP.cpp
+    TCP.cpp
 
     : -fno-pic
 ;