Rewrote how bind() works:

* there are now two hash tables: one for connections, and one for ports
* the first one is used to find the endpoint for incoming connections
* the latter is used to check if the address to bind() to is still available
  (incl. support for SO_REUSEADDR, and SO_REUSEPORT). Specialising an existing
  socket is not allowed, though; wildcard sockets need to be started last.
* the TCPConnection class now has a pointer to the next endpoint with the
  same port number - this list is scanned for the existing bound sockets
  on that port.


git-svn-id: file:///srv/svn/repos/haiku/haiku/trunk@19416 a95241bf-73f2-0310-859d-f6bbb57e9c96

src/add-ons/kernel/network/protocols/tcp/EndpointManager.cpp

@@ -0,0 +1,400 @@
+/*
+ * Copyright 2006, Haiku, Inc. All Rights Reserved.
+ * Distributed under the terms of the MIT License.
+ *
+ * Authors:
+ *		Axel Dörfler, axeld@pinc-software.de
+ */
+
+
+#include "EndpointManager.h"
+#include "TCPConnection.h"
+
+#include <NetUtilities.h>
+
+#include <util/AutoLock.h>
+
+#include <KernelExport.h>
+
+
+#define TRACE_ENDPOINT_MANAGER
+#ifdef TRACE_ENDPOINT_MANAGER
+#	define TRACE(x) dprintf x
+#else
+#	define TRACE(x)
+#endif
+
+
+struct connection_key {
+	const sockaddr	*local;
+	const sockaddr	*peer;
+};
+
+struct endpoint_key {
+	uint16	port;
+};
+
+
+static const uint32 kConnectionHashBuckets = 256;
+static const uint32 kEndpointHashBuckets = 256;
+
+static const uint16 kLastReservedPort = 1023;
+static const uint16 kFirstEphemeralPort = 40000;
+
+
+EndpointManager::EndpointManager()
+{
+	fConnectionHash = hash_init(kConnectionHashBuckets,
+		offsetof(TCPConnection, fConnectionHashNext),
+		&_ConnectionCompare, &_ConnectionHash);
+	fEndpointHash = hash_init(kEndpointHashBuckets,
+		offsetof(TCPConnection, fEndpointHashNext),
+		&_EndpointCompare, &_EndpointHash);
+
+	recursive_lock_init(&fLock, "endpoint manager");
+}
+
+
+EndpointManager::~EndpointManager()
+{
+	hash_uninit(fConnectionHash);
+	hash_uninit(fEndpointHash);
+
+	recursive_lock_destroy(&fLock);
+}
+
+
+status_t
+EndpointManager::InitCheck() const
+{
+	if (fConnectionHash == NULL
+		|| fEndpointHash == NULL)
+		return B_NO_MEMORY;
+
+	if (fLock.sem < B_OK)
+		return fLock.sem;
+
+	return B_OK;
+}
+
+
+//	#pragma mark - connections
+
+
+/*!
+	Returns the endpoint matching the connection.
+	You must hold the manager's lock when calling this method.
+*/
+TCPConnection *
+EndpointManager::_LookupConnection(sockaddr *local, sockaddr *peer)
+{
+	connection_key key;
+	key.local = local;
+	key.peer = peer;
+
+	return (TCPConnection *)hash_lookup(fConnectionHash, &key);
+}
+
+
+status_t
+EndpointManager::_RemoveConnection(TCPConnection *endpoint)
+{
+	RecursiveLocker locker(&fLock);
+	return hash_remove(fConnectionHash, endpoint);
+}
+
+
+void
+EndpointManager::_DumpConnections()
+{
+	RecursiveLocker lock(&fLock);
+
+	if (gDomain == NULL)
+		return;
+
+	struct hash_iterator iterator;
+	hash_open(fConnectionHash, &iterator);
+
+	TRACE(("Active TCP Connections:\n"));
+
+	TCPConnection *endpoint;
+	while ((endpoint = (TCPConnection *)hash_next(fConnectionHash, &iterator)) != NULL) {
+		TRACE(("  TCPConnection %p: local %s, peer %s\n", endpoint,
+			AddressString(gDomain, (sockaddr *)&endpoint->socket->address, true).Data(),
+			AddressString(gDomain, (sockaddr *)&endpoint->socket->peer, true).Data()));
+	}
+
+	hash_close(fConnectionHash, &iterator, false);
+}
+
+
+status_t
+EndpointManager::SetConnection(TCPConnection *endpoint,
+	const sockaddr *local, const sockaddr *peer, const sockaddr *interfaceLocal)
+{
+	RecursiveLocker locker(&fLock);
+	sockaddr localBuffer;
+
+	// need to associate this connection with a real address, not INADDR_ANY
+	if (gAddressModule->is_empty_address(local, false)) {
+		gAddressModule->set_to(&localBuffer, interfaceLocal);
+		gAddressModule->set_port(&localBuffer, gAddressModule->get_port(local));
+		local = &localBuffer;
+	}
+
+	connection_key key;
+	key.local = local;
+	key.peer = peer;
+
+	if (hash_lookup(fConnectionHash, &key) != NULL)
+		return EADDRINUSE;
+
+	_RemoveConnection(endpoint);
+
+	gAddressModule->set_to((sockaddr *)&endpoint->socket->address, local);
+	gAddressModule->set_to((sockaddr *)&endpoint->socket->peer, peer);
+
+	return hash_insert(fConnectionHash, endpoint);
+}
+
+
+TCPConnection *
+EndpointManager::FindConnection(sockaddr *local, sockaddr *peer)
+{
+	TCPConnection *endpoint = _LookupConnection(local, peer);
+	if (endpoint != NULL) {
+		TRACE(("TCP: Received packet corresponds to explicit endpoint %p\n", endpoint));
+		return endpoint;
+	}
+
+	// no explicit endpoint exists, check for wildcard endpoints
+
+	sockaddr wildcard;
+	gAddressModule->set_to_empty_address(&wildcard);
+
+	endpoint = _LookupConnection(local, &wildcard);
+	if (endpoint != NULL) {
+		TRACE(("TCP: Received packet corresponds to wildcard endpoint %p\n", endpoint));
+		return endpoint;
+	}
+
+	sockaddr localWildcard;
+	gAddressModule->set_to_empty_address(&localWildcard);
+	gAddressModule->set_port(&localWildcard, gAddressModule->get_port(local));
+
+	endpoint = _LookupConnection(&localWildcard, &wildcard);
+	if (endpoint != NULL) {
+		TRACE(("TCP: Received packet corresponds to local wildcard endpoint %p\n", endpoint));
+		return endpoint;
+	}
+
+	// no matching endpoint exists	
+	TRACE(("TCP: no matching endpoint!\n"));
+	_DumpConnections();
+
+	return NULL;
+}
+
+
+//	#pragma mark - endpoints
+
+
+TCPConnection *
+EndpointManager::_LookupEndpoint(uint16 port)
+{
+	endpoint_key key;
+	key.port = port;
+
+	return (TCPConnection *)hash_lookup(fEndpointHash, &key);
+}
+
+
+status_t
+EndpointManager::Bind(TCPConnection *endpoint, sockaddr *address)
+{
+	if (gAddressModule->is_empty_address(address, true))
+		return B_BAD_VALUE;
+
+	uint16 port = gAddressModule->get_port(address);
+
+	// TODO: check the root group instead?
+	if (ntohs(port) <= kLastReservedPort && geteuid() != 0)
+		return B_PERMISSION_DENIED;
+
+	RecursiveLocker locker(&fLock);
+
+	TCPConnection *first = _LookupEndpoint(port);
+
+	// If there is already an endpoint bound to that port, SO_REUSEADDR has to be
+	// specified by the new endpoint to be allowed to bind to that same port.
+	// Alternatively, all endpoints must have the SO_REUSEPORT option set.
+	if (first != NULL
+		&& (endpoint->socket->options & SO_REUSEADDR) == 0
+		&& ((endpoint->socket->options & SO_REUSEPORT) == 0
+			|| (first->socket->options & SO_REUSEPORT) == 0))
+		return EADDRINUSE;
+
+	if (first != NULL) {
+		TCPConnection *last = first;
+		while (true) {
+			// check if this endpoint binds to a wildcard address
+			if (gAddressModule->is_empty_address((sockaddr *)&last->socket->address, false)) {
+				// you cannot specialize a wildcard endpoint - you have to open the
+				// wildcard endpoint last
+				return B_PERMISSION_DENIED;
+			}
+
+			if (last->fEndpointNextWithSamePort == NULL)
+				break;
+
+			last = last->fEndpointNextWithSamePort;
+		}
+
+		// "first" stays the first item in the list
+		last->fEndpointNextWithSamePort = endpoint;
+	} else
+		hash_insert(fEndpointHash, endpoint);
+
+	endpoint->fEndpointNextWithSamePort = NULL;
+	hash_insert(fConnectionHash, endpoint);
+
+	return B_OK;
+}
+
+
+status_t
+EndpointManager::BindToEphemeral(TCPConnection *endpoint, sockaddr *address)
+{
+	RecursiveLocker locker(&fLock);
+
+	uint32 max = kFirstEphemeralPort + 65536;
+
+	for (int32 i = 1; i < 5; i++) {
+		// try to retrieve a more or less random port
+		uint32 counter = kFirstEphemeralPort;
+		uint32 step = i == 4 ? 1 : system_time() & 0x1f;
+
+		while (counter < max) {
+			uint16 port = counter & 0xffff;
+			if (port <= kLastReservedPort)
+				port += kLastReservedPort;
+
+			port = htons(port);
+
+			TCPConnection *other = _LookupEndpoint(port);
+			if (other == NULL) {
+				// found a port
+				gAddressModule->set_port((sockaddr *)&endpoint->socket->address, port);
+				hash_insert(fEndpointHash, endpoint);
+				hash_insert(fConnectionHash, endpoint);
+				return B_OK;
+			}
+
+			counter += step;
+		}
+	}
+
+	// could not find a port!
+	return EADDRINUSE;
+}
+
+
+status_t
+EndpointManager::Unbind(TCPConnection *endpoint)
+{
+	if (endpoint == NULL || !endpoint->IsBound())
+		return B_BAD_VALUE;
+
+	RecursiveLocker locker(&fLock);
+
+	if (!endpoint->IsBound())
+		return B_BAD_VALUE;
+
+	TCPConnection *other = _LookupEndpoint(gAddressModule->get_port((sockaddr *)&endpoint->socket->address));
+	if (other != endpoint) {
+		// remove endpoint from the list of endpoints with the same port
+		while (other->fEndpointNextWithSamePort != endpoint) {
+			other = other->fEndpointNextWithSamePort;
+		}
+
+		other->fEndpointNextWithSamePort = endpoint->fEndpointNextWithSamePort;
+	} else {
+		// we need to replace the first endpoint in the list
+		hash_remove(fEndpointHash, endpoint);
+
+		other = endpoint->fEndpointNextWithSamePort;
+		if (other != NULL)
+			hash_insert(fEndpointHash, other);
+	}
+
+	endpoint->fEndpointNextWithSamePort = NULL;
+	_RemoveConnection(endpoint);
+
+	endpoint->socket->address.ss_len = 0;
+
+	return B_OK;
+}
+
+
+//	#pragma mark - hash functions
+
+
+/*static*/ int
+EndpointManager::_ConnectionCompare(void *_endpoint, const void *_key)
+{
+	const connection_key *key = (connection_key *)_key;
+	TCPConnection *endpoint = (TCPConnection *)_endpoint;
+
+	if (gAddressModule->equal_addresses_and_ports(key->local,
+			(sockaddr *)&endpoint->socket->address)
+		&& gAddressModule->equal_addresses_and_ports(key->peer,
+			(sockaddr *)&endpoint->socket->peer))
+		return 0;
+
+	return 1;
+}
+
+
+/*static*/ uint32
+EndpointManager::_ConnectionHash(void *_endpoint, const void *_key, uint32 range)
+{
+	const sockaddr *local;
+	const sockaddr *peer;
+
+	if (_endpoint != NULL) {
+		TCPConnection *endpoint = (TCPConnection *)_endpoint;
+		local = (sockaddr *)&endpoint->socket->address;
+		peer = (sockaddr *)&endpoint->socket->peer;
+	} else {
+		const connection_key *key = (connection_key *)_key;
+		local = key->local;
+		peer = key->peer;
+	}
+
+	return gAddressModule->hash_address_pair(local, peer) % range;
+}
+
+
+/*static*/ int
+EndpointManager::_EndpointCompare(void *_endpoint, const void *_key)
+{
+	const endpoint_key *key = (endpoint_key *)_key;
+	TCPConnection *endpoint = (TCPConnection *)_endpoint;
+
+	return gAddressModule->get_port((sockaddr *)&endpoint->socket->address)
+		== key->port ? 0 : 1;
+}
+
+
+/*static*/ uint32
+EndpointManager::_EndpointHash(void *_endpoint, const void *_key, uint32 range)
+{
+	if (_endpoint != NULL) {
+		TCPConnection *endpoint = (TCPConnection *)_endpoint;
+		return gAddressModule->get_port((sockaddr *)&endpoint->socket->address) % range;
+	}
+
+	const endpoint_key *key = (endpoint_key *)_key;
+	return key->port % range;
+}
+

src/add-ons/kernel/network/protocols/tcp/EndpointManager.h

@@ -0,0 +1,53 @@
+/*
+ * Copyright 2006, Haiku, Inc. All Rights Reserved.
+ * Distributed under the terms of the MIT License.
+ *
+ * Authors:
+ *		Axel Dörfler, axeld@pinc-software.de
+ */
+#ifndef ENDPOINT_MANAGER_H
+#define ENDPOINT_MANAGER_H
+
+
+#include <lock.h>
+#include <util/khash.h>
+
+#include <sys/socket.h>
+
+
+class TCPConnection;
+
+class EndpointManager {
+	public:
+		EndpointManager();
+		~EndpointManager();
+
+		status_t InitCheck() const;
+
+		recursive_lock *Locker() { return &fLock; }
+
+		status_t SetConnection(TCPConnection *endpoint, const sockaddr *local,
+			const sockaddr *peer, const sockaddr *interfaceLocal);
+		TCPConnection *FindConnection(sockaddr *local, sockaddr *peer);
+
+		status_t Bind(TCPConnection *endpoint, sockaddr *address);
+		status_t BindToEphemeral(TCPConnection *endpoint, sockaddr *address);
+		status_t Unbind(TCPConnection *endpoint);
+
+	private:
+		TCPConnection *_LookupConnection(sockaddr *local, sockaddr *peer);
+		status_t _RemoveConnection(TCPConnection *endpoint);
+		TCPConnection *_LookupEndpoint(uint16 port);
+		void _DumpConnections();
+
+		static int _ConnectionCompare(void *_endpoint, const void *_key);
+		static uint32 _ConnectionHash(void *_endpoint, const void *_key, uint32 range);
+		static int _EndpointCompare(void *_endpoint, const void *_key);
+		static uint32 _EndpointHash(void *_endpoint, const void *_key, uint32 range);
+
+		hash_table *fConnectionHash;
+		hash_table *fEndpointHash;
+		recursive_lock fLock;
+};
+
+#endif	// ENDPOINT_MANAGER_H

src/add-ons/kernel/network/protocols/tcp/Jamfile

@@ -16,6 +16,7 @@ KernelAddon tcp :
 	tcp.cpp
 	TCPConnection.cpp
 	BufferQueue.cpp
+	EndpointManager.cpp
 ;
 
 # Installation

src/add-ons/kernel/network/protocols/tcp/TCPConnection.cpp

@@ -9,12 +9,19 @@
 
 
 #include "TCPConnection.h"
+#include "EndpointManager.h"
 
 #include <net_buffer.h>
 #include <net_datalink.h>
+#include <NetBufferUtilities.h>
+#include <NetUtilities.h>
+
+#include <lock.h>
+#include <util/AutoLock.h>
+#include <util/khash.h>
+#include <util/list.h>
 
 #include <KernelExport.h>
-#include <util/list.h>
 
 #include <netinet/in.h>
 #include <netinet/ip.h>
@@ -23,13 +30,6 @@
 #include <stdlib.h>
 #include <string.h>
 
-#include <lock.h>
-#include <util/AutoLock.h>
-#include <util/khash.h>
-
-#include <NetBufferUtilities.h>
-#include <NetUtilities.h>
-
 
 // Things this implementation currently doesn't implement:
 // TCP, RFC 793
@@ -197,7 +197,7 @@ TCPConnection::Connect(const struct sockaddr *address)
 	if (address->sa_family != AF_INET)
 		return EAFNOSUPPORT;
 
-	//BenaphoreLocker lock(&fSendLock);
+	RecursiveLocker locker(&fLock);
 
 	TRACE(("  TCP: Connect(): in state %d\n", fState));
 
@@ -218,24 +218,9 @@ TCPConnection::Connect(const struct sockaddr *address)
 			return ENETUNREACH;
 	}
 
-	remove_connection(this);
-		// we need to temporarily remove us from the connection list, as we're
-		// changing our addresses
-
-	// need to associate this connection with a real address, not INADDR_ANY
-	if (gAddressModule->is_empty_address((sockaddr *)&socket->address, false)) {
-		TRACE(("  TCP: Connect(): Local Address is INADDR_ANY\n"));
-		uint16 port = gAddressModule->get_port((sockaddr *)&socket->address);
-		gAddressModule->set_to((sockaddr *)&socket->address,
-			(sockaddr *)fRoute->interface->address);
-		gAddressModule->set_port((sockaddr *)&socket->address, port);
-			// need to reset the port after overwriting the address
-	}
-
-	gAddressModule->set_to((sockaddr *)&socket->peer, address);
-
 	// make sure connection does not already exist
-	status_t status = insert_connection(this);
+	status_t status = gEndpointManager->SetConnection(this,
+		(sockaddr *)&socket->address, address, fRoute->interface->address);
 	if (status < B_OK) {
 		TRACE(("  TCP: Connect(): could not add connection: %s!\n", strerror(status)));
 		return status;
@@ -251,7 +236,6 @@ TCPConnection::Connect(const struct sockaddr *address)
 		&& (0xffffUL << fReceiveWindowShift) < socket->receive.buffer_size) {
 		fReceiveWindowShift++;
 	}
-dprintf("************************* size = %ld, shift = %d\n", socket->receive.buffer_size, fReceiveWindowShift);
 
 	TRACE(("  TCP: Connect(): starting 3-way handshake...\n"));
 
@@ -274,6 +258,8 @@ dprintf("************************* size = %ld, shift = %d\n", socket->receive.bu
 		return EINPROGRESS;
 	}
 
+	locker.Unlock();
+
 	status = acquire_sem_etc(fSendLock, 1, B_RELATIVE_TIMEOUT | B_CAN_INTERRUPT, timeout);
 
 	TRACE(("  TCP: Connect(): Connection complete: %s\n", strerror(status)));
@@ -308,11 +294,7 @@ TCPConnection::Bind(sockaddr *address)
 	TRACE(("TCP:%p.Bind() on address %s\n", this,
 		AddressString(gDomain, address, true).Data()));
 
-	if (address->sa_family != AF_INET)
-		return EAFNOSUPPORT;
-
-	//BenaphoreLocker lock(&fSendLock);
-		// TODO: there is no lock yet for these things...
+	RecursiveLocker lock(fLock);
 
 	if (fState != CLOSED)
 		return EISCONN;
@@ -322,21 +304,10 @@ TCPConnection::Bind(sockaddr *address)
 	if (status < B_OK)
 		return status;
 
-	gAddressModule->set_to((sockaddr *)&socket->address, address);
-
-	if (gAddressModule->get_port((sockaddr *)&socket->address) == 0) {
-		// assign ephemeral port
-		// TODO: use port 40000 and following for now
-		static int e = 40000;
-		gAddressModule->set_port((sockaddr *)&socket->address, htons(e++));
-
-		TRACE(("TCP:%p.Bind() on address with ephemeral port %s, peer %s\n", this,
-			AddressString(gDomain, (sockaddr *)&socket->address, true).Data(),
-			AddressString(gDomain, (sockaddr *)&socket->peer, true).Data()));
-		status = insert_connection(this);
-	} else {
-		status = insert_connection(this);
-	}
+	if (gAddressModule->get_port(address) == 0)
+		status = gEndpointManager->BindToEphemeral(this, address);
+	else
+		status = gEndpointManager->Bind(this, address);
 
 	return status;
 }
@@ -345,18 +316,10 @@ TCPConnection::Bind(sockaddr *address)
 status_t
 TCPConnection::Unbind(struct sockaddr *address)
 {
-	TRACE(("TCP:%p.Unbind()\n", this ));
+	TRACE(("TCP:%p.Unbind()\n", this));
 
-	//BenaphoreLocker lock(&fSendLock);
-		// TODO: there is no lock yet for these things...
-
-	status_t status = remove_connection(this);
-	if (status != B_OK)
-		return status;
-
-	gAddressModule->set_to_empty_address((sockaddr *)&socket->address);
-	gAddressModule->set_port((sockaddr *)&socket->address, 0);
-	return B_OK;
+	RecursiveLocker lock(fLock);
+	return gEndpointManager->Unbind(this);
 }
 
 
@@ -364,9 +327,13 @@ status_t
 TCPConnection::Listen(int count)
 {
 	TRACE(("TCP:%p.Listen()\n", this));
-	//BenaphoreLocker lock(&fSendLock);
+
+	RecursiveLocker lock(fLock);
+
 	if (fState != CLOSED)
 		return B_BAD_VALUE;
+	if (!IsBound())
+		return EDESTADDRREQ;
 
 	fAcceptSemaphore = create_sem(0, "tcp accept");
 	fState = LISTEN;
@@ -504,6 +471,13 @@ TCPConnection::ReadAvailable()
 //	#pragma mark - misc
 
 
+bool
+TCPConnection::IsBound() const
+{
+	return !gAddressModule->is_empty_address((sockaddr *)&socket->address, true);
+}
+
+
 status_t
 TCPConnection::DelayedAcknowledge()
 {
@@ -581,7 +555,8 @@ TCPConnection::ListenReceive(tcp_segment_header &segment, net_buffer *buffer)
 	if (connection->fRoute == NULL)
 		return DROP;
 
-	if (insert_connection(connection) < B_OK)
+	if (gEndpointManager->SetConnection(connection, (sockaddr *)&buffer->destination,
+			(sockaddr *)&buffer->source, NULL) < B_OK)
 		return DROP;
 
 	connection->fInitialReceiveSequence = segment.sequence;
@@ -1041,7 +1016,6 @@ TCPConnection::_SendQueued(bool force)
 	uint32 sendWindow = fSendWindow;
 	uint32 available = fSendQueue.Available(fSendNext);
 	bool outstandingAcknowledge = fSendMax != fSendUnacknowledged;
-dprintf("fSendWindow = %lu, available = %lu, fSendNext = %lu, fSendUnacknowledged = %lu\n", fSendWindow, available, (uint32)fSendNext, (uint32)fSendUnacknowledged);
 
 	if (force && sendWindow == 0 && fSendNext <= fSendQueue.LastSequence()) {
 		// send one byte of data to ask for a window update
@@ -1073,7 +1047,6 @@ dprintf("fSendWindow = %lu, available = %lu, fSendNext = %lu, fSendUnacknowledge
 	segment.urgent_offset = 0;
 
 	while (true) {
-dprintf("length = %ld, segmentLength = %lu\n", length, segmentLength);
 		// Determine if we should really send this segment
 		if (!force && !_ShouldSendSegment(segment, segmentLength, outstandingAcknowledge)) {
 			if (fSendQueue.Available()
@@ -1127,6 +1100,9 @@ dprintf("length = %ld, segmentLength = %lu\n", length, segmentLength);
 			return status;
 		}
 
+		// TODO: we need to trim the segment to the max segment size in case
+		//		the options made it too large
+
 		status = next->module->send_routed_data(next, fRoute, buffer);
 		if (status < B_OK) {
 			gBufferModule->free(buffer);
@@ -1205,35 +1181,3 @@ TCPConnection::_TimeWait(struct net_timer *timer, void *data)
 {
 }
 
-
-//	#pragma mark - hash functions
-
-
-int
-TCPConnection::Compare(void *_connection, const void *_key)
-{
-	const tcp_connection_key *key = (tcp_connection_key *)_key;
-	TCPConnection *connection = ((TCPConnection *)_connection);
-
-	if (gAddressModule->equal_addresses_and_ports(key->local,
-			(sockaddr *)&connection->socket->address)
-		&& gAddressModule->equal_addresses_and_ports(key->peer,
-			(sockaddr *)&connection->socket->peer))
-		return 0;
-
-	return 1;
-}
-
-
-uint32
-TCPConnection::Hash(void *_connection, const void *_key, uint32 range)
-{
-	if (_connection != NULL) {
-		TCPConnection *connection = (TCPConnection *)_connection;
-		return gAddressModule->hash_address_pair((sockaddr *)&connection->socket->address,
-			(sockaddr *)&connection->socket->peer) % range;
-	}
-
-	const tcp_connection_key *key = (tcp_connection_key *)_key;
-	return gAddressModule->hash_address_pair(key->local, key->peer) % range;
-}

src/add-ons/kernel/network/protocols/tcp/TCPConnection.h

@@ -20,6 +20,8 @@
 #include <stddef.h>
 
 
+class EndpointManager;
+
 class TCPConnection : public net_protocol {
 	public:
 		TCPConnection(net_socket *socket);
@@ -44,6 +46,7 @@ class TCPConnection : public net_protocol {
 		size_t ReadAvailable();
 
 		tcp_state State() const { return fState; }
+		bool IsBound() const;
 
 		status_t DelayedAcknowledge();
 		status_t SendAcknowledge();
@@ -53,11 +56,9 @@ class TCPConnection : public net_protocol {
 			net_buffer *buffer);
 		int32 Receive(tcp_segment_header& segment, net_buffer *buffer);
 
-		static int Compare(void *_packet, const void *_key);
-		static uint32 Hash(void *_packet, const void *_key, uint32 range);
-		static int32 HashOffset() { return offsetof(TCPConnection, fHashNext); }
-
 	private:
+		friend class EndpointManager;
+
 		void _StartPersistTimer();
 		uint8 _CurrentFlags();
 		bool _ShouldSendSegment(tcp_segment_header &segment, uint32 length,
@@ -69,7 +70,9 @@ class TCPConnection : public net_protocol {
 		static void _PersistTimer(net_timer *timer, void *data);
 		static void _DelayedAcknowledgeTimer(net_timer *timer, void *data);
 
-		TCPConnection	*fHashNext;
+		TCPConnection	*fConnectionHashNext;
+		TCPConnection	*fEndpointHashNext;
+		TCPConnection	*fEndpointNextWithSamePort;
 
 		recursive_lock	fLock;
 		sem_id			fReceiveLock;

src/add-ons/kernel/network/protocols/tcp/tcp.cpp

@@ -8,6 +8,7 @@
  */
 
 
+#include "EndpointManager.h"
 #include "TCPConnection.h"
 
 #include <net_protocol.h>
@@ -37,45 +38,13 @@
 #endif
 
 
-#define MAX_HASH_TCP	64
-
-
 net_domain *gDomain;
 net_address_module_info *gAddressModule;
 net_buffer_module_info *gBufferModule;
 net_datalink_module_info *gDatalinkModule;
 net_socket_module_info *gSocketModule;
 net_stack_module_info *gStackModule;
-hash_table *gConnectionHash;
-recursive_lock gConnectionLock;
-
-
-#ifdef TRACE_TCP
-#	define DUMP_TCP_HASH tcp_dump_hash()
-// Dumps the TCP Connection hash.  gConnectionLock must NOT be held when calling
-void
-tcp_dump_hash()
-{
-	RecursiveLocker lock(&gConnectionLock);
-	if (gDomain == NULL) {
-		TRACE(("Unable to dump TCP Connections!\n"));
-		return;
-	}
-	struct hash_iterator iterator;
-	hash_open(gConnectionHash, &iterator);
-	TCPConnection *connection;
-	hash_rewind(gConnectionHash, &iterator);
-	TRACE(("Active TCP Connections:\n"));
-	while ((connection = (TCPConnection *)hash_next(gConnectionHash, &iterator)) != NULL) {
-		TRACE(("  TCPConnection %p: local %s, peer %s\n", connection,
-			AddressString(gDomain, (sockaddr *)&connection->socket->address, true).Data(),
-			AddressString(gDomain, (sockaddr *)&connection->socket->peer, true).Data()));
-	}
-	hash_close(gConnectionHash, &iterator, false);
-}
-#else
-#	define DUMP_TCP_HASH ;
-#endif
+EndpointManager *gEndpointManager;
 
 
 status_t
@@ -290,76 +259,6 @@ reply_with_reset(tcp_segment_header &segment, net_buffer *buffer)
 }
 
 
-TCPConnection *
-lookup_connection(sockaddr *local, sockaddr *peer)
-{
-	tcp_connection_key key;
-	key.local = local;
-	key.peer = peer;
-
-	return (TCPConnection *)hash_lookup(gConnectionHash, &key);
-}
-
-
-status_t
-remove_connection(TCPConnection *connection)
-{
-	RecursiveLocker hashLock(&gConnectionLock);
-	return hash_remove(gConnectionHash, connection);
-}
-
-
-status_t
-insert_connection(TCPConnection *connection)
-{
-	RecursiveLocker hashLock(&gConnectionLock);
-
-	tcp_connection_key key;
-	key.local = (sockaddr *)&connection->socket->address;
-	key.peer = (sockaddr *)&connection->socket->peer;
-
-	if (hash_lookup(gConnectionHash, &key) != NULL)
-		return EADDRINUSE;
-
-	return hash_insert(gConnectionHash, connection);
-}
-
-
-TCPConnection *
-find_connection(sockaddr *local, sockaddr *peer)
-{
-	TCPConnection *connection = lookup_connection(local, peer);
-	if (connection != NULL) {
-		TRACE(("TCP: Received packet corresponds to explicit connection %p\n", connection));
-		return connection;
-	}
-
-	// no explicit connection exists, check for wildcard connections
-
-	sockaddr wildcard;
-	gAddressModule->set_to_empty_address(&wildcard);
-
-	connection = lookup_connection(local, &wildcard);
-	if (connection != NULL) {
-		TRACE(("TCP: Received packet corresponds to wildcard connection %p\n", connection));
-		return connection;
-	}
-
-	sockaddr localWildcard;
-	gAddressModule->set_to_empty_address(&localWildcard);
-	gAddressModule->set_port(&localWildcard, gAddressModule->get_port(local));
-
-	connection = lookup_connection(&localWildcard, &wildcard);
-	if (connection != NULL) {
-		TRACE(("TCP: Received packet corresponds to local wildcard connection %p\n", connection));
-		return connection;
-	}
-
-	// no matching connection exists	
-	return NULL;
-}
-
-
 //	#pragma mark - protocol API
 
 
@@ -417,7 +316,6 @@ tcp_free(net_protocol *protocol)
 status_t
 tcp_connect(net_protocol *protocol, const struct sockaddr *address)
 {
-	DUMP_TCP_HASH;
 	return ((TCPConnection *)protocol)->Connect(address);
 }
 
@@ -569,11 +467,11 @@ tcp_receive_data(net_buffer *buffer)
 	bufferHeader.Remove(headerLength);
 		// we no longer need to keep the header around
 
-	RecursiveLocker hashLock(&gConnectionLock);
+	RecursiveLocker locker(gEndpointManager->Locker());
 	int32 segmentAction = DROP;
  
-	TCPConnection *connection = find_connection((struct sockaddr *)&buffer->destination,
-		(struct sockaddr *)&buffer->source);
+	TCPConnection *connection = gEndpointManager->FindConnection(
+		(struct sockaddr *)&buffer->destination, (struct sockaddr *)&buffer->source);
 	if (connection != NULL) {
 		RecursiveLocker locker(connection->Lock());
 
@@ -661,12 +559,12 @@ tcp_init()
 	if (status < B_OK)
 		goto err3;
 
-	gConnectionHash = hash_init(MAX_HASH_TCP, TCPConnection::HashOffset(),
-		&TCPConnection::Compare, &TCPConnection::Hash);
-	if (gConnectionHash == NULL)
+	gEndpointManager = new (std::nothrow) EndpointManager();
+	if (gEndpointManager == NULL) {
+		status = B_NO_MEMORY;
 		goto err4;
-
-	status = recursive_lock_init(&gConnectionLock, "tcp connection hash");
+	}
+	status = gEndpointManager->InitCheck();
 	if (status < B_OK)
 		goto err5;
 
@@ -675,26 +573,24 @@ tcp_init()
 		"network/protocols/ipv4/v1",
 		NULL);
 	if (status < B_OK)
-		goto err6;
+		goto err5;
 
 	status = gStackModule->register_domain_protocols(AF_INET, SOCK_STREAM, IPPROTO_TCP,
 		"network/protocols/tcp/v1",
 		"network/protocols/ipv4/v1",
 		NULL);
 	if (status < B_OK)
-		goto err6;
+		goto err5;
 
 	status = gStackModule->register_domain_receiving_protocol(AF_INET, IPPROTO_TCP,
 		"network/protocols/tcp/v1");
 	if (status < B_OK)
-		goto err6;
+		goto err5;
 
 	return B_OK;
 
-err6:
-	recursive_lock_destroy(&gConnectionLock);
 err5:
-	hash_uninit(gConnectionHash);
+	delete gEndpointManager;
 err4:
 	put_module(NET_DATALINK_MODULE_NAME);
 err3:
@@ -712,8 +608,8 @@ err1:
 static status_t
 tcp_uninit()
 {
-	recursive_lock_destroy(&gConnectionLock);
-	hash_uninit(gConnectionHash);
+	delete gEndpointManager;
+
 	put_module(NET_DATALINK_MODULE_NAME);
 	put_module(NET_SOCKET_MODULE_NAME);
 	put_module(NET_BUFFER_MODULE_NAME);

src/add-ons/kernel/network/protocols/tcp/tcp.h

@@ -21,7 +21,7 @@
 #include <sys/socket.h>
 
 
-class TCPConnection;
+class EndpointManager;
 
 typedef enum {
 	// establishing a connection
@@ -154,11 +154,6 @@ enum tcp_segment_action {
 	IMMEDIATE_ACKNOWLEDGE = 0x08,
 };
 
-struct tcp_connection_key {
-	const sockaddr	*local;
-	const sockaddr	*peer;
-};
-
 
 extern net_domain *gDomain;
 extern net_address_module_info *gAddressModule;
@@ -166,12 +161,9 @@ extern net_buffer_module_info *gBufferModule;
 extern net_datalink_module_info *gDatalinkModule;
 extern net_socket_module_info *gSocketModule;
 extern net_stack_module_info *gStackModule;
-//extern hash_table *gConnectionHash;
-//extern benaphore gConnectionLock;
+extern EndpointManager *gEndpointManager;
 
 
 status_t add_tcp_header(tcp_segment_header &segment, net_buffer *buffer);
-status_t remove_connection(TCPConnection *connection);
-status_t insert_connection(TCPConnection *connection);
 
 #endif	// TCP_H

src/tests/kits/net/Jamfile

@@ -36,6 +36,7 @@ SimpleTest tcp_tester :
 	tcp.cpp
 	TCPConnection.cpp
 	BufferQueue.cpp
+	EndpointManager.cpp
 
 	# misc
 	argv.c
@@ -45,7 +46,7 @@ SimpleTest tcp_tester :
 ;
 
 SEARCH on [ FGristFiles 
-		tcp.cpp TCPConnection.cpp BufferQueue.cpp
+		tcp.cpp TCPConnection.cpp BufferQueue.cpp EndpointManager.cpp
 	] = [ FDirName $(HAIKU_TOP) src add-ons kernel network protocols tcp ] ;
 
 SEARCH on [ FGristFiles