This repository has been archived on 2024-12-14. You can view files and clone it, but cannot push or open issues or pull requests.
PyBitmessage-2024-12-14/src/class_receiveDataThread.py
2015-01-28 14:45:29 -05:00

799 lines
44 KiB
Python

doTimingAttackMitigation = True
import time
import threading
import shared
import hashlib
import socket
import random
from struct import unpack, pack
import sys
import traceback
#import string
#from subprocess import call # used when the API must execute an outside program
#from pyelliptic.openssl import OpenSSL
#import highlevelcrypto
from addresses import *
from helper_generic import addDataPadding, isHostInPrivateIPRange
#import helper_bitcoin
#import helper_inbox
#import helper_sent
from helper_sql import *
#import tr
from debug import logger
# This thread is created either by the synSenderThread(for outgoing
# connections) or the singleListenerThread(for incoming connections).
class receiveDataThread(threading.Thread):
def __init__(self):
threading.Thread.__init__(self)
self.data = ''
self.verackSent = False
self.verackReceived = False
def setup(
self,
sock,
HOST,
port,
streamNumber,
someObjectsOfWhichThisRemoteNodeIsAlreadyAware,
selfInitiatedConnections,
sendDataThreadQueue):
self.sock = sock
self.peer = shared.Peer(HOST, port)
self.streamNumber = streamNumber
self.objectsThatWeHaveYetToGetFromThisPeer = {}
self.selfInitiatedConnections = selfInitiatedConnections
self.sendDataThreadQueue = sendDataThreadQueue # used to send commands and data to the sendDataThread
shared.connectedHostsList[
self.peer.host] = 0 # The very fact that this receiveData thread exists shows that we are connected to the remote host. Let's add it to this list so that an outgoingSynSender thread doesn't try to connect to it.
self.connectionIsOrWasFullyEstablished = False # set to true after the remote node and I accept each other's version messages. This is needed to allow the user interface to accurately reflect the current number of connections.
if self.streamNumber == -1: # This was an incoming connection. Send out a version message if we accept the other node's version message.
self.initiatedConnection = False
else:
self.initiatedConnection = True
self.selfInitiatedConnections[streamNumber][self] = 0
self.someObjectsOfWhichThisRemoteNodeIsAlreadyAware = someObjectsOfWhichThisRemoteNodeIsAlreadyAware
def run(self):
with shared.printLock:
print 'receiveDataThread starting. ID', str(id(self)) + '. The size of the shared.connectedHostsList is now', len(shared.connectedHostsList)
while True:
if shared.config.getint('bitmessagesettings', 'maxdownloadrate') == 0:
downloadRateLimitBytes = float("inf")
else:
downloadRateLimitBytes = shared.config.getint('bitmessagesettings', 'maxdownloadrate') * 1000
with shared.receiveDataLock:
while shared.numberOfBytesReceivedLastSecond >= downloadRateLimitBytes:
if int(time.time()) == shared.lastTimeWeResetBytesReceived:
# If it's still the same second that it was last time then sleep.
time.sleep(0.3)
else:
# It's a new second. Let us clear the shared.numberOfBytesReceivedLastSecond.
shared.lastTimeWeResetBytesReceived = int(time.time())
shared.numberOfBytesReceivedLastSecond = 0
dataLen = len(self.data)
try:
dataRecv = self.sock.recv(1024)
self.data += dataRecv
shared.numberOfBytesReceived += len(dataRecv) # for the 'network status' UI tab. The UI clears this value whenever it updates.
shared.numberOfBytesReceivedLastSecond += len(dataRecv) # for the download rate limit
except socket.timeout:
with shared.printLock:
print 'Timeout occurred waiting for data from', self.peer, '. Closing receiveData thread. (ID:', str(id(self)) + ')'
break
except Exception as err:
with shared.printLock:
print 'sock.recv error. Closing receiveData thread (' + str(self.peer) + ', Thread ID:', str(id(self)) + ').', err
break
# print 'Received', repr(self.data)
if len(self.data) == dataLen: # If self.sock.recv returned no data:
with shared.printLock:
print 'Connection to', self.peer, 'closed. Closing receiveData thread. (ID:', str(id(self)) + ')'
break
else:
self.processData()
try:
del self.selfInitiatedConnections[self.streamNumber][self]
with shared.printLock:
print 'removed self (a receiveDataThread) from selfInitiatedConnections'
except:
pass
self.sendDataThreadQueue.put((0, 'shutdown','no data')) # commands the corresponding sendDataThread to shut itself down.
try:
del shared.connectedHostsList[self.peer.host]
except Exception as err:
with shared.printLock:
print 'Could not delete', self.peer.host, 'from shared.connectedHostsList.', err
try:
del shared.numberOfObjectsThatWeHaveYetToGetPerPeer[
self.peer]
except:
pass
shared.UISignalQueue.put(('updateNetworkStatusTab', 'no data'))
with shared.printLock:
print 'receiveDataThread ending. ID', str(id(self)) + '. The size of the shared.connectedHostsList is now', len(shared.connectedHostsList)
def processData(self):
if len(self.data) < shared.Header.size: # if so little of the data has arrived that we can't even read the checksum then wait for more data.
return
magic,command,payloadLength,checksum = shared.Header.unpack(self.data[:shared.Header.size])
if magic != 0xE9BEB4D9:
self.data = ""
return
if payloadLength > 1600100: # ~1.6 MB which is the maximum possible size of an inv message.
logger.info('The incoming message, which we have not yet download, is too large. Ignoring it. (unfortunately there is no way to tell the other node to stop sending it except to disconnect.) Message size: %s' % payloadLength)
self.data = self.data[payloadLength + shared.Header.size:]
del magic,command,payloadLength,checksum # we don't need these anymore and better to clean them now before the recursive call rather than after
self.processData()
return
if len(self.data) < payloadLength + shared.Header.size: # check if the whole message has arrived yet.
return
payload = self.data[shared.Header.size:payloadLength + shared.Header.size]
if checksum != hashlib.sha512(payload).digest()[0:4]: # test the checksum in the message.
print 'Checksum incorrect. Clearing this message.'
self.data = self.data[payloadLength + shared.Header.size:]
del magic,command,payloadLength,checksum,payload # better to clean up before the recursive call
self.processData()
return
# The time we've last seen this node is obviously right now since we
# just received valid data from it. So update the knownNodes list so
# that other peers can be made aware of its existance.
if self.initiatedConnection and self.connectionIsOrWasFullyEstablished: # The remote port is only something we should share with others if it is the remote node's incoming port (rather than some random operating-system-assigned outgoing port).
shared.knownNodesLock.acquire()
shared.knownNodes[self.streamNumber][self.peer] = int(time.time())
shared.knownNodesLock.release()
#Strip the nulls
command = command.rstrip('\x00')
with shared.printLock:
print 'remoteCommand', repr(command), ' from', self.peer
try:
#TODO: Use a dispatcher here
if command == 'error':
self.recerror(payload)
elif not self.connectionIsOrWasFullyEstablished:
if command == 'version':
self.recversion(payload)
elif command == 'verack':
self.recverack()
else:
if command == 'addr':
self.recaddr(payload)
elif command == 'inv':
self.recinv(payload)
elif command == 'getdata':
self.recgetdata(payload)
elif command == 'object':
self.recobject(payload)
elif command == 'ping':
self.sendpong(payload)
#elif command == 'pong':
# pass
except varintDecodeError as e:
logger.debug("There was a problem with a varint while processing a message from the wire. Some details: %s" % e)
except Exception as e:
logger.critical("Critical error in a receiveDataThread: \n%s" % traceback.format_exc())
del payload
self.data = self.data[payloadLength + shared.Header.size:] # take this message out and then process the next message
if self.data == '': # if there are no more messages
while len(self.objectsThatWeHaveYetToGetFromThisPeer) > 0:
shared.numberOfInventoryLookupsPerformed += 1
objectHash, = random.sample(
self.objectsThatWeHaveYetToGetFromThisPeer, 1)
if objectHash in shared.inventory:
with shared.printLock:
print 'Inventory (in memory) already has object listed in inv message.'
del self.objectsThatWeHaveYetToGetFromThisPeer[
objectHash]
elif shared.isInSqlInventory(objectHash):
if shared.verbose >= 3:
with shared.printLock:
print 'Inventory (SQL on disk) already has object listed in inv message.'
del self.objectsThatWeHaveYetToGetFromThisPeer[
objectHash]
else:
# We don't have the object in our inventory. Let's request it.
self.sendgetdata(objectHash)
del self.objectsThatWeHaveYetToGetFromThisPeer[
objectHash] # It is possible that the remote node might not respond with the object. In that case, we'll very likely get it from someone else anyway.
if len(self.objectsThatWeHaveYetToGetFromThisPeer) == 0:
with shared.printLock:
print '(concerning', str(self.peer) + ')', 'number of objectsThatWeHaveYetToGetFromThisPeer is now 0'
try:
del shared.numberOfObjectsThatWeHaveYetToGetPerPeer[
self.peer] # this data structure is maintained so that we can keep track of how many total objects, across all connections, are currently outstanding. If it goes too high it can indicate that we are under attack by multiple nodes working together.
except:
pass
break
if len(self.objectsThatWeHaveYetToGetFromThisPeer) == 0:
# We had objectsThatWeHaveYetToGetFromThisPeer but the loop ran, they were all in our inventory, and now we don't have any to get anymore.
with shared.printLock:
print '(concerning', str(self.peer) + ')', 'number of objectsThatWeHaveYetToGetFromThisPeer is now 0'
try:
del shared.numberOfObjectsThatWeHaveYetToGetPerPeer[
self.peer] # this data structure is maintained so that we can keep track of how many total objects, across all connections, are currently outstanding. If it goes too high it can indicate that we are under attack by multiple nodes working together.
except:
pass
if len(self.objectsThatWeHaveYetToGetFromThisPeer) > 0:
with shared.printLock:
print '(concerning', str(self.peer) + ')', 'number of objectsThatWeHaveYetToGetFromThisPeer is now', len(self.objectsThatWeHaveYetToGetFromThisPeer)
shared.numberOfObjectsThatWeHaveYetToGetPerPeer[self.peer] = len(
self.objectsThatWeHaveYetToGetFromThisPeer) # this data structure is maintained so that we can keep track of how many total objects, across all connections, are currently outstanding. If it goes too high it can indicate that we are under attack by multiple nodes working together.
self.processData()
def sendpong(self):
with shared.printLock:
print 'Sending pong'
self.sendDataThreadQueue.put((0, 'sendRawData', shared.CreatePacket('pong')))
def recverack(self):
with shared.printLock:
print 'verack received'
self.verackReceived = True
if self.verackSent:
# We have thus both sent and received a verack.
self.connectionFullyEstablished()
def connectionFullyEstablished(self):
if self.connectionIsOrWasFullyEstablished:
# there is no reason to run this function a second time
return
self.connectionIsOrWasFullyEstablished = True
# Command the corresponding sendDataThread to set its own connectionIsOrWasFullyEstablished variable to True also
self.sendDataThreadQueue.put((0, 'connectionIsOrWasFullyEstablished', 'no data'))
if not self.initiatedConnection:
shared.clientHasReceivedIncomingConnections = True
shared.UISignalQueue.put(('setStatusIcon', 'green'))
self.sock.settimeout(
600) # We'll send out a pong every 5 minutes to make sure the connection stays alive if there has been no other traffic to send lately.
shared.UISignalQueue.put(('updateNetworkStatusTab', 'no data'))
with shared.printLock:
print 'Connection fully established with', self.peer
print 'The size of the connectedHostsList is now', len(shared.connectedHostsList)
print 'The length of sendDataQueues is now:', len(shared.sendDataQueues)
print 'broadcasting addr from within connectionFullyEstablished function.'
# Let all of our peers know about this new node.
dataToSend = (int(time.time()), self.streamNumber, 1, self.peer.host, self.remoteNodeIncomingPort)
shared.broadcastToSendDataQueues((
self.streamNumber, 'advertisepeer', dataToSend))
self.sendaddr() # This is one large addr message to this one peer.
if not self.initiatedConnection and len(shared.connectedHostsList) > 200:
with shared.printLock:
print 'We are connected to too many people. Closing connection.'
self.sendDataThreadQueue.put((0, 'shutdown','no data'))
return
self.sendBigInv()
def sendBigInv(self):
# Select all hashes for objects in this stream.
queryreturn = sqlQuery(
'''SELECT hash FROM inventory WHERE expirestime>? and streamnumber=?''',
int(time.time()),
self.streamNumber)
bigInvList = {}
for row in queryreturn:
hash, = row
if hash not in self.someObjectsOfWhichThisRemoteNodeIsAlreadyAware:
bigInvList[hash] = 0
# We also have messages in our inventory in memory (which is a python
# dictionary). Let's fetch those too.
with shared.inventoryLock:
for hash, storedValue in shared.inventory.items():
if hash not in self.someObjectsOfWhichThisRemoteNodeIsAlreadyAware:
objectType, streamNumber, payload, expiresTime, tag = storedValue
if streamNumber == self.streamNumber and expiresTime > int(time.time()):
bigInvList[hash] = 0
numberOfObjectsInInvMessage = 0
payload = ''
# Now let us start appending all of these hashes together. They will be
# sent out in a big inv message to our new peer.
for hash, storedValue in bigInvList.items():
payload += hash
numberOfObjectsInInvMessage += 1
if numberOfObjectsInInvMessage == 50000: # We can only send a max of 50000 items per inv message but we may have more objects to advertise. They must be split up into multiple inv messages.
self.sendinvMessageToJustThisOnePeer(
numberOfObjectsInInvMessage, payload)
payload = ''
numberOfObjectsInInvMessage = 0
if numberOfObjectsInInvMessage > 0:
self.sendinvMessageToJustThisOnePeer(
numberOfObjectsInInvMessage, payload)
# Used to send a big inv message when the connection with a node is
# first fully established. Notice that there is also a broadcastinv
# function for broadcasting invs to everyone in our stream.
def sendinvMessageToJustThisOnePeer(self, numberOfObjects, payload):
payload = encodeVarint(numberOfObjects) + payload
with shared.printLock:
print 'Sending huge inv message with', numberOfObjects, 'objects to just this one peer'
self.sendDataThreadQueue.put((0, 'sendRawData', shared.CreatePacket('inv', payload)))
def _sleepForTimingAttackMitigation(self, sleepTime):
# We don't need to do the timing attack mitigation if we are
# only connected to the trusted peer because we can trust the
# peer not to attack
if sleepTime > 0 and doTimingAttackMitigation and shared.trustedPeer == None:
with shared.printLock:
print 'Timing attack mitigation: Sleeping for', sleepTime, 'seconds.'
time.sleep(sleepTime)
def recerror(self, data):
"""
The remote node has been polite enough to send you an error message.
"""
fatalStatus, readPosition = decodeVarint(data[:10])
banTime, banTimeLength = decodeVarint(data[readPosition:readPosition+10])
readPosition += banTimeLength
inventoryVectorLength, inventoryVectorLengthLength = decodeVarint(data[readPosition:readPosition+10])
if inventoryVectorLength > 100:
return
readPosition += inventoryVectorLengthLength
inventoryVector = data[readPosition:readPosition+inventoryVectorLength]
readPosition += inventoryVectorLength
errorTextLength, errorTextLengthLength = decodeVarint(data[readPosition:readPosition+10])
if errorTextLength > 1000:
return
readPosition += errorTextLengthLength
errorText = data[readPosition:readPosition+errorTextLength]
if fatalStatus == 0:
fatalHumanFriendly = 'Warning'
elif fatalStatus == 1:
fatalHumanFriendly = 'Error'
elif fatalStatus == 2:
fatalHumanFriendly = 'Fatal'
message = '%s message received from %s: %s.' % (fatalHumanFriendly, self.peer, errorText)
if inventoryVector:
message += " This concerns object %s" % inventoryVector.encode('hex')
if banTime > 0:
message += " Remote node says that the ban time is %s" % banTime
logger.error(message)
def recobject(self, data):
self.messageProcessingStartTime = time.time()
lengthOfTimeWeShouldUseToProcessThisMessage = shared.checkAndShareObjectWithPeers(data)
"""
Sleeping will help guarantee that we can process messages faster than a
remote node can send them. If we fall behind, the attacker could observe
that we are are slowing down the rate at which we request objects from the
network which would indicate that we own a particular address (whichever
one to which they are sending all of their attack messages). Note
that if an attacker connects to a target with many connections, this
mitigation mechanism might not be sufficient.
"""
sleepTime = lengthOfTimeWeShouldUseToProcessThisMessage - (time.time() - self.messageProcessingStartTime)
self._sleepForTimingAttackMitigation(sleepTime)
# We have received an inv message
def recinv(self, data):
totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers = 0 # this counts duplicates separately because they take up memory
if len(shared.numberOfObjectsThatWeHaveYetToGetPerPeer) > 0:
for key, value in shared.numberOfObjectsThatWeHaveYetToGetPerPeer.items():
totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers += value
with shared.printLock:
print 'number of keys(hosts) in shared.numberOfObjectsThatWeHaveYetToGetPerPeer:', len(shared.numberOfObjectsThatWeHaveYetToGetPerPeer)
print 'totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers = ', totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers
numberOfItemsInInv, lengthOfVarint = decodeVarint(data[:10])
if numberOfItemsInInv > 50000:
sys.stderr.write('Too many items in inv message!')
return
if len(data) < lengthOfVarint + (numberOfItemsInInv * 32):
print 'inv message doesn\'t contain enough data. Ignoring.'
return
if numberOfItemsInInv == 1: # we'll just request this data from the person who advertised the object.
if totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers > 200000 and len(self.objectsThatWeHaveYetToGetFromThisPeer) > 1000: # inv flooding attack mitigation
with shared.printLock:
print 'We already have', totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers, 'items yet to retrieve from peers and over 1000 from this node in particular. Ignoring this inv message.'
return
self.someObjectsOfWhichThisRemoteNodeIsAlreadyAware[
data[lengthOfVarint:32 + lengthOfVarint]] = 0
shared.numberOfInventoryLookupsPerformed += 1
if data[lengthOfVarint:32 + lengthOfVarint] in shared.inventory:
with shared.printLock:
print 'Inventory (in memory) has inventory item already.'
elif shared.isInSqlInventory(data[lengthOfVarint:32 + lengthOfVarint]):
print 'Inventory (SQL on disk) has inventory item already.'
else:
self.sendgetdata(data[lengthOfVarint:32 + lengthOfVarint])
else:
# There are many items listed in this inv message. Let us create a
# 'set' of objects we are aware of and a set of objects in this inv
# message so that we can diff one from the other cheaply.
startTime = time.time()
advertisedSet = set()
for i in range(numberOfItemsInInv):
advertisedSet.add(data[lengthOfVarint + (32 * i):32 + lengthOfVarint + (32 * i)])
objectsNewToMe = advertisedSet - shared.inventorySets[self.streamNumber]
logger.info('inv message lists %s objects. Of those %s are new to me. It took %s seconds to figure that out.', numberOfItemsInInv, len(objectsNewToMe), time.time()-startTime)
for item in objectsNewToMe:
if totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers > 200000 and len(self.objectsThatWeHaveYetToGetFromThisPeer) > 1000: # inv flooding attack mitigation
with shared.printLock:
print 'We already have', totalNumberOfobjectsThatWeHaveYetToGetFromAllPeers, 'items yet to retrieve from peers and over', len(self.objectsThatWeHaveYetToGetFromThisPeer), 'from this node in particular. Ignoring the rest of this inv message.'
break
self.someObjectsOfWhichThisRemoteNodeIsAlreadyAware[item] = 0 # helps us keep from sending inv messages to peers that already know about the objects listed therein
self.objectsThatWeHaveYetToGetFromThisPeer[item] = 0 # upon finishing dealing with an incoming message, the receiveDataThread will request a random object of from peer out of this data structure. This way if we get multiple inv messages from multiple peers which list mostly the same objects, we will make getdata requests for different random objects from the various peers.
if len(self.objectsThatWeHaveYetToGetFromThisPeer) > 0:
shared.numberOfObjectsThatWeHaveYetToGetPerPeer[
self.peer] = len(self.objectsThatWeHaveYetToGetFromThisPeer)
# Send a getdata message to our peer to request the object with the given
# hash
def sendgetdata(self, hash):
with shared.printLock:
print 'sending getdata to retrieve object with hash:', hash.encode('hex')
payload = '\x01' + hash
self.sendDataThreadQueue.put((0, 'sendRawData', shared.CreatePacket('getdata', payload)))
# We have received a getdata request from our peer
def recgetdata(self, data):
numberOfRequestedInventoryItems, lengthOfVarint = decodeVarint(
data[:10])
if len(data) < lengthOfVarint + (32 * numberOfRequestedInventoryItems):
print 'getdata message does not contain enough data. Ignoring.'
return
for i in xrange(numberOfRequestedInventoryItems):
hash = data[lengthOfVarint + (
i * 32):32 + lengthOfVarint + (i * 32)]
with shared.printLock:
print 'received getdata request for item:', hash.encode('hex')
shared.numberOfInventoryLookupsPerformed += 1
shared.inventoryLock.acquire()
if hash in shared.inventory:
objectType, streamNumber, payload, expiresTime, tag = shared.inventory[hash]
shared.inventoryLock.release()
self.sendObject(payload)
else:
shared.inventoryLock.release()
queryreturn = sqlQuery(
'''select payload from inventory where hash=? and expirestime>=?''',
hash,
int(time.time()))
if queryreturn != []:
for row in queryreturn:
payload, = row
self.sendObject(payload)
else:
logger.warning('%s asked for an object with a getdata which is not in either our memory inventory or our SQL inventory. We probably cleaned it out after advertising it but before they got around to asking for it.' % (self.peer,))
# Our peer has requested (in a getdata message) that we send an object.
def sendObject(self, payload):
with shared.printLock:
print 'sending an object.'
self.sendDataThreadQueue.put((0, 'sendRawData', shared.CreatePacket('object',payload)))
def _checkIPv4Address(self, host, hostFromAddrMessage):
# print 'hostFromAddrMessage', hostFromAddrMessage
if host[0] == '\x7F':
print 'Ignoring IP address in loopback range:', hostFromAddrMessage
return False
if host[0] == '\x0A':
print 'Ignoring IP address in private range:', hostFromAddrMessage
return False
if host[0:2] == '\xC0\xA8':
print 'Ignoring IP address in private range:', hostFromAddrMessage
return False
return True
def _checkIPv6Address(self, host, hostFromAddrMessage):
if host == ('\x00' * 15) + '\x01':
print 'Ignoring loopback address:', hostFromAddrMessage
return False
if host[0] == '\xFE' and (ord(host[1]) & 0xc0) == 0x80:
print 'Ignoring local address:', hostFromAddrMessage
return False
if (ord(host[0]) & 0xfe) == 0xfc:
print 'Ignoring unique local address:', hostFromAddrMessage
return False
return True
# We have received an addr message.
def recaddr(self, data):
#listOfAddressDetailsToBroadcastToPeers = []
numberOfAddressesIncluded = 0
numberOfAddressesIncluded, lengthOfNumberOfAddresses = decodeVarint(
data[:10])
if shared.verbose >= 1:
with shared.printLock:
print 'addr message contains', numberOfAddressesIncluded, 'IP addresses.'
if numberOfAddressesIncluded > 1000 or numberOfAddressesIncluded == 0:
return
if len(data) != lengthOfNumberOfAddresses + (38 * numberOfAddressesIncluded):
print 'addr message does not contain the correct amount of data. Ignoring.'
return
for i in range(0, numberOfAddressesIncluded):
try:
fullHost = data[20 + lengthOfNumberOfAddresses + (38 * i):36 + lengthOfNumberOfAddresses + (38 * i)]
except Exception as err:
with shared.printLock:
sys.stderr.write(
'ERROR TRYING TO UNPACK recaddr (recaddrHost). Message: %s\n' % str(err))
break # giving up on unpacking any more. We should still be connected however.
try:
recaddrStream, = unpack('>I', data[8 + lengthOfNumberOfAddresses + (
38 * i):12 + lengthOfNumberOfAddresses + (38 * i)])
except Exception as err:
with shared.printLock:
sys.stderr.write(
'ERROR TRYING TO UNPACK recaddr (recaddrStream). Message: %s\n' % str(err))
break # giving up on unpacking any more. We should still be connected however.
if recaddrStream == 0:
continue
if recaddrStream != self.streamNumber and recaddrStream != (self.streamNumber * 2) and recaddrStream != ((self.streamNumber * 2) + 1): # if the embedded stream number is not in my stream or either of my child streams then ignore it. Someone might be trying funny business.
continue
try:
recaddrServices, = unpack('>Q', data[12 + lengthOfNumberOfAddresses + (
38 * i):20 + lengthOfNumberOfAddresses + (38 * i)])
except Exception as err:
with shared.printLock:
sys.stderr.write(
'ERROR TRYING TO UNPACK recaddr (recaddrServices). Message: %s\n' % str(err))
break # giving up on unpacking any more. We should still be connected however.
try:
recaddrPort, = unpack('>H', data[36 + lengthOfNumberOfAddresses + (
38 * i):38 + lengthOfNumberOfAddresses + (38 * i)])
except Exception as err:
with shared.printLock:
sys.stderr.write(
'ERROR TRYING TO UNPACK recaddr (recaddrPort). Message: %s\n' % str(err))
break # giving up on unpacking any more. We should still be connected however.
# print 'Within recaddr(): IP', recaddrIP, ', Port',
# recaddrPort, ', i', i
if fullHost[0:12] == '\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\xFF\xFF':
ipv4Host = fullHost[12:]
hostFromAddrMessage = socket.inet_ntop(socket.AF_INET, ipv4Host)
if not self._checkIPv4Address(ipv4Host, hostFromAddrMessage):
continue
else:
hostFromAddrMessage = socket.inet_ntop(socket.AF_INET6, fullHost)
if hostFromAddrMessage == "":
# This can happen on Windows systems which are not 64-bit compatible
# so let us drop the IPv6 address.
continue
if not self._checkIPv6Address(fullHost, hostFromAddrMessage):
continue
timeSomeoneElseReceivedMessageFromThisNode, = unpack('>Q', data[lengthOfNumberOfAddresses + (
38 * i):8 + lengthOfNumberOfAddresses + (38 * i)]) # This is the 'time' value in the received addr message. 64-bit.
if recaddrStream not in shared.knownNodes: # knownNodes is a dictionary of dictionaries with one outer dictionary for each stream. If the outer stream dictionary doesn't exist yet then we must make it.
shared.knownNodesLock.acquire()
shared.knownNodes[recaddrStream] = {}
shared.knownNodesLock.release()
peerFromAddrMessage = shared.Peer(hostFromAddrMessage, recaddrPort)
if peerFromAddrMessage not in shared.knownNodes[recaddrStream]:
if len(shared.knownNodes[recaddrStream]) < 20000 and timeSomeoneElseReceivedMessageFromThisNode > (int(time.time()) - 10800) and timeSomeoneElseReceivedMessageFromThisNode < (int(time.time()) + 10800): # If we have more than 20000 nodes in our list already then just forget about adding more. Also, make sure that the time that someone else received a message from this node is within three hours from now.
shared.knownNodesLock.acquire()
shared.knownNodes[recaddrStream][peerFromAddrMessage] = (
timeSomeoneElseReceivedMessageFromThisNode)
shared.knownNodesLock.release()
with shared.printLock:
print 'added new node', peerFromAddrMessage, 'to knownNodes in stream', recaddrStream
shared.needToWriteKnownNodesToDisk = True
hostDetails = (
timeSomeoneElseReceivedMessageFromThisNode,
recaddrStream, recaddrServices, hostFromAddrMessage, recaddrPort)
#listOfAddressDetailsToBroadcastToPeers.append(hostDetails)
shared.broadcastToSendDataQueues((
self.streamNumber, 'advertisepeer', hostDetails))
else:
timeLastReceivedMessageFromThisNode = shared.knownNodes[recaddrStream][
peerFromAddrMessage] # PORT in this case is either the port we used to connect to the remote node, or the port that was specified by someone else in a past addr message.
if (timeLastReceivedMessageFromThisNode < timeSomeoneElseReceivedMessageFromThisNode) and (timeSomeoneElseReceivedMessageFromThisNode < int(time.time())):
shared.knownNodesLock.acquire()
shared.knownNodes[recaddrStream][peerFromAddrMessage] = timeSomeoneElseReceivedMessageFromThisNode
shared.knownNodesLock.release()
#if listOfAddressDetailsToBroadcastToPeers != []:
# self.broadcastaddr(listOfAddressDetailsToBroadcastToPeers)
with shared.printLock:
print 'knownNodes currently has', len(shared.knownNodes[self.streamNumber]), 'nodes for this stream.'
# Send a huge addr message to our peer. This is only used
# when we fully establish a connection with a
# peer (with the full exchange of version and verack
# messages).
def sendaddr(self):
addrsInMyStream = {}
addrsInChildStreamLeft = {}
addrsInChildStreamRight = {}
# print 'knownNodes', shared.knownNodes
# We are going to share a maximum number of 1000 addrs with our peer.
# 500 from this stream, 250 from the left child stream, and 250 from
# the right child stream.
shared.knownNodesLock.acquire()
if len(shared.knownNodes[self.streamNumber]) > 0:
for i in range(500):
peer, = random.sample(shared.knownNodes[self.streamNumber], 1)
if isHostInPrivateIPRange(peer.host):
continue
addrsInMyStream[peer] = shared.knownNodes[
self.streamNumber][peer]
if len(shared.knownNodes[self.streamNumber * 2]) > 0:
for i in range(250):
peer, = random.sample(shared.knownNodes[
self.streamNumber * 2], 1)
if isHostInPrivateIPRange(peer.host):
continue
addrsInChildStreamLeft[peer] = shared.knownNodes[
self.streamNumber * 2][peer]
if len(shared.knownNodes[(self.streamNumber * 2) + 1]) > 0:
for i in range(250):
peer, = random.sample(shared.knownNodes[
(self.streamNumber * 2) + 1], 1)
if isHostInPrivateIPRange(peer.host):
continue
addrsInChildStreamRight[peer] = shared.knownNodes[
(self.streamNumber * 2) + 1][peer]
shared.knownNodesLock.release()
numberOfAddressesInAddrMessage = 0
payload = ''
# print 'addrsInMyStream.items()', addrsInMyStream.items()
for (HOST, PORT), value in addrsInMyStream.items():
timeLastReceivedMessageFromThisNode = value
if timeLastReceivedMessageFromThisNode > (int(time.time()) - shared.maximumAgeOfNodesThatIAdvertiseToOthers): # If it is younger than 3 hours old..
numberOfAddressesInAddrMessage += 1
payload += pack(
'>Q', timeLastReceivedMessageFromThisNode) # 64-bit time
payload += pack('>I', self.streamNumber)
payload += pack(
'>q', 1) # service bit flags offered by this node
payload += shared.encodeHost(HOST)
payload += pack('>H', PORT) # remote port
for (HOST, PORT), value in addrsInChildStreamLeft.items():
timeLastReceivedMessageFromThisNode = value
if timeLastReceivedMessageFromThisNode > (int(time.time()) - shared.maximumAgeOfNodesThatIAdvertiseToOthers): # If it is younger than 3 hours old..
numberOfAddressesInAddrMessage += 1
payload += pack(
'>Q', timeLastReceivedMessageFromThisNode) # 64-bit time
payload += pack('>I', self.streamNumber * 2)
payload += pack(
'>q', 1) # service bit flags offered by this node
payload += shared.encodeHost(HOST)
payload += pack('>H', PORT) # remote port
for (HOST, PORT), value in addrsInChildStreamRight.items():
timeLastReceivedMessageFromThisNode = value
if timeLastReceivedMessageFromThisNode > (int(time.time()) - shared.maximumAgeOfNodesThatIAdvertiseToOthers): # If it is younger than 3 hours old..
numberOfAddressesInAddrMessage += 1
payload += pack(
'>Q', timeLastReceivedMessageFromThisNode) # 64-bit time
payload += pack('>I', (self.streamNumber * 2) + 1)
payload += pack(
'>q', 1) # service bit flags offered by this node
payload += shared.encodeHost(HOST)
payload += pack('>H', PORT) # remote port
payload = encodeVarint(numberOfAddressesInAddrMessage) + payload
self.sendDataThreadQueue.put((0, 'sendRawData', shared.CreatePacket('addr', payload)))
# We have received a version message
def recversion(self, data):
if len(data) < 83:
# This version message is unreasonably short. Forget it.
return
if self.verackSent:
"""
We must have already processed the remote node's version message.
There might be a time in the future when we Do want to process
a new version message, like if the remote node wants to update
the streams in which they are interested. But for now we'll
ignore this version message
"""
return
self.remoteProtocolVersion, = unpack('>L', data[:4])
if self.remoteProtocolVersion < 3:
self.sendDataThreadQueue.put((0, 'shutdown','no data'))
with shared.printLock:
print 'Closing connection to old protocol version', self.remoteProtocolVersion, 'node: ', self.peer
return
timestamp, = unpack('>Q', data[12:20])
timeOffset = timestamp - int(time.time())
if timeOffset > 3600:
self.sendDataThreadQueue.put((0, 'sendRawData', shared.assembleErrorMessage(fatal=2, errorText="Your time is too far in the future compared to mine. Closing connection.")))
logger.info("%s's time is too far in the future (%s seconds). Closing connection to it." % (self.peer, timeOffset))
time.sleep(2)
self.sendDataThreadQueue.put((0, 'shutdown','no data'))
return
if timeOffset < -3600:
self.sendDataThreadQueue.put((0, 'sendRawData', shared.assembleErrorMessage(fatal=2, errorText="Your time is too far in the past compared to mine. Closing connection.")))
logger.info("%s's time is too far in the past (timeOffset %s seconds). Closing connection to it." % (self.peer, timeOffset))
time.sleep(2)
self.sendDataThreadQueue.put((0, 'shutdown','no data'))
return
self.myExternalIP = socket.inet_ntoa(data[40:44])
# print 'myExternalIP', self.myExternalIP
self.remoteNodeIncomingPort, = unpack('>H', data[70:72])
# print 'remoteNodeIncomingPort', self.remoteNodeIncomingPort
useragentLength, lengthOfUseragentVarint = decodeVarint(
data[80:84])
readPosition = 80 + lengthOfUseragentVarint
useragent = data[readPosition:readPosition + useragentLength]
readPosition += useragentLength
numberOfStreamsInVersionMessage, lengthOfNumberOfStreamsInVersionMessage = decodeVarint(
data[readPosition:])
readPosition += lengthOfNumberOfStreamsInVersionMessage
self.streamNumber, lengthOfRemoteStreamNumber = decodeVarint(
data[readPosition:])
with shared.printLock:
print 'Remote node useragent:', useragent, ' stream number:', self.streamNumber, ' time offset:', timeOffset, 'seconds.'
if self.streamNumber != 1:
self.sendDataThreadQueue.put((0, 'shutdown','no data'))
with shared.printLock:
print 'Closed connection to', self.peer, 'because they are interested in stream', self.streamNumber, '.'
return
shared.connectedHostsList[
self.peer.host] = 1 # We use this data structure to not only keep track of what hosts we are connected to so that we don't try to connect to them again, but also to list the connections count on the Network Status tab.
# If this was an incoming connection, then the sendDataThread
# doesn't know the stream. We have to set it.
if not self.initiatedConnection:
self.sendDataThreadQueue.put((0, 'setStreamNumber', self.streamNumber))
if data[72:80] == shared.eightBytesOfRandomDataUsedToDetectConnectionsToSelf:
self.sendDataThreadQueue.put((0, 'shutdown','no data'))
with shared.printLock:
print 'Closing connection to myself: ', self.peer
return
# The other peer's protocol version is of interest to the sendDataThread but we learn of it
# in this version message. Let us inform the sendDataThread.
self.sendDataThreadQueue.put((0, 'setRemoteProtocolVersion', self.remoteProtocolVersion))
shared.knownNodesLock.acquire()
shared.knownNodes[self.streamNumber][shared.Peer(self.peer.host, self.remoteNodeIncomingPort)] = int(time.time())
shared.needToWriteKnownNodesToDisk = True
shared.knownNodesLock.release()
self.sendverack()
if self.initiatedConnection == False:
self.sendversion()
# Sends a version message
def sendversion(self):
with shared.printLock:
print 'Sending version message'
self.sendDataThreadQueue.put((0, 'sendRawData', shared.assembleVersionMessage(
self.peer.host, self.peer.port, self.streamNumber)))
# Sends a verack message
def sendverack(self):
with shared.printLock:
print 'Sending verack'
self.sendDataThreadQueue.put((0, 'sendRawData', shared.CreatePacket('verack')))
self.verackSent = True
if self.verackReceived:
self.connectionFullyEstablished()