Yet more cleanup to TCP logic, this time adding a master switch and adding UDP preference in send().

This commit is contained in:
Adam Ierymenko 2014-04-09 17:08:35 -07:00
parent 73153b89b4
commit 8fb442d81a
6 changed files with 110 additions and 76 deletions

View File

@ -313,7 +313,7 @@ error_no_byte_order_defined;
#define ZT_PEER_PATH_ACTIVITY_TIMEOUT ((ZT_PEER_DIRECT_PING_DELAY * 2) + ZT_PING_CHECK_DELAY)
/**
* Close TCP tunnels if unused for this long
* Close TCP tunnels if unused for this long (used in SocketManager)
*/
#define ZT_TCP_TUNNEL_ACTIVITY_TIMEOUT ZT_PEER_PATH_ACTIVITY_TIMEOUT

View File

@ -591,13 +591,29 @@ Node::ReasonForTermination Node::run()
LOG("resynchronize forced by user, syncing with network");
}
if (resynchronize)
if (resynchronize) {
_r->tcpTunnelingEnabled = false; // turn off TCP tunneling master switch at first
_r->timeOfLastResynchronize = now;
}
/* Ping supernodes separately, and do so more aggressively if we haven't
* heard anything from anyone since our last resynchronize / startup. */
/* Supernodes are pinged separately and more aggressively. The
* ZT_STARTUP_AGGRO parameter sets a limit on how rapidly they are
* tried, while PingSupernodesThatNeedPing contains the logic for
* determining if they need PING. */
if ((now - lastSupernodePingCheck) >= ZT_STARTUP_AGGRO) {
lastSupernodePingCheck = now;
uint64_t lastReceiveFromAnySupernode = 0; // function object result paramter
_r->topology->eachSupernodePeer(Topology::FindMostRecentDirectReceiveTimestamp(lastReceiveFromAnySupernode));
// Turn on TCP tunneling master switch if we haven't heard anything since before
// the last resynchronize and we've been trying long enough.
uint64_t tlr = _r->timeOfLastResynchronize;
if ((lastReceiveFromAnySupernode < tlr)&&((now - tlr) >= ZT_TCP_TUNNEL_FAILOVER_TIMEOUT)) {
TRACE("network still unreachable after %u ms, TCP TUNNELING ENABLED",(unsigned int)ZT_TCP_TUNNEL_FAILOVER_TIMEOUT);
_r->tcpTunnelingEnabled = true;
}
_r->topology->eachSupernodePeer(Topology::PingSupernodesThatNeedPing(_r,now));
}

View File

@ -120,35 +120,44 @@ void Peer::receive(
bool Peer::send(const RuntimeEnvironment *_r,const void *data,unsigned int len,uint64_t now)
{
Mutex::Lock _l(_lock);
bool useTcpOut = _isTcpFailoverTime(_r,now);
std::vector<Path>::iterator p(_paths.begin());
if (!useTcpOut) {
// If we don't want to initiate TCP, seek past TCP paths if they are at the front
// to find the first UDP path as our default.
while ((p != _paths.end())&&(p->type() == Path::PATH_TYPE_TCP_OUT))
++p;
}
if (p == _paths.end())
return false;
// Treat first path as default and look for a better one based on time of
// last packet received.
std::vector<Path>::iterator bestPath = p;
uint64_t bestPathLastReceived = p->lastReceived();
while (++p != _paths.end()) {
Path *bestNormalPath = (Path *)0;
Path *bestTcpOutPath = (Path *)0;
uint64_t bestNormalPathLastReceived = 0;
uint64_t bestTcpOutPathLastReceived = 0;
for(std::vector<Path>::iterator p(_paths.begin());p!=_paths.end();++p) {
uint64_t lr = p->lastReceived();
if ( (lr > bestPathLastReceived) && ((useTcpOut)||(p->type() != Path::PATH_TYPE_TCP_OUT)) ) {
bestPathLastReceived = lr;
bestPath = p;
if (p->type() == Path::PATH_TYPE_TCP_OUT) { // TCP_OUT paths initiate TCP connections
if (lr >= bestTcpOutPathLastReceived) {
bestTcpOutPathLastReceived = lr;
bestTcpOutPath = &(*p);
}
} else { // paths other than TCP_OUT are considered "normal"
if (lr >= bestNormalPathLastReceived) {
bestNormalPathLastReceived = lr;
bestNormalPath = &(*p);
}
}
}
Path *bestPath = (Path *)0;
if (!_r->tcpTunnelingEnabled) { // TCP tunneling master switch is off, use normal path
bestPath = bestNormalPath;
} else if (bestNormalPath) { // we have a normal path, so use if it looks active
if ((bestNormalPathLastReceived > _r->timeOfLastResynchronize)&&((now - bestNormalPathLastReceived) < ZT_PEER_PATH_ACTIVITY_TIMEOUT))
bestPath = bestNormalPath;
else bestPath = bestTcpOutPath;
} else { // no normal path available
bestPath = bestTcpOutPath;
}
if (!bestPath)
return false;
if (_r->sm->send(bestPath->address(),bestPath->tcp(),bestPath->type() == Path::PATH_TYPE_TCP_OUT,data,len)) {
bestPath->sent(now);
return true;
}
return false;
}
@ -170,7 +179,18 @@ bool Peer::sendPing(const RuntimeEnvironment *_r,uint64_t now)
bool sent = false;
SharedPtr<Peer> self(this);
Mutex::Lock _l(_lock);
bool useTcpOut = _isTcpFailoverTime(_r,now);
uint64_t lastUdpPingSent = 0;
uint64_t lastUdpReceive = 0;
bool haveUdp = false;
for(std::vector<Path>::const_iterator p(_paths.begin());p!=_paths.end();++p) {
if (p->type() == Path::PATH_TYPE_UDP) {
lastUdpPingSent = std::max(lastUdpPingSent,p->lastPing());
lastUdpReceive = std::max(lastUdpReceive,p->lastReceived());
haveUdp = true;
}
}
bool useTcpOut = ( (!haveUdp) || ( (_r->tcpTunnelingEnabled) && (lastUdpPingSent > lastUdpReceive) && ((now - lastUdpReceive) >= ZT_TCP_TUNNEL_FAILOVER_TIMEOUT) ) );
TRACE("PING %s (useTcpOut==%d)",_id.address().toString().c_str(),(int)useTcpOut);
@ -199,46 +219,6 @@ void Peer::clean(uint64_t now)
_paths.resize(o);
}
bool Peer::_isTcpFailoverTime(const RuntimeEnvironment *_r,uint64_t now) const
throw()
{
// assumes _lock is locked
uint64_t lastResync = _r->timeOfLastResynchronize;
if ((now - lastResync) >= ZT_TCP_TUNNEL_FAILOVER_TIMEOUT) {
if ((now - _r->timeOfLastPacketReceived) >= ZT_TCP_TUNNEL_FAILOVER_TIMEOUT)
return true;
uint64_t lastUdpPingSent = 0;
uint64_t lastUdpReceive = 0;
bool haveUdp = false;
for(std::vector<Path>::const_iterator p(_paths.begin());p!=_paths.end();++p) {
if (p->type() == Path::PATH_TYPE_UDP) {
lastUdpPingSent = std::max(lastUdpPingSent,p->lastPing());
lastUdpReceive = std::max(lastUdpReceive,p->lastReceived());
haveUdp = true;
}
}
return ( (!haveUdp) || ( (lastUdpPingSent > lastResync) && ((now - lastUdpReceive) >= ZT_TCP_TUNNEL_FAILOVER_TIMEOUT) ) );
}
return false;
}
bool Peer::pingUnanswered(const RuntimeEnvironment *_r,uint64_t now)
{
uint64_t lp = 0;
uint64_t lr = 0;
{
Mutex::Lock _l(_lock);
for(std::vector<Path>::const_iterator p(_paths.begin());p!=_paths.end();++p) {
lp = std::max(p->lastPing(),lp);
lr = std::max(p->lastReceived(),lr);
}
}
return ( (lp > _r->timeOfLastResynchronize) && ((lr < lp)&&((lp - lr) >= ZT_PING_UNANSWERED_AFTER)) );
}
void Peer::getBestActiveUdpPathAddresses(uint64_t now,InetAddress &v4,InetAddress &v6) const
{
uint64_t bestV4 = 0,bestV6 = 0;

View File

@ -218,11 +218,20 @@ public:
}
/**
* @param _r Runtime environment
* @param now Current time
* @return True if the last ping is unanswered
* Get max timestamp of last ping and max timestamp of last receive in a single pass
*
* @param lp Last ping result parameter (init to 0 before calling)
* @param lr Last receive result parameter (init to 0 before calling)
*/
bool pingUnanswered(const RuntimeEnvironment *_r,uint64_t now);
inline void lastPingAndDirectReceive(uint64_t &lp,uint64_t &lr)
throw()
{
Mutex::Lock _l(_lock);
for(std::vector<Path>::const_iterator p(_paths.begin());p!=_paths.end();++p) {
lp = std::max(lp,p->lastPing());
lr = std::max(lr,p->lastReceived());
}
}
/**
* @return Time of most recent unicast frame received
@ -449,9 +458,6 @@ public:
}
private:
bool _isTcpFailoverTime(const RuntimeEnvironment *_r,uint64_t now) const
throw();
unsigned char _key[ZT_PEER_SECRET_KEY_LENGTH];
Identity _id;

View File

@ -64,6 +64,7 @@ class RuntimeEnvironment
public:
RuntimeEnvironment() :
shutdownInProgress(false),
tcpTunnelingEnabled(false),
timeOfLastResynchronize(0),
timeOfLastPacketReceived(0),
log((Logger *)0),
@ -90,6 +91,9 @@ public:
// Indicates that we are shutting down -- this is hacky, want to factor out
volatile bool shutdownInProgress;
// Are we in outgoing TCP failover mode?
volatile bool tcpTunnelingEnabled;
// Time network environment (e.g. fingerprint) last changed -- used to determine online-ness
volatile uint64_t timeOfLastResynchronize;

View File

@ -216,7 +216,13 @@ public:
};
/**
* Pings all peers that need a ping sent, excluding supernodes (which are pinged separately)
* Pings all peers that need a ping sent, excluding supernodes
*
* Ordinary peers are pinged if we haven't heard from them recently. Receive
* time rather than send time as OK is returned on success and we want to
* keep trying if a packet is lost. Ordinary peers are subject to a frame
* inactivity timeout. We give up if we haven't actually transferred any
* data to them recently, and eventually Topology purges them from memory.
*/
class PingPeersThatNeedPing
{
@ -235,7 +241,7 @@ public:
* than time of last send in order to only count full round trips. */
if ( (!_supernodeAddresses.count(p->address())) &&
((_now - p->lastFrame()) < ZT_PEER_PATH_ACTIVITY_TIMEOUT) &&
((_now - p->lastDirectReceive()) > ZT_PEER_DIRECT_PING_DELAY) ) {
((_now - p->lastDirectReceive()) >= ZT_PEER_DIRECT_PING_DELAY) ) {
p->sendPing(_r,_now);
}
}
@ -247,7 +253,13 @@ public:
};
/**
* Ping peers that need ping according to supernode rules (slightly more aggressive)
* Ping peers that need ping according to supernode rules
*
* Supernodes ping aggressively if a ping is unanswered and they are not
* subject to the activity timeout. In other words: we assume they are
* always there and always try to reach them.
*
* The ultimate rate limit for this is controlled up in the Node main loop.
*/
class PingSupernodesThatNeedPing
{
@ -261,7 +273,11 @@ public:
/* For supernodes we always ping even if no frames have been seen, and
* we ping aggressively if pings are unanswered. The limit to this
* frequency is set in the main loop to no more than ZT_STARTUP_AGGRO. */
if ( (p->pingUnanswered(_r,_now)) || ((_now - p->lastDirectReceive()) > ZT_PEER_DIRECT_PING_DELAY) || (p->lastDirectReceive() < _r->timeOfLastResynchronize) )
uint64_t lp = 0;
uint64_t lr = 0;
p->lastPingAndDirectReceive(lp,lr);
if ( (lr < _r->timeOfLastResynchronize) || ((lr < lp)&&((lp - lr) >= ZT_PING_UNANSWERED_AFTER)) || ((_now - lr) >= ZT_PEER_DIRECT_PING_DELAY) )
p->sendPing(_r,_now);
}
@ -270,6 +286,18 @@ public:
const RuntimeEnvironment *_r;
};
/**
* Computes most recent timestamp of direct packet receive over a list of peers
*/
class FindMostRecentDirectReceiveTimestamp
{
public:
FindMostRecentDirectReceiveTimestamp(uint64_t &ts) throw() : _ts(ts) {}
inline void operator()(Topology &t,const SharedPtr<Peer> &p) throw() { _ts = std::max(p->lastDirectReceive(),_ts); }
private:
uint64_t &_ts;
};
/**
* Function object to forget direct links to active peers and then ping them indirectly
*/