Add multi-core concurrent packet processing

osdep/BSDEthernetTap.cpp

@@ -39,7 +39,9 @@
 #include <net/if_dl.h>
 #include <net/if_media.h>
 #include <net/route.h>
+#include <pthread_np.h>
 
+#include <sched.h>
 #include <string>
 #include <map>
 #include <set>
@@ -53,6 +55,7 @@
 #include "BSDEthernetTap.hpp"
 
 #define ZT_BASE32_CHARS "0123456789abcdefghijklmnopqrstuv"
+#define ZT_TAP_BUF_SIZE (1024 * 16)
 
 // ff:ff:ff:ff:ff:ff with no ADI
 static const ZeroTier::MulticastGroup _blindWildcardMulticastGroup(ZeroTier::MAC(0xff),0);
@@ -61,6 +64,7 @@ namespace ZeroTier {
 
 BSDEthernetTap::BSDEthernetTap(
 	const char *homePath,
+	unsigned int concurrency,
 	const MAC &mac,
 	unsigned int mtu,
 	unsigned int metric,
@@ -69,6 +73,7 @@ BSDEthernetTap::BSDEthernetTap(
 	void (*handler)(void *,void *,uint64_t,const MAC &,const MAC &,unsigned int,unsigned int,const void *,unsigned int),
 	void *arg) :
 	_handler(handler),
+	_concurrency(concurrency),
 	_arg(arg),
 	_nwid(nwid),
 	_mtu(mtu),
@@ -195,11 +200,9 @@ BSDEthernetTap::BSDEthernetTap(
 BSDEthernetTap::~BSDEthernetTap()
 {
 	::write(_shutdownSignalPipe[1],"\0",1); // causes thread to exit
-	Thread::join(_thread);
 	::close(_fd);
 	::close(_shutdownSignalPipe[0]);
 	::close(_shutdownSignalPipe[1]);
-
 	long cpid = (long)vfork();
 	if (cpid == 0) {
 #ifdef ZT_TRACE
@@ -211,6 +214,10 @@ BSDEthernetTap::~BSDEthernetTap()
 		int exitcode = -1;
 		::waitpid(cpid,&exitcode,0);
 	}
+	Thread::join(_thread);
+	for (std::thread &t : _rxThreads) {
+		t.join();
+	}
 }
 
 void BSDEthernetTap::setEnabled(bool en)
@@ -418,20 +425,49 @@ void BSDEthernetTap::setMtu(unsigned int mtu)
 void BSDEthernetTap::threadMain()
 	throw()
 {
-	fd_set readfds,nullfds;
+	bool _enablePinning = false;
-	MAC to,from;
+	char* envvar = std::getenv("ZT_CPU_PINNING");
-	int n,nfds,r;
+	if (envvar) {
-	char getBuf[ZT_MAX_MTU + 64];
+		int tmp = atoi(envvar);
+		if (tmp > 0) {
+			_enablePinning = true;
+		}
+	}
 
 	// Wait for a moment after startup -- wait for Network to finish
 	// constructing itself.
 	Thread::sleep(500);
 
+	for (unsigned int i = 0; i < _concurrency; ++i) {
+		_rxThreads.push_back(std::thread([this, i, _enablePinning] {
+
+			if (_enablePinning) {
+				int pinCore = i % _concurrency;
+				fprintf(stderr, "pinning thread %d to core %d\n", i, pinCore);
+				pthread_t self = pthread_self();
+				cpu_set_t cpuset;
+				CPU_ZERO(&cpuset);
+				CPU_SET(pinCore, &cpuset);
+				//int rc = sched_setaffinity(self, sizeof(cpu_set_t), &cpuset);
+				int rc = pthread_setaffinity_np(self, sizeof(cpu_set_t), &cpuset);
+				if (rc != 0)
+				{
+					fprintf(stderr, "failed to pin thread %d to core %d: %s\n", i, pinCore, strerror(errno));
+					exit(1);
+				}
+			}
+
+			uint8_t b[ZT_TAP_BUF_SIZE];
+			MAC to, from;
+			fd_set readfds, nullfds;
+			int n, nfds, r;
+
 			FD_ZERO(&readfds);
 			FD_ZERO(&nullfds);
 			nfds = (int)std::max(_shutdownSignalPipe[0],_fd) + 1;
 
 			r = 0;
+
 			for(;;) {
 				FD_SET(_shutdownSignalPipe[0],&readfds);
 				FD_SET(_fd,&readfds);
@@ -441,7 +477,7 @@ void BSDEthernetTap::threadMain()
 					break;
 
 				if (FD_ISSET(_fd,&readfds)) {
-			n = (int)::read(_fd,getBuf + r,sizeof(getBuf) - r);
+					n = (int)::read(_fd,b + r,sizeof(b) - r);
 					if (n < 0) {
 						if ((errno != EINTR)&&(errno != ETIMEDOUT))
 							break;
@@ -455,10 +491,10 @@ void BSDEthernetTap::threadMain()
 								r = _mtu + 14;
 
 							if (_enabled) {
-						to.setTo(getBuf,6);
+								to.setTo(b,6);
-						from.setTo(getBuf + 6,6);
+								from.setTo(b + 6,6);
-						unsigned int etherType = ntohs(((const uint16_t *)getBuf)[6]);
+								unsigned int etherType = ntohs(((const uint16_t *)b)[6]);
-						_handler(_arg,(void *)0,_nwid,from,to,etherType,0,(const void *)(getBuf + 14),r - 14);
+								_handler(_arg,(void *)0,_nwid,from,to,etherType,0,(const void *)(b + 14),r - 14);
 							}
 
 							r = 0;
@@ -466,6 +502,8 @@ void BSDEthernetTap::threadMain()
 					}
 				}
 			}
+		}));
+	}
 }
 
 } // namespace ZeroTier

osdep/BSDEthernetTap.hpp

@@ -20,6 +20,7 @@
 #include <string>
 #include <vector>
 #include <stdexcept>
+#include <thread>
 
 #include "../node/Constants.hpp"
 #include "../node/MulticastGroup.hpp"
@@ -34,6 +35,7 @@ class BSDEthernetTap : public EthernetTap
 public:
 	BSDEthernetTap(
 		const char *homePath,
+		unsigned int concurrency,
 		const MAC &mac,
 		unsigned int mtu,
 		unsigned int metric,
@@ -62,6 +64,7 @@ public:
 private:
 	void (*_handler)(void *,void *,uint64_t,const MAC &,const MAC &,unsigned int,unsigned int,const void *,unsigned int);
 	void *_arg;
+	unsigned int _concurrency;
 	uint64_t _nwid;
 	Thread _thread;
 	std::string _dev;
@@ -73,6 +76,7 @@ private:
 	volatile bool _enabled;
 	mutable std::vector<InetAddress> _ifaddrs;
 	mutable uint64_t _lastIfAddrsUpdate;
+	std::vector<std::thread> _rxThreads;
 };
 
 } // namespace ZeroTier

osdep/EthernetTap.cpp

@@ -58,6 +58,7 @@ namespace ZeroTier {
 std::shared_ptr<EthernetTap> EthernetTap::newInstance(
 	const char *tapDeviceType, // OS-specific, NULL for default
 	const char *homePath,
+	unsigned int concurrency,
 	const MAC &mac,
 	unsigned int mtu,
 	unsigned int metric,
@@ -83,16 +84,16 @@ std::shared_ptr<EthernetTap> EthernetTap::newInstance(
 			// The "feth" virtual Ethernet device type appeared in Darwin 17.x.x. Older versions
 			// (Sierra and earlier) must use the a kernel extension.
 			if (strtol(osrelease,(char **)0,10) < 17) {
-				return std::shared_ptr<EthernetTap>(new MacKextEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+				return std::shared_ptr<EthernetTap>(new MacKextEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 			} else {
-				return std::shared_ptr<EthernetTap>(new MacEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+				return std::shared_ptr<EthernetTap>(new MacEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 			}
 		}
 	}
 #endif // __APPLE__
 
 #ifdef __LINUX__
-	return std::shared_ptr<EthernetTap>(new LinuxEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+	return std::shared_ptr<EthernetTap>(new LinuxEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 #endif // __LINUX__
 
 #ifdef __WINDOWS__
@@ -126,19 +127,19 @@ std::shared_ptr<EthernetTap> EthernetTap::newInstance(
 			_comInit = true;
 		}
 	}
-	return std::shared_ptr<EthernetTap>(new WindowsEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+	return std::shared_ptr<EthernetTap>(new WindowsEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 #endif // __WINDOWS__
 
 #ifdef __FreeBSD__
-	return std::shared_ptr<EthernetTap>(new BSDEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+	return std::shared_ptr<EthernetTap>(new BSDEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 #endif // __FreeBSD__
 
 #ifdef __NetBSD__
-	return std::shared_ptr<EthernetTap>(new NetBSDEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+	return std::shared_ptr<EthernetTap>(new NetBSDEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 #endif // __NetBSD__
 
 #ifdef __OpenBSD__
-	return std::shared_ptr<EthernetTap>(new BSDEthernetTap(homePath,mac,mtu,metric,nwid,friendlyName,handler,arg));
+	return std::shared_ptr<EthernetTap>(new BSDEthernetTap(homePath,concurrency,mac,mtu,metric,nwid,friendlyName,handler,arg));
 #endif // __OpenBSD__
 
 #endif // ZT_SDK?

osdep/EthernetTap.hpp

@@ -33,6 +33,7 @@ public:
 	static std::shared_ptr<EthernetTap> newInstance(
 		const char *tapDeviceType, // OS-specific, NULL for default
 		const char *homePath,
+		unsigned int concurrency,
 		const MAC &mac,
 		unsigned int mtu,
 		unsigned int metric,

osdep/LinuxEthernetTap.cpp

@@ -60,7 +60,7 @@
 #define IFNAMSIZ 16
 #endif
 
-#define ZT_TAP_BUF_SIZE 16384
+#define ZT_TAP_BUF_SIZE (1024 * 16)
 
 // ff:ff:ff:ff:ff:ff with no ADI
 static const ZeroTier::MulticastGroup _blindWildcardMulticastGroup(ZeroTier::MAC(0xff),0);
@@ -111,6 +111,7 @@ static void _base32_5_to_8(const uint8_t *in,char *out)
 
 LinuxEthernetTap::LinuxEthernetTap(
 	const char *homePath,
+	unsigned int concurrency,
 	const MAC &mac,
 	unsigned int mtu,
 	unsigned int metric,
@@ -127,6 +128,7 @@ LinuxEthernetTap::LinuxEthernetTap(
 	_fd(0),
 	_enabled(true),
 	_run(true),
+	_concurrency(concurrency),
 	_lastIfAddrsUpdate(0)
 {
 	static std::mutex s_tapCreateLock;
@@ -220,29 +222,54 @@ LinuxEthernetTap::LinuxEthernetTap(
 
 	(void)::pipe(_shutdownSignalPipe);
 
-	_tapReaderThread = std::thread([this]{
+	bool _enablePinning = false;
+	char* envvar = std::getenv("ZT_CPU_PINNING");
+	if (envvar) {
+		int tmp = atoi(envvar);
+		if (tmp > 0) {
+			_enablePinning = true;
+		}
+	}
+
+	for (unsigned int i = 0; i < _concurrency; ++i) {
+		_rxThreads.push_back(std::thread([this, i, _enablePinning] {
+
+			if (_enablePinning) {
+				int pinCore = i % _concurrency;
+				fprintf(stderr, "pinning thread %d to core %d\n", i, pinCore);
+				pthread_t self = pthread_self();
+				cpu_set_t cpuset;
+				CPU_ZERO(&cpuset);
+				CPU_SET(pinCore, &cpuset);
+				int rc = pthread_setaffinity_np(self, sizeof(cpu_set_t), &cpuset);
+				if (rc != 0)
+				{
+					fprintf(stderr, "failed to pin thread %d to core %d: %s\n", i, pinCore, strerror(errno));
+					exit(1);
+				}
+			}
+
 			uint8_t b[ZT_TAP_BUF_SIZE];
-		fd_set readfds,nullfds;
+			fd_set readfds, nullfds;
-		int n,nfds,r;
+			int n, nfds, r;
-		std::vector<void *> buffers;
+			if (i == 0) {
 				struct ifreq ifr;
+				memset(&ifr, 0, sizeof(ifr));
+				strcpy(ifr.ifr_name, _dev.c_str());
 
-		memset(&ifr,0,sizeof(ifr));
+				const int sock = socket(AF_INET, SOCK_DGRAM, 0);
-		strcpy(ifr.ifr_name,_dev.c_str());
-
-		const int sock = socket(AF_INET,SOCK_DGRAM,0);
 				if (sock <= 0)
 					return;
 
-		if (ioctl(sock,SIOCGIFFLAGS,(void *)&ifr) < 0) {
+				if (ioctl(sock, SIOCGIFFLAGS, (void*)&ifr) < 0) {
 					::close(sock);
 					printf("WARNING: ioctl() failed setting up Linux tap device (bring interface up)\n");
 					return;
 				}
 
 				ifr.ifr_ifru.ifru_hwaddr.sa_family = ARPHRD_ETHER;
-		_mac.copyTo(ifr.ifr_ifru.ifru_hwaddr.sa_data,6);
+				_mac.copyTo(ifr.ifr_ifru.ifru_hwaddr.sa_data, 6);
-		if (ioctl(sock,SIOCSIFHWADDR,(void *)&ifr) < 0) {
+				if (ioctl(sock, SIOCSIFHWADDR, (void*)&ifr) < 0) {
 					::close(sock);
 					printf("WARNING: ioctl() failed setting up Linux tap device (set MAC)\n");
 					return;
@@ -252,7 +279,7 @@ LinuxEthernetTap::LinuxEthernetTap(
 
 				if (isOldLinuxKernel()) {
 					ifr.ifr_ifru.ifru_mtu = (int)_mtu;
-			if (ioctl(sock,SIOCSIFMTU,(void *)&ifr) < 0) {
+					if (ioctl(sock, SIOCSIFMTU, (void*)&ifr) < 0) {
 						::close(sock);
 						printf("WARNING: ioctl() failed setting up Linux tap device (set MTU)\n");
 						return;
@@ -261,10 +288,9 @@ LinuxEthernetTap::LinuxEthernetTap(
 					usleep(100000);
 				}
 
-
 				ifr.ifr_flags |= IFF_MULTICAST;
 				ifr.ifr_flags |= IFF_UP;
-		if (ioctl(sock,SIOCSIFFLAGS,(void *)&ifr) < 0) {
+				if (ioctl(sock, SIOCSIFFLAGS, (void*)&ifr) < 0) {
 					::close(sock);
 					printf("WARNING: ioctl() failed setting up Linux tap device (bring interface up)\n");
 					return;
@@ -272,46 +298,49 @@ LinuxEthernetTap::LinuxEthernetTap(
 
 				usleep(100000);
 
-		if (!isOldLinuxKernel()) {
+				if (! isOldLinuxKernel()) {
 					ifr.ifr_ifru.ifru_hwaddr.sa_family = ARPHRD_ETHER;
-			_mac.copyTo(ifr.ifr_ifru.ifru_hwaddr.sa_data,6);
+					_mac.copyTo(ifr.ifr_ifru.ifru_hwaddr.sa_data, 6);
-			if (ioctl(sock,SIOCSIFHWADDR,(void *)&ifr) < 0) {
+					if (ioctl(sock, SIOCSIFHWADDR, (void*)&ifr) < 0) {
 						::close(sock);
 						printf("WARNING: ioctl() failed setting up Linux tap device (set MAC)\n");
 						return;
 					}
 
 					ifr.ifr_ifru.ifru_mtu = (int)_mtu;
-			if (ioctl(sock,SIOCSIFMTU,(void *)&ifr) < 0) {
+					if (ioctl(sock, SIOCSIFMTU, (void*)&ifr) < 0) {
 						::close(sock);
 						printf("WARNING: ioctl() failed setting up Linux tap device (set MTU)\n");
 						return;
 					}
 				}
 
-		fcntl(_fd,F_SETFL,O_NONBLOCK);
+				fcntl(_fd, F_SETFL, O_NONBLOCK);
 
 				::close(sock);
+			}
 
-		if (!_run)
+			if (! _run) {
 				return;
+			}
 
 			FD_ZERO(&readfds);
 			FD_ZERO(&nullfds);
-		nfds = (int)std::max(_shutdownSignalPipe[0],_fd) + 1;
+			nfds = (int)std::max(_shutdownSignalPipe[0], _fd) + 1;
 
 			r = 0;
-		for(;;) {
+			for (;;) {
-			FD_SET(_shutdownSignalPipe[0],&readfds);
+				FD_SET(_shutdownSignalPipe[0], &readfds);
-			FD_SET(_fd,&readfds);
+				FD_SET(_fd, &readfds);
-			select(nfds,&readfds,&nullfds,&nullfds,(struct timeval *)0);
+				select(nfds, &readfds, &nullfds, &nullfds, (struct timeval*)0);
 
-			if (FD_ISSET(_shutdownSignalPipe[0],&readfds))
+				if (FD_ISSET(_shutdownSignalPipe[0], &readfds)) {
 					break;
-
+				}
-			if (FD_ISSET(_fd,&readfds)) {
+				if (FD_ISSET(_fd, &readfds)) {
-				for(;;) { // read until there are no more packets, then return to outer select() loop
+					for (;;) {
-					n = (int)::read(_fd,b + r,ZT_TAP_BUF_SIZE - r);
+						// read until there are no more packets, then return to outer select() loop
+						n = (int)::read(_fd, b + r, ZT_TAP_BUF_SIZE - r);
 						if (n > 0) {
 							// Some tap drivers like to send the ethernet frame and the
 							// payload in two chunks, so handle that by accumulating
@@ -322,33 +351,35 @@ LinuxEthernetTap::LinuxEthernetTap(
 									r = _mtu + 14;
 
 								if (_enabled) {
-								//_tapq.post(std::pair<void *,int>(buf,r));
+									MAC to(b, 6), from(b + 6, 6);
-								//buf = nullptr;
+									unsigned int etherType = Utils::ntoh(((const uint16_t*)b)[6]);
-								MAC to(b, 6),from(b + 6, 6);
+									_handler(_arg, nullptr, _nwid, from, to, etherType, 0, (const void*)(b + 14), (unsigned int)(r - 14));
-								unsigned int etherType = Utils::ntoh(((const uint16_t *)b)[6]);
-								_handler(_arg, nullptr, _nwid, from, to, etherType, 0, (const void *)(b + 14),(unsigned int)(r - 14));
 								}
 
 								r = 0;
 							}
-					} else {
+						}
+						else {
 							r = 0;
 							break;
 						}
 					}
 				}
 			}
-	});
+		}));
+	}
 }
 
 LinuxEthernetTap::~LinuxEthernetTap()
 {
 	_run = false;
 	(void)::write(_shutdownSignalPipe[1],"\0",1);
-	_tapReaderThread.join();
 	::close(_fd);
 	::close(_shutdownSignalPipe[0]);
 	::close(_shutdownSignalPipe[1]);
+	for (std::thread &t : _rxThreads) {
+		t.join();
+	}
 }
 
 void LinuxEthernetTap::setEnabled(bool en)

osdep/LinuxEthernetTap.hpp

@@ -26,6 +26,7 @@
 #include <mutex>
 #include "../node/MulticastGroup.hpp"
 #include "EthernetTap.hpp"
+#include "BlockingQueue.hpp"
 
 namespace ZeroTier {
 
@@ -34,6 +35,7 @@ class LinuxEthernetTap : public EthernetTap
 public:
 	LinuxEthernetTap(
 		const char *homePath,
+		unsigned int _concurrency,
 		const MAC &mac,
 		unsigned int mtu,
 		unsigned int metric,
@@ -57,9 +59,6 @@ public:
 	virtual void setMtu(unsigned int mtu);
 	virtual void setDns(const char *domain, const std::vector<InetAddress> &servers) {}
 
-
-
-
 private:
 	void (*_handler)(void *,void *,uint64_t,const MAC &,const MAC &,unsigned int,unsigned int,const void *,unsigned int);
 	void *_arg;
@@ -69,13 +68,14 @@ private:
 	std::string _dev;
 	std::vector<MulticastGroup> _multicastGroups;
 	unsigned int _mtu;
+	unsigned int _concurrency;
 	int _fd;
 	int _shutdownSignalPipe[2];
 	std::atomic_bool _enabled;
 	std::atomic_bool _run;
-	std::thread _tapReaderThread;
 	mutable std::vector<InetAddress> _ifaddrs;
 	mutable uint64_t _lastIfAddrsUpdate;
+	std::vector<std::thread> _rxThreads;
 };
 
 } // namespace ZeroTier

osdep/MacEthernetTap.cpp

@@ -69,6 +69,7 @@ static bool fethMaxMtuAdjusted = false;
 
 MacEthernetTap::MacEthernetTap(
 	const char *homePath,
+	unsigned int concurrency,
 	const MAC &mac,
 	unsigned int mtu,
 	unsigned int metric,
@@ -77,6 +78,7 @@ MacEthernetTap::MacEthernetTap(
 	void (*handler)(void *,void *,uint64_t,const MAC &,const MAC &,unsigned int,unsigned int,const void *data,unsigned int len),
 	void *arg) :
 	_handler(handler),
+	_concurrency(concurrency),
 	_arg(arg),
 	_nwid(nwid),
 	_homePath(homePath),
@@ -286,6 +288,9 @@ MacEthernetTap::~MacEthernetTap()
 	}
 
 	Thread::join(_thread);
+	for (std::thread &t : _rxThreads) {
+		t.join();
+	}
 }
 
 void MacEthernetTap::setEnabled(bool en) { _enabled = en; }
@@ -474,17 +479,25 @@ void MacEthernetTap::setMtu(unsigned int mtu)
 void MacEthernetTap::threadMain()
 	throw()
 {
+	Thread::sleep(250);
+
+	for (unsigned int i = 0; i < _concurrency; ++i) {
+		_rxThreads.push_back(std::thread([this, i] {
+
+			fprintf(stderr, "starting thread %d\n", i);
+
 	char agentReadBuf[ZT_MACETHERNETTAP_AGENT_READ_BUF_SIZE];
 	char agentStderrBuf[256];
 	fd_set readfds,nullfds;
 	MAC to,from;
 
-	Thread::sleep(250);
-
 	const int nfds = std::max(std::max(_shutdownSignalPipe[0],_agentStdout),_agentStderr) + 1;
 	long agentReadPtr = 0;
+
+	if (i == 0) {
 		fcntl(_agentStdout,F_SETFL,fcntl(_agentStdout,F_GETFL)|O_NONBLOCK);
 		fcntl(_agentStderr,F_SETFL,fcntl(_agentStderr,F_GETFL)|O_NONBLOCK);
+	}
 
 	FD_ZERO(&readfds);
 	FD_ZERO(&nullfds);
@@ -533,6 +546,7 @@ void MacEthernetTap::threadMain()
 			*/
 		}
 	}
+		}));}
 
 	::close(_agentStdin);
 	::close(_agentStdout);

osdep/MacEthernetTap.hpp

@@ -28,6 +28,7 @@
 #include <stdexcept>
 #include <string>
 #include <vector>
+#include <thread>
 
 namespace ZeroTier {
 
@@ -36,6 +37,7 @@ class MacEthernetTap : public EthernetTap
 public:
 	MacEthernetTap(
 		const char *homePath,
+		unsigned int concurrency,
 		const MAC &mac,
 		unsigned int mtu,
 		unsigned int metric,
@@ -67,6 +69,7 @@ private:
 	uint64_t _nwid;
 	Thread _thread;
 	std::string _homePath;
+	unsigned int _concurrency;
 	std::string _dev;
 	std::vector<MulticastGroup> _multicastGroups;
 	Mutex _putLock;
@@ -79,6 +82,7 @@ private:
 	volatile bool _enabled;
 	mutable std::vector<InetAddress> _ifaddrs;
 	mutable uint64_t _lastIfAddrsUpdate;
+	std::vector<std::thread> _rxThreads;
 
 };
 

osdep/MacKextEthernetTap.cpp

@@ -306,6 +306,7 @@ static Mutex globalTapCreateLock;
 
 MacKextEthernetTap::MacKextEthernetTap(
 	const char *homePath,
+	unsigned int concurrency,
 	const MAC &mac,
 	unsigned int mtu,
 	unsigned int metric,
@@ -317,6 +318,7 @@ MacKextEthernetTap::MacKextEthernetTap(
 	_arg(arg),
 	_nwid(nwid),
 	_homePath(homePath),
+	_concurrency(concurrency),
 	_mtu(mtu),
 	_metric(metric),
 	_fd(0),
@@ -447,7 +449,9 @@ MacKextEthernetTap::~MacKextEthernetTap()
 
 	::write(_shutdownSignalPipe[1],"\0",1); // causes thread to exit
 	Thread::join(_thread);
-
+		for (std::thread &t : _rxThreads) {
+		t.join();
+	}
 	::close(_fd);
 	::close(_shutdownSignalPipe[0]);
 	::close(_shutdownSignalPipe[1]);

osdep/MacKextEthernetTap.hpp

@@ -20,6 +20,7 @@
 #include <stdexcept>
 #include <string>
 #include <vector>
+#include <thread>
 
 #include "../node/Constants.hpp"
 #include "../node/MAC.hpp"
@@ -36,6 +37,7 @@ class MacKextEthernetTap : public EthernetTap
 public:
 	MacKextEthernetTap(
 		const char *homePath,
+		unsigned int concurrency,
 		const MAC &mac,
 		unsigned int mtu,
 		unsigned int metric,
@@ -70,11 +72,13 @@ private:
 	std::string _homePath;
 	std::string _dev;
 	std::vector<MulticastGroup> _multicastGroups;
+	unsigned int _concurrency;
 	unsigned int _mtu;
 	unsigned int _metric;
 	int _fd;
 	int _shutdownSignalPipe[2];
 	volatile bool _enabled;
+	std::vector<std::thread> _rxThreads;
 };
 
 } // namespace ZeroTier

service/OneService.cpp

@@ -16,7 +16,6 @@
 #include <stdlib.h>
 #include <string.h>
 #include <stdint.h>
-
 #include <string>
 #include <map>
 #include <vector>
@@ -26,6 +25,11 @@
 #include <mutex>
 #include <condition_variable>
 
+#ifdef __FreeBSD__
+#include <sched.h>
+#include <pthread_np.h>
+#endif
+
 #include "../version.h"
 #include "../include/ZeroTierOne.h"
 
@@ -758,7 +762,7 @@ struct TcpConnection
 	Mutex writeq_m;
 };
 
-struct OneServiceIncomingPacket
+struct PacketRecord
 {
 	uint64_t now;
 	int64_t sock;
@@ -792,6 +796,12 @@ public:
 	bool _serverThreadRunning;
 	bool _serverThreadRunningV6;
 
+	unsigned int _rxThreadCount;
+	BlockingQueue<PacketRecord *> _rxPacketQueue;
+	std::vector<PacketRecord *> _rxPacketVector;
+	std::vector<std::thread> _rxPacketThreads;
+	Mutex _rxPacketVector_m,_rxPacketThreads_m;
+
 	bool _allowTcpFallbackRelay;
 	bool _forceTcpRelay;
 	bool _allowSecondaryPort;
@@ -842,8 +852,6 @@ public:
 	// Deadline for the next background task service function
 	volatile int64_t _nextBackgroundTaskDeadline;
 
-
-
 	std::map<uint64_t,NetworkState> _nets;
 	Mutex _nets_m;
 
@@ -926,6 +934,76 @@ public:
 		_ports[1] = 0;
 		_ports[2] = 0;
 
+		bool _enablePinning = false;
+		char* pinningVar = std::getenv("ZT_CPU_PINNING");
+		if (pinningVar) {
+			int tmp = atoi(pinningVar);
+			if (tmp > 0) {
+				_enablePinning = true;
+			}
+		}
+		char* concurrencyVar = std::getenv("ZT_PACKET_PROCESSING_CONCURRENCY");
+		if (concurrencyVar) {
+			int tmp = atoi(concurrencyVar);
+			if (tmp > 0) {
+				_rxThreadCount = tmp;
+			}
+			else {
+				_rxThreadCount = std::thread::hardware_concurrency();
+			}
+		}
+		else {
+			_rxThreadCount = std::thread::hardware_concurrency();
+		}
+		for (unsigned int i = 0; i < _rxThreadCount; ++i) {
+			_rxPacketThreads.push_back(std::thread([this, i]() {
+
+#if defined(__LINUX__) || defined(__FreeBSD__) /* || defined(__APPLE__) */
+				int pinCore = i % _rxThreadCount;
+				fprintf(stderr, "pinning thread %d to core %d\n", i, pinCore);
+				pthread_t self = pthread_self();
+				cpu_set_t cpuset;
+				CPU_ZERO(&cpuset);
+				CPU_SET(pinCore, &cpuset);
+#endif
+#ifdef __LINUX__
+				int rc = pthread_setaffinity_np(self, sizeof(cpu_set_t), &cpuset);
+#elif __FreeBSD__
+				int rc = pthread_setaffinity_np(self, sizeof(cpu_set_t), &cpuset);
+#endif
+#if defined(__LINUX__) || defined(__FreeBSD__) /* || defined(__APPLE__) */
+				if (rc != 0)
+				{
+					fprintf(stderr, "failed to pin thread %d to core %d: %s\n", i, pinCore, strerror(errno));
+					exit(1);
+				}
+#endif
+				PacketRecord* packet = nullptr;
+				for (;;) {
+					if (! _rxPacketQueue.get(packet)) {
+						break;
+					}
+					if (! packet) {
+						break;
+					}
+					const ZT_ResultCode err = _node->processWirePacket(nullptr, packet->now, packet->sock, &(packet->from), packet->data, packet->size, &_nextBackgroundTaskDeadline);
+					{
+						Mutex::Lock l(_rxPacketVector_m);
+						_rxPacketVector.push_back(packet);
+					}
+					if (ZT_ResultCode_isFatal(err)) {
+						char tmp[256];
+						OSUtils::ztsnprintf(tmp, sizeof(tmp), "error processing packet: %d", (int)err);
+						Mutex::Lock _l(_termReason_m);
+						_termReason = ONE_UNRECOVERABLE_ERROR;
+						_fatalErrorMessage = tmp;
+						this->terminate();
+						break;
+					}
+				}
+			}));
+		}
+
 		prometheus::simpleapi::saver.set_registry(prometheus::simpleapi::registry_ptr);
 		prometheus::simpleapi::saver.set_delay(std::chrono::seconds(5));
 		prometheus::simpleapi::saver.set_out_file(_homePath + ZT_PATH_SEPARATOR + "metrics.prom");
@@ -940,6 +1018,13 @@ public:
 #ifdef __WINDOWS__
 		WinFWHelper::removeICMPRules();
 #endif
+
+		_rxPacketQueue.stop();
+		_rxPacketThreads_m.lock();
+		for(auto t=_rxPacketThreads.begin();t!=_rxPacketThreads.end();++t) {
+			t->join();
+		}
+		_rxPacketThreads_m.unlock();
 		_binder.closeAll(_phy);
 
 #if ZT_VAULT_SUPPORT
@@ -954,6 +1039,13 @@ public:
 		if (_serverThreadRunningV6) {
 			_serverThreadV6.join();
 		}
+		_rxPacketVector_m.lock();
+		while (!_rxPacketVector.empty()) {
+			delete _rxPacketVector.back();
+			_rxPacketVector.pop_back();
+		}
+		_rxPacketVector_m.unlock();
+
 
 #ifdef ZT_USE_MINIUPNPC
 		delete _portMapper;
@@ -1270,6 +1362,9 @@ public:
 				const unsigned long delay = (dl > now) ? (unsigned long)(dl - now) : 500;
 				clockShouldBe = now + (int64_t)delay;
 				_phy.poll(delay);
+
+
+
 			}
 		} catch (std::exception &e) {
 			Mutex::Lock _l(_termReason_m);
@@ -2756,25 +2851,37 @@ public:
 	// Handlers for Node and Phy<> callbacks
 	// =========================================================================
 
-	inline void phyOnDatagram(PhySocket *sock,void **uptr,const struct sockaddr *localAddr,const struct sockaddr *from,void *data,unsigned long len)
+
+
+
+	inline void phyOnDatagram(PhySocket* sock, void** uptr, const struct sockaddr* localAddr, const struct sockaddr* from, void* data, unsigned long len)
 	{
 		if (_forceTcpRelay) {
 			return;
 		}
 		Metrics::udp_recv += len;
 		const uint64_t now = OSUtils::now();
-		if ((len >= 16)&&(reinterpret_cast<const InetAddress *>(from)->ipScope() == InetAddress::IP_SCOPE_GLOBAL)) {
+		if ((len >= 16) && (reinterpret_cast<const InetAddress*>(from)->ipScope() == InetAddress::IP_SCOPE_GLOBAL)) {
 			_lastDirectReceiveFromGlobal = now;
 		}
-		const ZT_ResultCode rc = _node->processWirePacket(nullptr,now,reinterpret_cast<int64_t>(sock),reinterpret_cast<const struct sockaddr_storage *>(from),data,len,&_nextBackgroundTaskDeadline);
+
-		if (ZT_ResultCode_isFatal(rc)) {
+		PacketRecord* packet;
-			char tmp[256];
+		_rxPacketVector_m.lock();
-			OSUtils::ztsnprintf(tmp,sizeof(tmp),"fatal error code from processWirePacket: %d",(int)rc);
+		if (_rxPacketVector.empty()) {
-			Mutex::Lock _l(_termReason_m);
+			packet = new PacketRecord;
-			_termReason = ONE_UNRECOVERABLE_ERROR;
-			_fatalErrorMessage = tmp;
-			this->terminate();
 		}
+		else {
+			packet = _rxPacketVector.back();
+			_rxPacketVector.pop_back();
+		}
+		_rxPacketVector_m.unlock();
+
+		packet->sock = reinterpret_cast<int64_t>(sock);
+		packet->now = now;
+		memcpy(&(packet->from), from, sizeof(struct sockaddr_storage));
+		packet->size = (unsigned int)len;
+		memcpy(packet->data, data, len);
+		_rxPacketQueue.postLimit(packet, 256 * _rxThreadCount);
 	}
 
 	inline void phyOnTcpConnect(PhySocket *sock,void **uptr,bool success)
@@ -2996,6 +3103,7 @@ public:
 						n.setTap(EthernetTap::newInstance(
 							nullptr,
 							_homePath.c_str(),
+							_rxThreadCount,
 							MAC(nwc->mac),
 							nwc->mtu,
 							(unsigned int)ZT_IF_METRIC,
@@ -3509,8 +3617,9 @@ public:
 	inline void nodeVirtualNetworkFrameFunction(uint64_t nwid,void **nuptr,uint64_t sourceMac,uint64_t destMac,unsigned int etherType,unsigned int vlanId,const void *data,unsigned int len)
 	{
 		NetworkState *n = reinterpret_cast<NetworkState *>(*nuptr);
-		if ((!n)||(!n->tap()))
+		if ((!n)||(!n->tap())) {
 			return;
+		}
 		n->tap()->put(MAC(sourceMac),MAC(destMac),etherType,data,len);
 	}