Implement SO_BINDTODEVICE sockopt

PiperOrigin-RevId: 271644926

2 files changed, 189 insertions, 38 deletions

diff --git a/pkg/tcpip/ports/ports.go b/pkg/tcpip/ports/ports.go
index 315780c0c..40e202717 100644
--- a/pkg/tcpip/ports/ports.go
+++ b/pkg/tcpip/ports/ports.go
@@ -47,43 +47,76 @@ type portNode struct {
 	refs  int
 }
 
-// bindAddresses is a set of IP addresses.
-type bindAddresses map[tcpip.Address]portNode
-
-// isAvailable checks whether an IP address is available to bind to.
-func (b bindAddresses) isAvailable(addr tcpip.Address, reuse bool) bool {
-	if addr == anyIPAddress {
-		if len(b) == 0 {
-			return true
-		}
+// deviceNode is never empty. When it has no elements, it is removed from the
+// map that references it.
+type deviceNode map[tcpip.NICID]portNode
+
+// isAvailable checks whether binding is possible by device. If not binding to a
+// device, check against all portNodes. If binding to a specific device, check
+// against the unspecified device and the provided device.
+func (d deviceNode) isAvailable(reuse bool, bindToDevice tcpip.NICID) bool {
+	if bindToDevice == 0 {
+		// Trying to binding all devices.
 		if !reuse {
+			// Can't bind because the (addr,port) is already bound.
 			return false
 		}
-		for _, n := range b {
-			if !n.reuse {
+		for _, p := range d {
+			if !p.reuse {
+				// Can't bind because the (addr,port) was previously bound without reuse.
 				return false
 			}
 		}
 		return true
 	}
 
-	// If all addresses for this portDescriptor are already bound, no
-	// address is available.
-	if n, ok := b[anyIPAddress]; ok {
-		if !reuse {
+	if p, ok := d[0]; ok {
+		if !reuse || !p.reuse {
 			return false
 		}
-		if !n.reuse {
+	}
+
+	if p, ok := d[bindToDevice]; ok {
+		if !reuse || !p.reuse {
 			return false
 		}
 	}
 
-	if n, ok := b[addr]; ok {
-		if !reuse {
+	return true
+}
+
+// bindAddresses is a set of IP addresses.
+type bindAddresses map[tcpip.Address]deviceNode
+
+// isAvailable checks whether an IP address is available to bind to. If the
+// address is the "any" address, check all other addresses. Otherwise, just
+// check against the "any" address and the provided address.
+func (b bindAddresses) isAvailable(addr tcpip.Address, reuse bool, bindToDevice tcpip.NICID) bool {
+	if addr == anyIPAddress {
+		// If binding to the "any" address then check that there are no conflicts
+		// with all addresses.
+		for _, d := range b {
+			if !d.isAvailable(reuse, bindToDevice) {
+				return false
+			}
+		}
+		return true
+	}
+
+	// Check that there is no conflict with the "any" address.
+	if d, ok := b[anyIPAddress]; ok {
+		if !d.isAvailable(reuse, bindToDevice) {
+			return false
+		}
+	}
+
+	// Check that this is no conflict with the provided address.
+	if d, ok := b[addr]; ok {
+		if !d.isAvailable(reuse, bindToDevice) {
 			return false
 		}
-		return n.reuse
 	}
+
 	return true
 }
 
@@ -116,17 +149,17 @@ func (s *PortManager) PickEphemeralPort(testPort func(p uint16) (bool, *tcpip.Er
 }
 
 // IsPortAvailable tests if the given port is available on all given protocols.
-func (s *PortManager) IsPortAvailable(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool) bool {
+func (s *PortManager) IsPortAvailable(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) bool {
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	return s.isPortAvailableLocked(networks, transport, addr, port, reuse)
+	return s.isPortAvailableLocked(networks, transport, addr, port, reuse, bindToDevice)
 }
 
-func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool) bool {
+func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) bool {
 	for _, network := range networks {
 		desc := portDescriptor{network, transport, port}
 		if addrs, ok := s.allocatedPorts[desc]; ok {
-			if !addrs.isAvailable(addr, reuse) {
+			if !addrs.isAvailable(addr, reuse, bindToDevice) {
 				return false
 			}
 		}
@@ -138,14 +171,14 @@ func (s *PortManager) isPortAvailableLocked(networks []tcpip.NetworkProtocolNumb
 // reserved by another endpoint. If port is zero, ReservePort will search for
 // an unreserved ephemeral port and reserve it, returning its value in the
 // "port" return value.
-func (s *PortManager) ReservePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool) (reservedPort uint16, err *tcpip.Error) {
+func (s *PortManager) ReservePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) (reservedPort uint16, err *tcpip.Error) {
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
 	// If a port is specified, just try to reserve it for all network
 	// protocols.
 	if port != 0 {
-		if !s.reserveSpecificPort(networks, transport, addr, port, reuse) {
+		if !s.reserveSpecificPort(networks, transport, addr, port, reuse, bindToDevice) {
 			return 0, tcpip.ErrPortInUse
 		}
 		return port, nil
@@ -153,13 +186,13 @@ func (s *PortManager) ReservePort(networks []tcpip.NetworkProtocolNumber, transp
 
 	// A port wasn't specified, so try to find one.
 	return s.PickEphemeralPort(func(p uint16) (bool, *tcpip.Error) {
-		return s.reserveSpecificPort(networks, transport, addr, p, reuse), nil
+		return s.reserveSpecificPort(networks, transport, addr, p, reuse, bindToDevice), nil
 	})
 }
 
 // reserveSpecificPort tries to reserve the given port on all given protocols.
-func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool) bool {
-	if !s.isPortAvailableLocked(networks, transport, addr, port, reuse) {
+func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, reuse bool, bindToDevice tcpip.NICID) bool {
+	if !s.isPortAvailableLocked(networks, transport, addr, port, reuse, bindToDevice) {
 		return false
 	}
 
@@ -171,11 +204,16 @@ func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber
 			m = make(bindAddresses)
 			s.allocatedPorts[desc] = m
 		}
-		if n, ok := m[addr]; ok {
+		d, ok := m[addr]
+		if !ok {
+			d = make(deviceNode)
+			m[addr] = d
+		}
+		if n, ok := d[bindToDevice]; ok {
 			n.refs++
-			m[addr] = n
+			d[bindToDevice] = n
 		} else {
-			m[addr] = portNode{reuse: reuse, refs: 1}
+			d[bindToDevice] = portNode{reuse: reuse, refs: 1}
 		}
 	}
 
@@ -184,22 +222,28 @@ func (s *PortManager) reserveSpecificPort(networks []tcpip.NetworkProtocolNumber
 
 // ReleasePort releases the reservation on a port/IP combination so that it can
 // be reserved by other endpoints.
-func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16) {
+func (s *PortManager) ReleasePort(networks []tcpip.NetworkProtocolNumber, transport tcpip.TransportProtocolNumber, addr tcpip.Address, port uint16, bindToDevice tcpip.NICID) {
 	s.mu.Lock()
 	defer s.mu.Unlock()
 
 	for _, network := range networks {
 		desc := portDescriptor{network, transport, port}
 		if m, ok := s.allocatedPorts[desc]; ok {
-			n, ok := m[addr]
+			d, ok := m[addr]
+			if !ok {
+				continue
+			}
+			n, ok := d[bindToDevice]
 			if !ok {
 				continue
 			}
 			n.refs--
+			d[bindToDevice] = n
 			if n.refs == 0 {
+				delete(d, bindToDevice)
+			}
+			if len(d) == 0 {
 				delete(m, addr)
-			} else {
-				m[addr] = n
 			}
 			if len(m) == 0 {
 				delete(s.allocatedPorts, desc)
diff --git a/pkg/tcpip/ports/ports_test.go b/pkg/tcpip/ports/ports_test.go
index 689401661..a67e283f1 100644
--- a/pkg/tcpip/ports/ports_test.go
+++ b/pkg/tcpip/ports/ports_test.go
@@ -34,6 +34,7 @@ type portReserveTestAction struct {
 	want    *tcpip.Error
 	reuse   bool
 	release bool
+	device  tcpip.NICID
 }
 
 func TestPortReservation(t *testing.T) {
@@ -100,6 +101,112 @@ func TestPortReservation(t *testing.T) {
 				{port: 24, ip: anyIPAddress, release: true},
 				{port: 24, ip: anyIPAddress, reuse: false, want: nil},
 			},
+		}, {
+			tname: "bind twice with device fails",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 3, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 3, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "bind to device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 1, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 2, want: nil},
+			},
+		}, {
+			tname: "bind to device and then without device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "bind without device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 123, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, reuse: true, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "bind with device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 123, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 0, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 456, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 789, want: nil},
+				{port: 24, ip: fakeIPAddress, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, reuse: true, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "bind with reuse",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 123, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: nil},
+			},
+		}, {
+			tname: "binding with reuse and device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 123, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 456, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 789, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 999, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "mixing reuse and not reuse by binding to device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 456, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 789, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 999, want: nil},
+			},
+		}, {
+			tname: "can't bind to 0 after mixing reuse and not reuse",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 456, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "bind and release",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 345, reuse: false, want: tcpip.ErrPortInUse},
+				{port: 24, ip: fakeIPAddress, device: 789, reuse: true, want: nil},
+
+				// Release the bind to device 0 and try again.
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: nil, release: true},
+				{port: 24, ip: fakeIPAddress, device: 345, reuse: false, want: nil},
+			},
+		}, {
+			tname: "bind twice with reuse once",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: false, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 0, reuse: true, want: tcpip.ErrPortInUse},
+			},
+		}, {
+			tname: "release an unreserved device",
+			actions: []portReserveTestAction{
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: false, want: nil},
+				{port: 24, ip: fakeIPAddress, device: 456, reuse: false, want: nil},
+				// The below don't exist.
+				{port: 24, ip: fakeIPAddress, device: 345, reuse: false, want: nil, release: true},
+				{port: 9999, ip: fakeIPAddress, device: 123, reuse: false, want: nil, release: true},
+				// Release all.
+				{port: 24, ip: fakeIPAddress, device: 123, reuse: false, want: nil, release: true},
+				{port: 24, ip: fakeIPAddress, device: 456, reuse: false, want: nil, release: true},
+			},
 		},
 	} {
 		t.Run(test.tname, func(t *testing.T) {
@@ -108,12 +215,12 @@ func TestPortReservation(t *testing.T) {
 
 			for _, test := range test.actions {
 				if test.release {
-					pm.ReleasePort(net, fakeTransNumber, test.ip, test.port)
+					pm.ReleasePort(net, fakeTransNumber, test.ip, test.port, test.device)
 					continue
 				}
-				gotPort, err := pm.ReservePort(net, fakeTransNumber, test.ip, test.port, test.reuse)
+				gotPort, err := pm.ReservePort(net, fakeTransNumber, test.ip, test.port, test.reuse, test.device)
 				if err != test.want {
-					t.Fatalf("ReservePort(.., .., %s, %d, %t) = %v, want %v", test.ip, test.port, test.release, err, test.want)
+					t.Fatalf("ReservePort(.., .., %s, %d, %t, %d) = %v, want %v", test.ip, test.port, test.reuse, test.device, err, test.want)
 				}
 				if test.port == 0 && (gotPort == 0 || gotPort < FirstEphemeral) {
 					t.Fatalf("ReservePort(.., .., .., 0) = %d, want port number >= %d to be picked", gotPort, FirstEphemeral)