Allow dual stack sockets to operate on AF_INET

Fixes #1490
Fixes #1495

PiperOrigin-RevId: 289523250

7 files changed, 120 insertions, 81 deletions

diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go
index 099319327..c020c11cb 100644
--- a/pkg/sentry/socket/netstack/netstack.go
+++ b/pkg/sentry/socket/netstack/netstack.go
@@ -324,22 +324,15 @@ func bytesToIPAddress(addr []byte) tcpip.Address {
 // converts it to the FullAddress format. It supports AF_UNIX, AF_INET,
 // AF_INET6, and AF_PACKET addresses.
 //
-// strict indicates whether addresses with the AF_UNSPEC family are accepted of not.
-//
 // AddressAndFamily returns an address and its family.
-func AddressAndFamily(sfamily int, addr []byte, strict bool) (tcpip.FullAddress, uint16, *syserr.Error) {
+func AddressAndFamily(addr []byte) (tcpip.FullAddress, uint16, *syserr.Error) {
 	// Make sure we have at least 2 bytes for the address family.
 	if len(addr) < 2 {
 		return tcpip.FullAddress{}, 0, syserr.ErrInvalidArgument
 	}
 
-	family := usermem.ByteOrder.Uint16(addr)
-	if family != uint16(sfamily) && (strict || family != linux.AF_UNSPEC) {
-		return tcpip.FullAddress{}, family, syserr.ErrAddressFamilyNotSupported
-	}
-
 	// Get the rest of the fields based on the address family.
-	switch family {
+	switch family := usermem.ByteOrder.Uint16(addr); family {
 	case linux.AF_UNIX:
 		path := addr[2:]
 		if len(path) > linux.UnixPathMax {
@@ -638,10 +631,40 @@ func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
 	return r
 }
 
+func (s *SocketOperations) checkFamily(family uint16, exact bool) *syserr.Error {
+	if family == uint16(s.family) {
+		return nil
+	}
+	if !exact && family == linux.AF_INET && s.family == linux.AF_INET6 {
+		v, err := s.Endpoint.GetSockOptBool(tcpip.V6OnlyOption)
+		if err != nil {
+			return syserr.TranslateNetstackError(err)
+		}
+		if !v {
+			return nil
+		}
+	}
+	return syserr.ErrInvalidArgument
+}
+
+// mapFamily maps the AF_INET ANY address to the IPv4-mapped IPv6 ANY if the
+// receiver's family is AF_INET6.
+//
+// This is a hack to work around the fact that both IPv4 and IPv6 ANY are
+// represented by the empty string.
+//
+// TODO(gvisor.dev/issues/1556): remove this function.
+func (s *SocketOperations) mapFamily(addr tcpip.FullAddress, family uint16) tcpip.FullAddress {
+	if len(addr.Addr) == 0 && s.family == linux.AF_INET6 && family == linux.AF_INET {
+		addr.Addr = "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\xff\xff\x00\x00\x00\x00"
+	}
+	return addr
+}
+
 // Connect implements the linux syscall connect(2) for sockets backed by
 // tpcip.Endpoint.
 func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
-	addr, family, err := AddressAndFamily(s.family, sockaddr, false /* strict */)
+	addr, family, err := AddressAndFamily(sockaddr)
 	if err != nil {
 		return err
 	}
@@ -653,6 +676,12 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo
 		}
 		return syserr.TranslateNetstackError(err)
 	}
+
+	if err := s.checkFamily(family, false /* exact */); err != nil {
+		return err
+	}
+	addr = s.mapFamily(addr, family)
+
 	// Always return right away in the non-blocking case.
 	if !blocking {
 		return syserr.TranslateNetstackError(s.Endpoint.Connect(addr))
@@ -681,10 +710,14 @@ func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking boo
 // Bind implements the linux syscall bind(2) for sockets backed by
 // tcpip.Endpoint.
 func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
-	addr, _, err := AddressAndFamily(s.family, sockaddr, true /* strict */)
+	addr, family, err := AddressAndFamily(sockaddr)
 	if err != nil {
 		return err
 	}
+	if err := s.checkFamily(family, true /* exact */); err != nil {
+		return err
+	}
+	addr = s.mapFamily(addr, family)
 
 	// Issue the bind request to the endpoint.
 	return syserr.TranslateNetstackError(s.Endpoint.Bind(addr))
@@ -2080,8 +2113,8 @@ func ConvertAddress(family int, addr tcpip.FullAddress) (linux.SockAddr, uint32)
 
 	case linux.AF_INET6:
 		var out linux.SockAddrInet6
-		if len(addr.Addr) == 4 {
-			// Copy address is v4-mapped format.
+		if len(addr.Addr) == header.IPv4AddressSize {
+			// Copy address in v4-mapped format.
 			copy(out.Addr[12:], addr.Addr)
 			out.Addr[10] = 0xff
 			out.Addr[11] = 0xff
@@ -2395,10 +2428,14 @@ func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []
 
 	var addr *tcpip.FullAddress
 	if len(to) > 0 {
-		addrBuf, _, err := AddressAndFamily(s.family, to, true /* strict */)
+		addrBuf, family, err := AddressAndFamily(to)
 		if err != nil {
 			return 0, err
 		}
+		if err := s.checkFamily(family, false /* exact */); err != nil {
+			return 0, err
+		}
+		addrBuf = s.mapFamily(addrBuf, family)
 
 		addr = &addrBuf
 	}
diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go
index 91effe89a..7f49ba864 100644
--- a/pkg/sentry/socket/unix/unix.go
+++ b/pkg/sentry/socket/unix/unix.go
@@ -116,13 +116,16 @@ func (s *SocketOperations) Endpoint() transport.Endpoint {
 
 // extractPath extracts and validates the address.
 func extractPath(sockaddr []byte) (string, *syserr.Error) {
-	addr, _, err := netstack.AddressAndFamily(linux.AF_UNIX, sockaddr, true /* strict */)
+	addr, family, err := netstack.AddressAndFamily(sockaddr)
 	if err != nil {
 		if err == syserr.ErrAddressFamilyNotSupported {
 			err = syserr.ErrInvalidArgument
 		}
 		return "", err
 	}
+	if family != linux.AF_UNIX {
+		return "", syserr.ErrInvalidArgument
+	}
 
 	// The address is trimmed by GetAddress.
 	p := string(addr.Addr)
diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go
index 51f2efb39..b6d7177f4 100644
--- a/pkg/sentry/strace/socket.go
+++ b/pkg/sentry/strace/socket.go
@@ -341,7 +341,7 @@ func sockAddr(t *kernel.Task, addr usermem.Addr, length uint32) string {
 
 	switch family {
 	case linux.AF_INET, linux.AF_INET6, linux.AF_UNIX:
-		fa, _, err := netstack.AddressAndFamily(int(family), b, true /* strict */)
+		fa, _, err := netstack.AddressAndFamily(b)
 		if err != nil {
 			return fmt.Sprintf("%#x {Family: %s, error extracting address: %v}", addr, familyStr, err)
 		}
diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go
index a47ceba54..113b457fb 100644
--- a/pkg/tcpip/stack/stack.go
+++ b/pkg/tcpip/stack/stack.go
@@ -547,6 +547,49 @@ type TransportEndpointInfo struct {
 	RegisterNICID tcpip.NICID
 }
 
+// AddrNetProto unwraps the specified address if it is a V4-mapped V6 address
+// and returns the network protocol number to be used to communicate with the
+// specified address. It returns an error if the passed address is incompatible
+// with the receiver.
+func (e *TransportEndpointInfo) AddrNetProto(addr tcpip.FullAddress, v6only bool) (tcpip.FullAddress, tcpip.NetworkProtocolNumber, *tcpip.Error) {
+	netProto := e.NetProto
+	switch len(addr.Addr) {
+	case header.IPv4AddressSize:
+		netProto = header.IPv4ProtocolNumber
+	case header.IPv6AddressSize:
+		if header.IsV4MappedAddress(addr.Addr) {
+			netProto = header.IPv4ProtocolNumber
+			addr.Addr = addr.Addr[header.IPv6AddressSize-header.IPv4AddressSize:]
+			if addr.Addr == header.IPv4Any {
+				addr.Addr = ""
+			}
+		}
+	}
+
+	switch len(e.ID.LocalAddress) {
+	case header.IPv4AddressSize:
+		if len(addr.Addr) == header.IPv6AddressSize {
+			return tcpip.FullAddress{}, 0, tcpip.ErrInvalidEndpointState
+		}
+	case header.IPv6AddressSize:
+		if len(addr.Addr) == header.IPv4AddressSize {
+			return tcpip.FullAddress{}, 0, tcpip.ErrNetworkUnreachable
+		}
+	}
+
+	switch {
+	case netProto == e.NetProto:
+	case netProto == header.IPv4ProtocolNumber && e.NetProto == header.IPv6ProtocolNumber:
+		if v6only {
+			return tcpip.FullAddress{}, 0, tcpip.ErrNoRoute
+		}
+	default:
+		return tcpip.FullAddress{}, 0, tcpip.ErrInvalidEndpointState
+	}
+
+	return addr, netProto, nil
+}
+
 // IsEndpointInfo is an empty method to implement the tcpip.EndpointInfo
 // marker interface.
 func (*TransportEndpointInfo) IsEndpointInfo() {}
diff --git a/pkg/tcpip/transport/icmp/endpoint.go b/pkg/tcpip/transport/icmp/endpoint.go
index 330786f4c..42afb3f5b 100644
--- a/pkg/tcpip/transport/icmp/endpoint.go
+++ b/pkg/tcpip/transport/icmp/endpoint.go
@@ -288,7 +288,7 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c
 
 		toCopy := *to
 		to = &toCopy
-		netProto, err := e.checkV4Mapped(to, true)
+		netProto, err := e.checkV4Mapped(to)
 		if err != nil {
 			return 0, nil, err
 		}
@@ -475,18 +475,12 @@ func send6(r *stack.Route, ident uint16, data buffer.View, ttl uint8) *tcpip.Err
 	})
 }
 
-func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress, allowMismatch bool) (tcpip.NetworkProtocolNumber, *tcpip.Error) {
-	netProto := e.NetProto
-	if header.IsV4MappedAddress(addr.Addr) {
-		return 0, tcpip.ErrNoRoute
-	}
-
-	// Fail if we're bound to an address length different from the one we're
-	// checking.
-	if l := len(e.ID.LocalAddress); !allowMismatch && l != 0 && l != len(addr.Addr) {
-		return 0, tcpip.ErrInvalidEndpointState
+func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) {
+	unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, false /* v6only */)
+	if err != nil {
+		return 0, err
 	}
-
+	*addr = unwrapped
 	return netProto, nil
 }
 
@@ -518,7 +512,7 @@ func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error {
 		return tcpip.ErrInvalidEndpointState
 	}
 
-	netProto, err := e.checkV4Mapped(&addr, false)
+	netProto, err := e.checkV4Mapped(&addr)
 	if err != nil {
 		return err
 	}
@@ -631,7 +625,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) *tcpip.Error {
 		return tcpip.ErrInvalidEndpointState
 	}
 
-	netProto, err := e.checkV4Mapped(&addr, false)
+	netProto, err := e.checkV4Mapped(&addr)
 	if err != nil {
 		return err
 	}
diff --git a/pkg/tcpip/transport/tcp/endpoint.go b/pkg/tcpip/transport/tcp/endpoint.go
index cca511fb9..cc8b533c8 100644
--- a/pkg/tcpip/transport/tcp/endpoint.go
+++ b/pkg/tcpip/transport/tcp/endpoint.go
@@ -1691,26 +1691,11 @@ func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error {
 }
 
 func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) {
-	netProto := e.NetProto
-	if header.IsV4MappedAddress(addr.Addr) {
-		// Fail if using a v4 mapped address on a v6only endpoint.
-		if e.v6only {
-			return 0, tcpip.ErrNoRoute
-		}
-
-		netProto = header.IPv4ProtocolNumber
-		addr.Addr = addr.Addr[header.IPv6AddressSize-header.IPv4AddressSize:]
-		if addr.Addr == header.IPv4Any {
-			addr.Addr = ""
-		}
-	}
-
-	// Fail if we're bound to an address length different from the one we're
-	// checking.
-	if l := len(e.ID.LocalAddress); l != 0 && len(addr.Addr) != 0 && l != len(addr.Addr) {
-		return 0, tcpip.ErrInvalidEndpointState
+	unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, e.v6only)
+	if err != nil {
+		return 0, err
 	}
-
+	*addr = unwrapped
 	return netProto, nil
 }
 
diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go
index a4ff29a7d..13446f5d9 100644
--- a/pkg/tcpip/transport/udp/endpoint.go
+++ b/pkg/tcpip/transport/udp/endpoint.go
@@ -402,7 +402,7 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, <-c
 			return 0, nil, tcpip.ErrBroadcastDisabled
 		}
 
-		netProto, err := e.checkV4Mapped(to, false)
+		netProto, err := e.checkV4Mapped(to)
 		if err != nil {
 			return 0, nil, err
 		}
@@ -501,7 +501,7 @@ func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error {
 		defer e.mu.Unlock()
 
 		fa := tcpip.FullAddress{Addr: v.InterfaceAddr}
-		netProto, err := e.checkV4Mapped(&fa, false)
+		netProto, err := e.checkV4Mapped(&fa)
 		if err != nil {
 			return err
 		}
@@ -839,35 +839,12 @@ func sendUDP(r *stack.Route, data buffer.VectorisedView, localPort, remotePort u
 	return nil
 }
 
-func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress, allowMismatch bool) (tcpip.NetworkProtocolNumber, *tcpip.Error) {
-	netProto := e.NetProto
-	if len(addr.Addr) == 0 {
-		return netProto, nil
-	}
-	if header.IsV4MappedAddress(addr.Addr) {
-		// Fail if using a v4 mapped address on a v6only endpoint.
-		if e.v6only {
-			return 0, tcpip.ErrNoRoute
-		}
-
-		netProto = header.IPv4ProtocolNumber
-		addr.Addr = addr.Addr[header.IPv6AddressSize-header.IPv4AddressSize:]
-		if addr.Addr == header.IPv4Any {
-			addr.Addr = ""
-		}
-
-		// Fail if we are bound to an IPv6 address.
-		if !allowMismatch && len(e.ID.LocalAddress) == 16 {
-			return 0, tcpip.ErrNetworkUnreachable
-		}
-	}
-
-	// Fail if we're bound to an address length different from the one we're
-	// checking.
-	if l := len(e.ID.LocalAddress); l != 0 && l != len(addr.Addr) {
-		return 0, tcpip.ErrInvalidEndpointState
+func (e *endpoint) checkV4Mapped(addr *tcpip.FullAddress) (tcpip.NetworkProtocolNumber, *tcpip.Error) {
+	unwrapped, netProto, err := e.TransportEndpointInfo.AddrNetProto(*addr, e.v6only)
+	if err != nil {
+		return 0, err
 	}
-
+	*addr = unwrapped
 	return netProto, nil
 }
 
@@ -916,7 +893,7 @@ func (e *endpoint) Disconnect() *tcpip.Error {
 
 // Connect connects the endpoint to its peer. Specifying a NIC is optional.
 func (e *endpoint) Connect(addr tcpip.FullAddress) *tcpip.Error {
-	netProto, err := e.checkV4Mapped(&addr, false)
+	netProto, err := e.checkV4Mapped(&addr)
 	if err != nil {
 		return err
 	}
@@ -1074,7 +1051,7 @@ func (e *endpoint) bindLocked(addr tcpip.FullAddress) *tcpip.Error {
 		return tcpip.ErrInvalidEndpointState
 	}
 
-	netProto, err := e.checkV4Mapped(&addr, true)
+	netProto, err := e.checkV4Mapped(&addr)
 	if err != nil {
 		return err
 	}