Support sending with packet sockets

...through the loopback interface, only.

This change only supports sending on packet sockets through the loopback
interface as the loopback interface is the only interface used in packet
socket syscall tests - the other link endpoints are not excercised with
the existing test infrastructure.

Support for sending on packet sockets through the other interfaces will
be added as needed.

BUG: https://fxbug.dev/81592
PiperOrigin-RevId: 394368899

6 files changed, 52 insertions, 14 deletions

diff --git a/pkg/tcpip/transport/icmp/endpoint.go b/pkg/tcpip/transport/icmp/endpoint.go
index f9a15efb2..00497bf07 100644
--- a/pkg/tcpip/transport/icmp/endpoint.go
+++ b/pkg/tcpip/transport/icmp/endpoint.go
@@ -329,6 +329,7 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, tcp
 		route = r
 	}
 
+	// TODO(https://gvisor.dev/issue/6538): Avoid this allocation.
 	v := make([]byte, p.Len())
 	if _, err := io.ReadFull(p, v); err != nil {
 		return 0, &tcpip.ErrBadBuffer{}
diff --git a/pkg/tcpip/transport/packet/endpoint.go b/pkg/tcpip/transport/packet/endpoint.go
index 8e7bb6c6e..89b4720aa 100644
--- a/pkg/tcpip/transport/packet/endpoint.go
+++ b/pkg/tcpip/transport/packet/endpoint.go
@@ -207,8 +207,52 @@ func (ep *endpoint) Read(dst io.Writer, opts tcpip.ReadOptions) (tcpip.ReadResul
 	return res, nil
 }
 
-func (*endpoint) Write(tcpip.Payloader, tcpip.WriteOptions) (int64, tcpip.Error) {
-	return 0, &tcpip.ErrInvalidOptionValue{}
+func (ep *endpoint) Write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, tcpip.Error) {
+	if !ep.stack.PacketEndpointWriteSupported() {
+		return 0, &tcpip.ErrNotSupported{}
+	}
+
+	ep.mu.Lock()
+	closed := ep.closed
+	nicID := ep.boundNIC
+	ep.mu.Unlock()
+	if closed {
+		return 0, &tcpip.ErrClosedForSend{}
+	}
+
+	var remote tcpip.LinkAddress
+	proto := ep.netProto
+	if to := opts.To; to != nil {
+		remote = tcpip.LinkAddress(to.Addr)
+
+		if n := to.NIC; n != 0 {
+			nicID = n
+		}
+
+		if p := to.Port; p != 0 {
+			proto = tcpip.NetworkProtocolNumber(p)
+		}
+	}
+
+	if nicID == 0 {
+		return 0, &tcpip.ErrInvalidOptionValue{}
+	}
+
+	// TODO(https://gvisor.dev/issue/6538): Avoid this allocation.
+	payloadBytes := make(buffer.View, p.Len())
+	if _, err := io.ReadFull(p, payloadBytes); err != nil {
+		return 0, &tcpip.ErrBadBuffer{}
+	}
+
+	if err := func() tcpip.Error {
+		if ep.cooked {
+			return ep.stack.WritePacketToRemote(nicID, remote, proto, payloadBytes.ToVectorisedView())
+		}
+		return ep.stack.WriteRawPacket(nicID, proto, payloadBytes.ToVectorisedView())
+	}(); err != nil {
+		return 0, err
+	}
+	return int64(len(payloadBytes)), nil
 }
 
 // Disconnect implements tcpip.Endpoint.Disconnect. Packet sockets cannot be
diff --git a/pkg/tcpip/transport/packet/endpoint_state.go b/pkg/tcpip/transport/packet/endpoint_state.go
index e729921db..5c688d286 100644
--- a/pkg/tcpip/transport/packet/endpoint_state.go
+++ b/pkg/tcpip/transport/packet/endpoint_state.go
@@ -34,17 +34,11 @@ func (p *packet) loadReceivedAt(nsec int64) {
 
 // saveData saves packet.data field.
 func (p *packet) saveData() buffer.VectorisedView {
-	// We cannot save p.data directly as p.data.views may alias to p.views,
-	// which is not allowed by state framework (in-struct pointer).
 	return p.data.Clone(nil)
 }
 
 // loadData loads packet.data field.
 func (p *packet) loadData(data buffer.VectorisedView) {
-	// NOTE: We cannot do the p.data = data.Clone(p.views[:]) optimization
-	// here because data.views is not guaranteed to be loaded by now. Plus,
-	// data.views will be allocated anyway so there really is little point
-	// of utilizing p.views for data.views.
 	p.data = data
 }
 
diff --git a/pkg/tcpip/transport/raw/endpoint.go b/pkg/tcpip/transport/raw/endpoint.go
index 55854ba59..3bf6c0a8f 100644
--- a/pkg/tcpip/transport/raw/endpoint.go
+++ b/pkg/tcpip/transport/raw/endpoint.go
@@ -281,6 +281,7 @@ func (e *endpoint) write(p tcpip.Payloader, opts tcpip.WriteOptions) (int64, tcp
 			return nil, nil, nil, &tcpip.ErrInvalidEndpointState{}
 		}
 
+		// TODO(https://gvisor.dev/issue/6538): Avoid this allocation.
 		payloadBytes := make([]byte, p.Len())
 		if _, err := io.ReadFull(p, payloadBytes); err != nil {
 			return nil, nil, nil, &tcpip.ErrBadBuffer{}
@@ -600,6 +601,9 @@ func (e *endpoint) HandlePacket(pkt *stack.PacketBuffer) {
 	// We copy headers' underlying bytes because pkt.*Header may point to
 	// the middle of a slice, and another struct may point to the "outer"
 	// slice. Save/restore doesn't support overlapping slices and will fail.
+	//
+	// TODO(https://gvisor.dev/issue/6517): Avoid the copy once S/R supports
+	// overlapping slices.
 	var combinedVV buffer.VectorisedView
 	if e.TransportEndpointInfo.NetProto == header.IPv4ProtocolNumber {
 		network, transport := pkt.NetworkHeader().View(), pkt.TransportHeader().View()
diff --git a/pkg/tcpip/transport/udp/endpoint.go b/pkg/tcpip/transport/udp/endpoint.go
index ac7ecb5f8..4b6bdc3be 100644
--- a/pkg/tcpip/transport/udp/endpoint.go
+++ b/pkg/tcpip/transport/udp/endpoint.go
@@ -366,6 +366,7 @@ func (e *endpoint) prepareForWrite(p tcpip.Payloader, opts tcpip.WriteOptions) (
 		return udpPacketInfo{}, err
 	}
 
+	// TODO(https://gvisor.dev/issue/6538): Avoid this allocation.
 	v := make([]byte, p.Len())
 	if _, err := io.ReadFull(p, v); err != nil {
 		ctx.Release()
diff --git a/pkg/tcpip/transport/udp/endpoint_state.go b/pkg/tcpip/transport/udp/endpoint_state.go
index 20c45ab87..2ff8b0482 100644
--- a/pkg/tcpip/transport/udp/endpoint_state.go
+++ b/pkg/tcpip/transport/udp/endpoint_state.go
@@ -36,17 +36,11 @@ func (p *udpPacket) loadReceivedAt(nsec int64) {
 
 // saveData saves udpPacket.data field.
 func (p *udpPacket) saveData() buffer.VectorisedView {
-	// We cannot save p.data directly as p.data.views may alias to p.views,
-	// which is not allowed by state framework (in-struct pointer).
 	return p.data.Clone(nil)
 }
 
 // loadData loads udpPacket.data field.
 func (p *udpPacket) loadData(data buffer.VectorisedView) {
-	// NOTE: We cannot do the p.data = data.Clone(p.views[:]) optimization
-	// here because data.views is not guaranteed to be loaded by now. Plus,
-	// data.views will be allocated anyway so there really is little point
-	// of utilizing p.views for data.views.
 	p.data = data
 }