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

10 files changed, 162 insertions, 55 deletions

diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go
index 658e90bb9..83b9d9389 100644
--- a/pkg/sentry/socket/socket.go
+++ b/pkg/sentry/socket/socket.go
@@ -750,6 +750,7 @@ func AddressAndFamily(addr []byte) (tcpip.FullAddress, uint16, *syserr.Error) {
 		return tcpip.FullAddress{
 			NIC:  tcpip.NICID(a.InterfaceIndex),
 			Addr: tcpip.Address(a.HardwareAddr[:header.EthernetAddressSize]),
+			Port: Ntohs(a.Protocol),
 		}, family, nil
 
 	case linux.AF_UNSPEC:
diff --git a/pkg/tcpip/link/loopback/loopback.go b/pkg/tcpip/link/loopback/loopback.go
index d7bbfa639..ca1f9c08d 100644
--- a/pkg/tcpip/link/loopback/loopback.go
+++ b/pkg/tcpip/link/loopback/loopback.go
@@ -76,19 +76,8 @@ func (*endpoint) Wait() {}
 
 // WritePacket implements stack.LinkEndpoint.WritePacket. It delivers outbound
 // packets to the network-layer dispatcher.
-func (e *endpoint) WritePacket(_ stack.RouteInfo, protocol tcpip.NetworkProtocolNumber, pkt *stack.PacketBuffer) tcpip.Error {
-	// Construct data as the unparsed portion for the loopback packet.
-	data := buffer.NewVectorisedView(pkt.Size(), pkt.Views())
-
-	// Because we're immediately turning around and writing the packet back
-	// to the rx path, we intentionally don't preserve the remote and local
-	// link addresses from the stack.Route we're passed.
-	newPkt := stack.NewPacketBuffer(stack.PacketBufferOptions{
-		Data: data,
-	})
-	e.dispatcher.DeliverNetworkPacket("" /* remote */, "" /* local */, protocol, newPkt)
-
-	return nil
+func (e *endpoint) WritePacket(_ stack.RouteInfo, _ tcpip.NetworkProtocolNumber, pkt *stack.PacketBuffer) tcpip.Error {
+	return e.WriteRawPacket(pkt)
 }
 
 // WritePackets implements stack.LinkEndpoint.WritePackets.
@@ -105,4 +94,17 @@ func (e *endpoint) AddHeader(local, remote tcpip.LinkAddress, protocol tcpip.Net
 }
 
 // WriteRawPacket implements stack.LinkEndpoint.
-func (*endpoint) WriteRawPacket(*stack.PacketBuffer) tcpip.Error { return &tcpip.ErrNotSupported{} }
+func (e *endpoint) WriteRawPacket(pkt *stack.PacketBuffer) tcpip.Error {
+	// Construct data as the unparsed portion for the loopback packet.
+	data := buffer.NewVectorisedView(pkt.Size(), pkt.Views())
+
+	// Because we're immediately turning around and writing the packet back
+	// to the rx path, we intentionally don't preserve the remote and local
+	// link addresses from the stack.Route we're passed.
+	newPkt := stack.NewPacketBuffer(stack.PacketBufferOptions{
+		Data: data,
+	})
+	e.dispatcher.DeliverNetworkPacket("" /* remote */, "" /* local */, pkt.NetworkProtocolNumber, newPkt)
+
+	return nil
+}
diff --git a/pkg/tcpip/stack/nic.go b/pkg/tcpip/stack/nic.go
index b854d868c..ddc1ddab6 100644
--- a/pkg/tcpip/stack/nic.go
+++ b/pkg/tcpip/stack/nic.go
@@ -734,10 +734,29 @@ func (n *nic) DeliverNetworkPacket(remote, local tcpip.LinkAddress, protocol tcp
 	n.mu.RUnlock()
 
 	// Deliver to interested packet endpoints without holding NIC lock.
+	var packetEPPkt *PacketBuffer
 	deliverPacketEPs := func(ep PacketEndpoint) {
-		p := pkt.Clone()
-		p.PktType = tcpip.PacketHost
-		ep.HandlePacket(n.id, local, protocol, p)
+		if packetEPPkt == nil {
+			// Packet endpoints hold the full packet.
+			//
+			// We perform a deep copy because higher-level endpoints may point to
+			// the middle of a view that is held by a packet endpoint. Save/Restore
+			// does not support overlapping slices and will panic in this case.
+			//
+			// TODO(https://gvisor.dev/issue/6517): Avoid this copy once S/R supports
+			// overlapping slices (e.g. by passing a shallow copy of pkt to the packet
+			// endpoint).
+			packetEPPkt = NewPacketBuffer(PacketBufferOptions{
+				Data: PayloadSince(pkt.LinkHeader()).ToVectorisedView(),
+			})
+			// If a link header was populated in the original packet buffer, then
+			// populate it in the packet buffer we provide to packet endpoints as
+			// packet endpoints inspect link headers.
+			packetEPPkt.LinkHeader().Consume(pkt.LinkHeader().View().Size())
+			packetEPPkt.PktType = tcpip.PacketHost
+		}
+
+		ep.HandlePacket(n.id, local, protocol, packetEPPkt.Clone())
 	}
 	if protoEPs != nil {
 		protoEPs.forEach(deliverPacketEPs)
@@ -758,13 +777,30 @@ func (n *nic) DeliverOutboundPacket(remote, local tcpip.LinkAddress, protocol tc
 	eps := n.mu.packetEPs[header.EthernetProtocolAll]
 	n.mu.RUnlock()
 
+	var packetEPPkt *PacketBuffer
 	eps.forEach(func(ep PacketEndpoint) {
-		p := pkt.Clone()
-		p.PktType = tcpip.PacketOutgoing
-		// Add the link layer header as outgoing packets are intercepted
-		// before the link layer header is created.
-		n.LinkEndpoint.AddHeader(local, remote, protocol, p)
-		ep.HandlePacket(n.id, local, protocol, p)
+		if packetEPPkt == nil {
+			// Packet endpoints hold the full packet.
+			//
+			// We perform a deep copy because higher-level endpoints may point to
+			// the middle of a view that is held by a packet endpoint. Save/Restore
+			// does not support overlapping slices and will panic in this case.
+			//
+			// TODO(https://gvisor.dev/issue/6517): Avoid this copy once S/R supports
+			// overlapping slices (e.g. by passing a shallow copy of pkt to the packet
+			// endpoint).
+			packetEPPkt = NewPacketBuffer(PacketBufferOptions{
+				ReserveHeaderBytes: pkt.AvailableHeaderBytes(),
+				Data:               PayloadSince(pkt.NetworkHeader()).ToVectorisedView(),
+			})
+			// Add the link layer header as outgoing packets are intercepted before
+			// the link layer header is created and packet endpoints are interested
+			// in the link header.
+			n.LinkEndpoint.AddHeader(local, remote, protocol, packetEPPkt)
+			packetEPPkt.PktType = tcpip.PacketOutgoing
+		}
+
+		ep.HandlePacket(n.id, local, protocol, packetEPPkt.Clone())
 	})
 }
 
diff --git a/pkg/tcpip/stack/stack.go b/pkg/tcpip/stack/stack.go
index cfa8a2e8f..cb741e540 100644
--- a/pkg/tcpip/stack/stack.go
+++ b/pkg/tcpip/stack/stack.go
@@ -72,7 +72,8 @@ type Stack struct {
 
 	// rawFactory creates raw endpoints. If nil, raw endpoints are
 	// disabled. It is set during Stack creation and is immutable.
-	rawFactory RawFactory
+	rawFactory                   RawFactory
+	packetEndpointWriteSupported bool
 
 	demux *transportDemuxer
 
@@ -218,6 +219,10 @@ type Options struct {
 	// this is non-nil.
 	RawFactory RawFactory
 
+	// AllowPacketEndpointWrite determines if packet endpoints support write
+	// operations.
+	AllowPacketEndpointWrite bool
+
 	// RandSource is an optional source to use to generate random
 	// numbers. If omitted it defaults to a Source seeded by the data
 	// returned by the stack secure RNG.
@@ -359,23 +364,24 @@ func New(opts Options) *Stack {
 	opts.NUDConfigs.resetInvalidFields()
 
 	s := &Stack{
-		transportProtocols:       make(map[tcpip.TransportProtocolNumber]*transportProtocolState),
-		networkProtocols:         make(map[tcpip.NetworkProtocolNumber]NetworkProtocol),
-		nics:                     make(map[tcpip.NICID]*nic),
-		defaultForwardingEnabled: make(map[tcpip.NetworkProtocolNumber]struct{}),
-		cleanupEndpoints:         make(map[TransportEndpoint]struct{}),
-		PortManager:              ports.NewPortManager(),
-		clock:                    clock,
-		stats:                    opts.Stats.FillIn(),
-		handleLocal:              opts.HandleLocal,
-		tables:                   opts.IPTables,
-		icmpRateLimiter:          NewICMPRateLimiter(),
-		seed:                     seed,
-		nudConfigs:               opts.NUDConfigs,
-		uniqueIDGenerator:        opts.UniqueID,
-		nudDisp:                  opts.NUDDisp,
-		randomGenerator:          randomGenerator,
-		secureRNG:                opts.SecureRNG,
+		transportProtocols:           make(map[tcpip.TransportProtocolNumber]*transportProtocolState),
+		networkProtocols:             make(map[tcpip.NetworkProtocolNumber]NetworkProtocol),
+		nics:                         make(map[tcpip.NICID]*nic),
+		packetEndpointWriteSupported: opts.AllowPacketEndpointWrite,
+		defaultForwardingEnabled:     make(map[tcpip.NetworkProtocolNumber]struct{}),
+		cleanupEndpoints:             make(map[TransportEndpoint]struct{}),
+		PortManager:                  ports.NewPortManager(),
+		clock:                        clock,
+		stats:                        opts.Stats.FillIn(),
+		handleLocal:                  opts.HandleLocal,
+		tables:                       opts.IPTables,
+		icmpRateLimiter:              NewICMPRateLimiter(),
+		seed:                         seed,
+		nudConfigs:                   opts.NUDConfigs,
+		uniqueIDGenerator:            opts.UniqueID,
+		nudDisp:                      opts.NUDDisp,
+		randomGenerator:              randomGenerator,
+		secureRNG:                    opts.SecureRNG,
 		sendBufferSize: tcpip.SendBufferSizeOption{
 			Min:     MinBufferSize,
 			Default: DefaultBufferSize,
@@ -1653,9 +1659,27 @@ func (s *Stack) WritePacketToRemote(nicID tcpip.NICID, remote tcpip.LinkAddress,
 		ReserveHeaderBytes: int(nic.MaxHeaderLength()),
 		Data:               payload,
 	})
+	pkt.NetworkProtocolNumber = netProto
 	return nic.WritePacketToRemote(remote, netProto, pkt)
 }
 
+// WriteRawPacket writes data directly to the specified NIC without adding any
+// headers.
+func (s *Stack) WriteRawPacket(nicID tcpip.NICID, proto tcpip.NetworkProtocolNumber, payload buffer.VectorisedView) tcpip.Error {
+	s.mu.RLock()
+	nic, ok := s.nics[nicID]
+	s.mu.RUnlock()
+	if !ok {
+		return &tcpip.ErrUnknownNICID{}
+	}
+
+	pkt := NewPacketBuffer(PacketBufferOptions{
+		Data: payload,
+	})
+	pkt.NetworkProtocolNumber = proto
+	return nic.WriteRawPacket(pkt)
+}
+
 // NetworkProtocolInstance returns the protocol instance in the stack for the
 // specified network protocol. This method is public for protocol implementers
 // and tests to use.
@@ -1947,3 +1971,9 @@ func (s *Stack) IsSubnetBroadcast(nicID tcpip.NICID, protocol tcpip.NetworkProto
 
 	return false
 }
+
+// PacketEndpointWriteSupported returns true iff packet endpoints support write
+// operations.
+func (s *Stack) PacketEndpointWriteSupported() bool {
+	return s.packetEndpointWriteSupported
+}
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
 }