Initial network namespace support.

TCP/IP will work with netstack networking. hostinet doesn't work, and sockets
will have the same behavior as it is now.

Before the userspace is able to create device, the default loopback device can
be used to test.

/proc/net and /sys/net will still be connected to the root network stack; this
is the same behavior now.

Issue #1833

PiperOrigin-RevId: 296309389

6 files changed, 131 insertions, 47 deletions

diff --git a/runsc/boot/BUILD b/runsc/boot/BUILD
index ae4dd102a..26f68fe3d 100644
--- a/runsc/boot/BUILD
+++ b/runsc/boot/BUILD
@@ -19,7 +19,6 @@ go_library(
         "loader_amd64.go",
         "loader_arm64.go",
         "network.go",
-        "pprof.go",
         "strace.go",
         "user.go",
     ],
@@ -91,6 +90,7 @@ go_library(
         "//pkg/usermem",
         "//runsc/boot/filter",
         "//runsc/boot/platforms",
+        "//runsc/boot/pprof",
         "//runsc/specutils",
         "@com_github_golang_protobuf//proto:go_default_library",
         "@com_github_opencontainers_runtime-spec//specs-go:go_default_library",
diff --git a/runsc/boot/controller.go b/runsc/boot/controller.go
index 9c9e94864..17e774e0c 100644
--- a/runsc/boot/controller.go
+++ b/runsc/boot/controller.go
@@ -32,6 +32,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/watchdog"
 	"gvisor.dev/gvisor/pkg/tcpip/stack"
 	"gvisor.dev/gvisor/pkg/urpc"
+	"gvisor.dev/gvisor/runsc/boot/pprof"
 	"gvisor.dev/gvisor/runsc/specutils"
 )
 
@@ -142,7 +143,7 @@ func newController(fd int, l *Loader) (*controller, error) {
 	}
 	srv.Register(manager)
 
-	if eps, ok := l.k.NetworkStack().(*netstack.Stack); ok {
+	if eps, ok := l.k.RootNetworkNamespace().Stack().(*netstack.Stack); ok {
 		net := &Network{
 			Stack: eps.Stack,
 		}
@@ -341,7 +342,7 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error {
 		return fmt.Errorf("creating memory file: %v", err)
 	}
 	k.SetMemoryFile(mf)
-	networkStack := cm.l.k.NetworkStack()
+	networkStack := cm.l.k.RootNetworkNamespace().Stack()
 	cm.l.k = k
 
 	// Set up the restore environment.
@@ -365,9 +366,9 @@ func (cm *containerManager) Restore(o *RestoreOpts, _ *struct{}) error {
 	}
 
 	if cm.l.conf.ProfileEnable {
-		// initializePProf opens /proc/self/maps, so has to be
-		// called before installing seccomp filters.
-		initializePProf()
+		// pprof.Initialize opens /proc/self/maps, so has to be called before
+		// installing seccomp filters.
+		pprof.Initialize()
 	}
 
 	// Seccomp filters have to be applied before parsing the state file.
diff --git a/runsc/boot/loader.go b/runsc/boot/loader.go
index eef43b9df..e7ca98134 100644
--- a/runsc/boot/loader.go
+++ b/runsc/boot/loader.go
@@ -49,6 +49,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/watchdog"
 	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/tcpip"
+	"gvisor.dev/gvisor/pkg/tcpip/link/loopback"
 	"gvisor.dev/gvisor/pkg/tcpip/link/sniffer"
 	"gvisor.dev/gvisor/pkg/tcpip/network/arp"
 	"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
@@ -60,6 +61,7 @@ import (
 	"gvisor.dev/gvisor/pkg/tcpip/transport/udp"
 	"gvisor.dev/gvisor/runsc/boot/filter"
 	_ "gvisor.dev/gvisor/runsc/boot/platforms" // register all platforms.
+	"gvisor.dev/gvisor/runsc/boot/pprof"
 	"gvisor.dev/gvisor/runsc/specutils"
 
 	// Include supported socket providers.
@@ -230,11 +232,8 @@ func New(args Args) (*Loader, error) {
 		return nil, fmt.Errorf("enabling strace: %v", err)
 	}
 
-	// Create an empty network stack because the network namespace may be empty at
-	// this point. Netns is configured before Run() is called. Netstack is
-	// configured using a control uRPC message. Host network is configured inside
-	// Run().
-	networkStack, err := newEmptyNetworkStack(args.Conf, k, k)
+	// Create root network namespace/stack.
+	netns, err := newRootNetworkNamespace(args.Conf, k, k)
 	if err != nil {
 		return nil, fmt.Errorf("creating network: %v", err)
 	}
@@ -277,7 +276,7 @@ func New(args Args) (*Loader, error) {
 		FeatureSet:                  cpuid.HostFeatureSet(),
 		Timekeeper:                  tk,
 		RootUserNamespace:           creds.UserNamespace,
-		NetworkStack:                networkStack,
+		RootNetworkNamespace:        netns,
 		ApplicationCores:            uint(args.NumCPU),
 		Vdso:                        vdso,
 		RootUTSNamespace:            kernel.NewUTSNamespace(args.Spec.Hostname, args.Spec.Hostname, creds.UserNamespace),
@@ -466,7 +465,7 @@ func (l *Loader) run() error {
 		// Delay host network configuration to this point because network namespace
 		// is configured after the loader is created and before Run() is called.
 		log.Debugf("Configuring host network")
-		stack := l.k.NetworkStack().(*hostinet.Stack)
+		stack := l.k.RootNetworkNamespace().Stack().(*hostinet.Stack)
 		if err := stack.Configure(); err != nil {
 			return err
 		}
@@ -485,7 +484,7 @@ func (l *Loader) run() error {
 	// l.restore is set by the container manager when a restore call is made.
 	if !l.restore {
 		if l.conf.ProfileEnable {
-			initializePProf()
+			pprof.Initialize()
 		}
 
 		// Finally done with all configuration. Setup filters before user code
@@ -908,48 +907,92 @@ func (l *Loader) WaitExit() kernel.ExitStatus {
 	return l.k.GlobalInit().ExitStatus()
 }
 
-func newEmptyNetworkStack(conf *Config, clock tcpip.Clock, uniqueID stack.UniqueID) (inet.Stack, error) {
+func newRootNetworkNamespace(conf *Config, clock tcpip.Clock, uniqueID stack.UniqueID) (*inet.Namespace, error) {
+	// Create an empty network stack because the network namespace may be empty at
+	// this point. Netns is configured before Run() is called. Netstack is
+	// configured using a control uRPC message. Host network is configured inside
+	// Run().
 	switch conf.Network {
 	case NetworkHost:
-		return hostinet.NewStack(), nil
+		// No network namespacing support for hostinet yet, hence creator is nil.
+		return inet.NewRootNamespace(hostinet.NewStack(), nil), nil
 
 	case NetworkNone, NetworkSandbox:
-		// NetworkNone sets up loopback using netstack.
-		netProtos := []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol(), arp.NewProtocol()}
-		transProtos := []stack.TransportProtocol{tcp.NewProtocol(), udp.NewProtocol(), icmp.NewProtocol4()}
-		s := netstack.Stack{stack.New(stack.Options{
-			NetworkProtocols:   netProtos,
-			TransportProtocols: transProtos,
-			Clock:              clock,
-			Stats:              netstack.Metrics,
-			HandleLocal:        true,
-			// Enable raw sockets for users with sufficient
-			// privileges.
-			RawFactory: raw.EndpointFactory{},
-			UniqueID:   uniqueID,
-		})}
-
-		// Enable SACK Recovery.
-		if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(true)); err != nil {
-			return nil, fmt.Errorf("failed to enable SACK: %v", err)
+		s, err := newEmptySandboxNetworkStack(clock, uniqueID)
+		if err != nil {
+			return nil, err
 		}
+		creator := &sandboxNetstackCreator{
+			clock:    clock,
+			uniqueID: uniqueID,
+		}
+		return inet.NewRootNamespace(s, creator), nil
 
-		// Set default TTLs as required by socket/netstack.
-		s.Stack.SetNetworkProtocolOption(ipv4.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
-		s.Stack.SetNetworkProtocolOption(ipv6.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
+	default:
+		panic(fmt.Sprintf("invalid network configuration: %v", conf.Network))
+	}
 
-		// Enable Receive Buffer Auto-Tuning.
-		if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.ModerateReceiveBufferOption(true)); err != nil {
-			return nil, fmt.Errorf("SetTransportProtocolOption failed: %v", err)
-		}
+}
 
-		s.FillDefaultIPTables()
+func newEmptySandboxNetworkStack(clock tcpip.Clock, uniqueID stack.UniqueID) (inet.Stack, error) {
+	netProtos := []stack.NetworkProtocol{ipv4.NewProtocol(), ipv6.NewProtocol(), arp.NewProtocol()}
+	transProtos := []stack.TransportProtocol{tcp.NewProtocol(), udp.NewProtocol(), icmp.NewProtocol4()}
+	s := netstack.Stack{stack.New(stack.Options{
+		NetworkProtocols:   netProtos,
+		TransportProtocols: transProtos,
+		Clock:              clock,
+		Stats:              netstack.Metrics,
+		HandleLocal:        true,
+		// Enable raw sockets for users with sufficient
+		// privileges.
+		RawFactory: raw.EndpointFactory{},
+		UniqueID:   uniqueID,
+	})}
 
-		return &s, nil
+	// Enable SACK Recovery.
+	if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(true)); err != nil {
+		return nil, fmt.Errorf("failed to enable SACK: %v", err)
+	}
 
-	default:
-		panic(fmt.Sprintf("invalid network configuration: %v", conf.Network))
+	// Set default TTLs as required by socket/netstack.
+	s.Stack.SetNetworkProtocolOption(ipv4.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
+	s.Stack.SetNetworkProtocolOption(ipv6.ProtocolNumber, tcpip.DefaultTTLOption(netstack.DefaultTTL))
+
+	// Enable Receive Buffer Auto-Tuning.
+	if err := s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcpip.ModerateReceiveBufferOption(true)); err != nil {
+		return nil, fmt.Errorf("SetTransportProtocolOption failed: %v", err)
+	}
+
+	s.FillDefaultIPTables()
+
+	return &s, nil
+}
+
+// sandboxNetstackCreator implements kernel.NetworkStackCreator.
+//
+// +stateify savable
+type sandboxNetstackCreator struct {
+	clock    tcpip.Clock
+	uniqueID stack.UniqueID
+}
+
+// CreateStack implements kernel.NetworkStackCreator.CreateStack.
+func (f *sandboxNetstackCreator) CreateStack() (inet.Stack, error) {
+	s, err := newEmptySandboxNetworkStack(f.clock, f.uniqueID)
+	if err != nil {
+		return nil, err
 	}
+
+	// Setup loopback.
+	n := &Network{Stack: s.(*netstack.Stack).Stack}
+	nicID := tcpip.NICID(f.uniqueID.UniqueID())
+	link := DefaultLoopbackLink
+	linkEP := loopback.New()
+	if err := n.createNICWithAddrs(nicID, link.Name, linkEP, link.Addresses); err != nil {
+		return nil, err
+	}
+
+	return s, nil
 }
 
 // signal sends a signal to one or more processes in a container. If PID is 0,
diff --git a/runsc/boot/network.go b/runsc/boot/network.go
index 6a8765ec8..bee6ee336 100644
--- a/runsc/boot/network.go
+++ b/runsc/boot/network.go
@@ -17,6 +17,7 @@ package boot
 import (
 	"fmt"
 	"net"
+	"strings"
 	"syscall"
 
 	"gvisor.dev/gvisor/pkg/log"
@@ -31,6 +32,32 @@ import (
 	"gvisor.dev/gvisor/pkg/urpc"
 )
 
+var (
+	// DefaultLoopbackLink contains IP addresses and routes of "127.0.0.1/8" and
+	// "::1/8" on "lo" interface.
+	DefaultLoopbackLink = LoopbackLink{
+		Name: "lo",
+		Addresses: []net.IP{
+			net.IP("\x7f\x00\x00\x01"),
+			net.IPv6loopback,
+		},
+		Routes: []Route{
+			{
+				Destination: net.IPNet{
+					IP:   net.IPv4(0x7f, 0, 0, 0),
+					Mask: net.IPv4Mask(0xff, 0, 0, 0),
+				},
+			},
+			{
+				Destination: net.IPNet{
+					IP:   net.IPv6loopback,
+					Mask: net.IPMask(strings.Repeat("\xff", net.IPv6len)),
+				},
+			},
+		},
+	}
+)
+
 // Network exposes methods that can be used to configure a network stack.
 type Network struct {
 	Stack *stack.Stack
diff --git a/runsc/boot/pprof/BUILD b/runsc/boot/pprof/BUILD
new file mode 100644
index 000000000..29cb42b2f
--- /dev/null
+++ b/runsc/boot/pprof/BUILD
@@ -0,0 +1,11 @@
+load("//tools:defs.bzl", "go_library")
+
+package(licenses = ["notice"])
+
+go_library(
+    name = "pprof",
+    srcs = ["pprof.go"],
+    visibility = [
+        "//runsc:__subpackages__",
+    ],
+)
diff --git a/runsc/boot/pprof.go b/runsc/boot/pprof/pprof.go
index 463362f02..1ded20dee 100644
--- a/runsc/boot/pprof.go
+++ b/runsc/boot/pprof/pprof.go
@@ -12,7 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-package boot
+// Package pprof provides a stub to initialize custom profilers.
+package pprof
 
-func initializePProf() {
+// Initialize will be called at boot for initializing custom profilers.
+func Initialize() {
 }