Add multi-fd support to fdbased endpoint.

This allows an fdbased endpoint to have multiple underlying fd's from which
packets can be read and dispatched/written to.

This should allow for higher throughput as well as better scalability of the
network stack as number of connections increases.

Updates #231

PiperOrigin-RevId: 251852825

5 files changed, 123 insertions, 67 deletions

diff --git a/runsc/boot/config.go b/runsc/boot/config.go
index 15f624f9b..8564c502d 100644
--- a/runsc/boot/config.go
+++ b/runsc/boot/config.go
@@ -221,6 +221,11 @@ type Config struct {
 	// user, and without chrooting the sandbox process. This can be
 	// necessary in test environments that have limited capabilities.
 	TestOnlyAllowRunAsCurrentUserWithoutChroot bool
+
+	// NumNetworkChannels controls the number of AF_PACKET sockets that map
+	// to the same underlying network device. This allows netstack to better
+	// scale for high throughput use cases.
+	NumNetworkChannels int
 }
 
 // ToFlags returns a slice of flags that correspond to the given Config.
@@ -244,6 +249,7 @@ func (c *Config) ToFlags() []string {
 		"--panic-signal=" + strconv.Itoa(c.PanicSignal),
 		"--profile=" + strconv.FormatBool(c.ProfileEnable),
 		"--net-raw=" + strconv.FormatBool(c.EnableRaw),
+		"--num-network-channels=" + strconv.Itoa(c.NumNetworkChannels),
 	}
 	if c.TestOnlyAllowRunAsCurrentUserWithoutChroot {
 		// Only include if set since it is never to be used by users.
diff --git a/runsc/boot/network.go b/runsc/boot/network.go
index 0a154d90b..82c259f47 100644
--- a/runsc/boot/network.go
+++ b/runsc/boot/network.go
@@ -57,6 +57,10 @@ type FDBasedLink struct {
 	Routes      []Route
 	GSOMaxSize  uint32
 	LinkAddress []byte
+
+	// NumChannels controls how many underlying FD's are to be used to
+	// create this endpoint.
+	NumChannels int
 }
 
 // LoopbackLink configures a loopback li nk.
@@ -68,8 +72,9 @@ type LoopbackLink struct {
 
 // CreateLinksAndRoutesArgs are arguments to CreateLinkAndRoutes.
 type CreateLinksAndRoutesArgs struct {
-	// FilePayload contains the fds associated with the FDBasedLinks.  The
-	// two slices must have the same length.
+	// FilePayload contains the fds associated with the FDBasedLinks. The
+	// number of fd's should match the sum of the NumChannels field of the
+	// FDBasedLink entries below.
 	urpc.FilePayload
 
 	LoopbackLinks []LoopbackLink
@@ -95,8 +100,12 @@ func (r *Route) toTcpipRoute(id tcpip.NICID) tcpip.Route {
 // CreateLinksAndRoutes creates links and routes in a network stack.  It should
 // only be called once.
 func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct{}) error {
-	if len(args.FilePayload.Files) != len(args.FDBasedLinks) {
-		return fmt.Errorf("FilePayload must be same length at FDBasedLinks")
+	wantFDs := 0
+	for _, l := range args.FDBasedLinks {
+		wantFDs += l.NumChannels
+	}
+	if got := len(args.FilePayload.Files); got != wantFDs {
+		return fmt.Errorf("args.FilePayload.Files has %d FD's but we need %d entries based on FDBasedLinks", got, wantFDs)
 	}
 
 	var nicID tcpip.NICID
@@ -123,20 +132,26 @@ func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct
 		}
 	}
 
-	for i, link := range args.FDBasedLinks {
+	fdOffset := 0
+	for _, link := range args.FDBasedLinks {
 		nicID++
 		nicids[link.Name] = nicID
 
-		// Copy the underlying FD.
-		oldFD := args.FilePayload.Files[i].Fd()
-		newFD, err := syscall.Dup(int(oldFD))
-		if err != nil {
-			return fmt.Errorf("failed to dup FD %v: %v", oldFD, err)
+		FDs := []int{}
+		for j := 0; j < link.NumChannels; j++ {
+			// Copy the underlying FD.
+			oldFD := args.FilePayload.Files[fdOffset].Fd()
+			newFD, err := syscall.Dup(int(oldFD))
+			if err != nil {
+				return fmt.Errorf("failed to dup FD %v: %v", oldFD, err)
+			}
+			FDs = append(FDs, newFD)
+			fdOffset++
 		}
 
 		mac := tcpip.LinkAddress(link.LinkAddress)
 		linkEP, err := fdbased.New(&fdbased.Options{
-			FD:                 newFD,
+			FDs:                FDs,
 			MTU:                uint32(link.MTU),
 			EthernetHeader:     true,
 			Address:            mac,
@@ -148,7 +163,7 @@ func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct
 			return err
 		}
 
-		log.Infof("Enabling interface %q with id %d on addresses %+v (%v)", link.Name, nicID, link.Addresses, mac)
+		log.Infof("Enabling interface %q with id %d on addresses %+v (%v) w/ %d channels", link.Name, nicID, link.Addresses, mac, link.NumChannels)
 		if err := n.createNICWithAddrs(nicID, link.Name, linkEP, link.Addresses, false /* loopback */); err != nil {
 			return err
 		}
diff --git a/runsc/main.go b/runsc/main.go
index 11bc73f75..44ad23cba 100644
--- a/runsc/main.go
+++ b/runsc/main.go
@@ -60,16 +60,16 @@ var (
 	straceLogSize  = flag.Uint("strace-log-size", 1024, "default size (in bytes) to log data argument blobs")
 
 	// Flags that control sandbox runtime behavior.
-	platform       = flag.String("platform", "ptrace", "specifies which platform to use: ptrace (default), kvm")
-	network        = flag.String("network", "sandbox", "specifies which network to use: sandbox (default), host, none. Using network inside the sandbox is more secure because it's isolated from the host network.")
-	gso            = flag.Bool("gso", true, "enable generic segmenation offload")
-	fileAccess     = flag.String("file-access", "exclusive", "specifies which filesystem to use for the root mount: exclusive (default), shared. Volume mounts are always shared.")
-	overlay        = flag.Bool("overlay", false, "wrap filesystem mounts with writable overlay. All modifications are stored in memory inside the sandbox.")
-	watchdogAction = flag.String("watchdog-action", "log", "sets what action the watchdog takes when triggered: log (default), panic.")
-	panicSignal    = flag.Int("panic-signal", -1, "register signal handling that panics. Usually set to SIGUSR2(12) to troubleshoot hangs. -1 disables it.")
-	profile        = flag.Bool("profile", false, "prepares the sandbox to use Golang profiler. Note that enabling profiler loosens the seccomp protection added to the sandbox (DO NOT USE IN PRODUCTION).")
-	netRaw         = flag.Bool("net-raw", false, "enable raw sockets. When false, raw sockets are disabled by removing CAP_NET_RAW from containers (`runsc exec` will still be able to utilize raw sockets). Raw sockets allow malicious containers to craft packets and potentially attack the network.")
-
+	platform                                   = flag.String("platform", "ptrace", "specifies which platform to use: ptrace (default), kvm")
+	network                                    = flag.String("network", "sandbox", "specifies which network to use: sandbox (default), host, none. Using network inside the sandbox is more secure because it's isolated from the host network.")
+	gso                                        = flag.Bool("gso", true, "enable generic segmenation offload")
+	fileAccess                                 = flag.String("file-access", "exclusive", "specifies which filesystem to use for the root mount: exclusive (default), shared. Volume mounts are always shared.")
+	overlay                                    = flag.Bool("overlay", false, "wrap filesystem mounts with writable overlay. All modifications are stored in memory inside the sandbox.")
+	watchdogAction                             = flag.String("watchdog-action", "log", "sets what action the watchdog takes when triggered: log (default), panic.")
+	panicSignal                                = flag.Int("panic-signal", -1, "register signal handling that panics. Usually set to SIGUSR2(12) to troubleshoot hangs. -1 disables it.")
+	profile                                    = flag.Bool("profile", false, "prepares the sandbox to use Golang profiler. Note that enabling profiler loosens the seccomp protection added to the sandbox (DO NOT USE IN PRODUCTION).")
+	netRaw                                     = flag.Bool("net-raw", false, "enable raw sockets. When false, raw sockets are disabled by removing CAP_NET_RAW from containers (`runsc exec` will still be able to utilize raw sockets). Raw sockets allow malicious containers to craft packets and potentially attack the network.")
+	numNetworkChannels                         = flag.Int("num-network-channels", 1, "number of underlying channels(FDs) to use for network link endpoints.")
 	testOnlyAllowRunAsCurrentUserWithoutChroot = flag.Bool("TESTONLY-unsafe-nonroot", false, "TEST ONLY; do not ever use! This skips many security measures that isolate the host from the sandbox.")
 )
 
@@ -141,6 +141,10 @@ func main() {
 		cmd.Fatalf("%v", err)
 	}
 
+	if *numNetworkChannels <= 0 {
+		cmd.Fatalf("num_network_channels must be > 0, got: %d", *numNetworkChannels)
+	}
+
 	// Create a new Config from the flags.
 	conf := &boot.Config{
 		RootDir:        *rootDir,
@@ -162,6 +166,7 @@ func main() {
 		ProfileEnable:  *profile,
 		EnableRaw:      *netRaw,
 		TestOnlyAllowRunAsCurrentUserWithoutChroot: *testOnlyAllowRunAsCurrentUserWithoutChroot,
+		NumNetworkChannels:                         *numNetworkChannels,
 	}
 	if len(*straceSyscalls) != 0 {
 		conf.StraceSyscalls = strings.Split(*straceSyscalls, ",")
diff --git a/runsc/sandbox/network.go b/runsc/sandbox/network.go
index 0460d5f1a..1fd091514 100644
--- a/runsc/sandbox/network.go
+++ b/runsc/sandbox/network.go
@@ -68,7 +68,7 @@ func setupNetwork(conn *urpc.Client, pid int, spec *specs.Spec, conf *boot.Confi
 		// Build the path to the net namespace of the sandbox process.
 		// This is what we will copy.
 		nsPath := filepath.Join("/proc", strconv.Itoa(pid), "ns/net")
-		if err := createInterfacesAndRoutesFromNS(conn, nsPath, conf.GSO); err != nil {
+		if err := createInterfacesAndRoutesFromNS(conn, nsPath, conf.GSO, conf.NumNetworkChannels); err != nil {
 			return fmt.Errorf("creating interfaces from net namespace %q: %v", nsPath, err)
 		}
 	case boot.NetworkHost:
@@ -138,7 +138,7 @@ func isRootNS() (bool, error) {
 // createInterfacesAndRoutesFromNS scrapes the interface and routes from the
 // net namespace with the given path, creates them in the sandbox, and removes
 // them from the host.
-func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO bool) error {
+func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO bool, numNetworkChannels int) error {
 	// Join the network namespace that we will be copying.
 	restore, err := joinNetNS(nsPath)
 	if err != nil {
@@ -202,25 +202,6 @@ func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO
 			continue
 		}
 
-		// Create the socket.
-		const protocol = 0x0300 // htons(ETH_P_ALL)
-		fd, err := syscall.Socket(syscall.AF_PACKET, syscall.SOCK_RAW, protocol)
-		if err != nil {
-			return fmt.Errorf("unable to create raw socket: %v", err)
-		}
-		deviceFile := os.NewFile(uintptr(fd), "raw-device-fd")
-
-		// Bind to the appropriate device.
-		ll := syscall.SockaddrLinklayer{
-			Protocol: protocol,
-			Ifindex:  iface.Index,
-			Hatype:   0, // No ARP type.
-			Pkttype:  syscall.PACKET_OTHERHOST,
-		}
-		if err := syscall.Bind(fd, &ll); err != nil {
-			return fmt.Errorf("unable to bind to %q: %v", iface.Name, err)
-		}
-
 		// Scrape the routes before removing the address, since that
 		// will remove the routes as well.
 		routes, def, err := routesForIface(iface)
@@ -236,9 +217,10 @@ func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO
 		}
 
 		link := boot.FDBasedLink{
-			Name:   iface.Name,
-			MTU:    iface.MTU,
-			Routes: routes,
+			Name:        iface.Name,
+			MTU:         iface.MTU,
+			Routes:      routes,
+			NumChannels: numNetworkChannels,
 		}
 
 		// Get the link for the interface.
@@ -248,30 +230,23 @@ func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO
 		}
 		link.LinkAddress = []byte(ifaceLink.Attrs().HardwareAddr)
 
-		if enableGSO {
-			gso, err := isGSOEnabled(fd, iface.Name)
+		log.Debugf("Setting up network channels")
+		// Create the socket for the device.
+		for i := 0; i < link.NumChannels; i++ {
+			log.Debugf("Creating Channel %d", i)
+			socketEntry, err := createSocket(iface, ifaceLink, enableGSO)
 			if err != nil {
-				return fmt.Errorf("getting GSO for interface %q: %v", iface.Name, err)
+				return fmt.Errorf("failed to createSocket for %s : %v", iface.Name, err)
 			}
-			if gso {
-				if err := syscall.SetsockoptInt(fd, syscall.SOL_PACKET, unix.PACKET_VNET_HDR, 1); err != nil {
-					return fmt.Errorf("unable to enable the PACKET_VNET_HDR option: %v", err)
-				}
-				link.GSOMaxSize = ifaceLink.Attrs().GSOMaxSize
+			if i == 0 {
+				link.GSOMaxSize = socketEntry.gsoMaxSize
 			} else {
-				log.Infof("GSO not available in host.")
+				if link.GSOMaxSize != socketEntry.gsoMaxSize {
+					return fmt.Errorf("inconsistent gsoMaxSize %d and %d when creating multiple channels for same interface: %s",
+						link.GSOMaxSize, socketEntry.gsoMaxSize, iface.Name)
+				}
 			}
-		}
-
-		// Use SO_RCVBUFFORCE because on linux the receive buffer for an
-		// AF_PACKET socket is capped by "net.core.rmem_max". rmem_max
-		// defaults to a unusually low value of 208KB. This is too low
-		// for gVisor to be able to receive packets at high throughputs
-		// without incurring packet drops.
-		const rcvBufSize = 4 << 20 // 4MB.
-
-		if err := syscall.SetsockoptInt(fd, syscall.SOL_SOCKET, syscall.SO_RCVBUFFORCE, rcvBufSize); err != nil {
-			return fmt.Errorf("failed to increase socket rcv buffer to %d: %v", rcvBufSize, err)
+			args.FilePayload.Files = append(args.FilePayload.Files, socketEntry.deviceFile)
 		}
 
 		// Collect the addresses for the interface, enable forwarding,
@@ -285,7 +260,6 @@ func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO
 			}
 		}
 
-		args.FilePayload.Files = append(args.FilePayload.Files, deviceFile)
 		args.FDBasedLinks = append(args.FDBasedLinks, link)
 	}
 
@@ -296,6 +270,61 @@ func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, enableGSO
 	return nil
 }
 
+type socketEntry struct {
+	deviceFile *os.File
+	gsoMaxSize uint32
+}
+
+// createSocket creates an underlying AF_PACKET socket and configures it for use by
+// the sentry and returns an *os.File that wraps the underlying socket fd.
+func createSocket(iface net.Interface, ifaceLink netlink.Link, enableGSO bool) (*socketEntry, error) {
+	// Create the socket.
+	const protocol = 0x0300 // htons(ETH_P_ALL)
+	fd, err := syscall.Socket(syscall.AF_PACKET, syscall.SOCK_RAW, protocol)
+	if err != nil {
+		return nil, fmt.Errorf("unable to create raw socket: %v", err)
+	}
+	deviceFile := os.NewFile(uintptr(fd), "raw-device-fd")
+	// Bind to the appropriate device.
+	ll := syscall.SockaddrLinklayer{
+		Protocol: protocol,
+		Ifindex:  iface.Index,
+		Hatype:   0, // No ARP type.
+		Pkttype:  syscall.PACKET_OTHERHOST,
+	}
+	if err := syscall.Bind(fd, &ll); err != nil {
+		return nil, fmt.Errorf("unable to bind to %q: %v", iface.Name, err)
+	}
+
+	gsoMaxSize := uint32(0)
+	if enableGSO {
+		gso, err := isGSOEnabled(fd, iface.Name)
+		if err != nil {
+			return nil, fmt.Errorf("getting GSO for interface %q: %v", iface.Name, err)
+		}
+		if gso {
+			if err := syscall.SetsockoptInt(fd, syscall.SOL_PACKET, unix.PACKET_VNET_HDR, 1); err != nil {
+				return nil, fmt.Errorf("unable to enable the PACKET_VNET_HDR option: %v", err)
+			}
+			gsoMaxSize = ifaceLink.Attrs().GSOMaxSize
+		} else {
+			log.Infof("GSO not available in host.")
+		}
+	}
+
+	// Use SO_RCVBUFFORCE because on linux the receive buffer for an
+	// AF_PACKET socket is capped by "net.core.rmem_max". rmem_max
+	// defaults to a unusually low value of 208KB. This is too low
+	// for gVisor to be able to receive packets at high throughputs
+	// without incurring packet drops.
+	const rcvBufSize = 4 << 20 // 4MB.
+
+	if err := syscall.SetsockoptInt(fd, syscall.SOL_SOCKET, syscall.SO_RCVBUFFORCE, rcvBufSize); err != nil {
+		return nil, fmt.Errorf("failed to increase socket rcv buffer to %d: %v", rcvBufSize, err)
+	}
+	return &socketEntry{deviceFile, gsoMaxSize}, nil
+}
+
 // loopbackLinks collects the links for a loopback interface.
 func loopbackLinks(iface net.Interface, addrs []net.Addr) ([]boot.LoopbackLink, error) {
 	var links []boot.LoopbackLink
diff --git a/runsc/test/testutil/testutil.go b/runsc/test/testutil/testutil.go
index 9efb1ba8e..727b648a6 100644
--- a/runsc/test/testutil/testutil.go
+++ b/runsc/test/testutil/testutil.go
@@ -136,6 +136,7 @@ func TestConfig() *boot.Config {
 		Strace:     true,
 		FileAccess: boot.FileAccessExclusive,
 		TestOnlyAllowRunAsCurrentUserWithoutChroot: true,
+		NumNetworkChannels:                         1,
 	}
 }