FIFO QDisc implementation

Updates #231

PiperOrigin-RevId: 309323808

5 files changed, 66 insertions, 3 deletions

diff --git a/runsc/boot/BUILD b/runsc/boot/BUILD
index ed3c8f546..abcaf4206 100644
--- a/runsc/boot/BUILD
+++ b/runsc/boot/BUILD
@@ -88,6 +88,7 @@ go_library(
         "//pkg/tcpip",
         "//pkg/tcpip/link/fdbased",
         "//pkg/tcpip/link/loopback",
+        "//pkg/tcpip/link/qdisc/fifo",
         "//pkg/tcpip/link/sniffer",
         "//pkg/tcpip/network/arp",
         "//pkg/tcpip/network/ipv4",
diff --git a/runsc/boot/config.go b/runsc/boot/config.go
index 715a19112..6d6a705f8 100644
--- a/runsc/boot/config.go
+++ b/runsc/boot/config.go
@@ -187,6 +187,10 @@ type Config struct {
 	// SoftwareGSO indicates that software segmentation offload is enabled.
 	SoftwareGSO bool
 
+	// QDisc indicates the type of queuening discipline to use by default
+	// for non-loopback interfaces.
+	QDisc QueueingDiscipline
+
 	// LogPackets indicates that all network packets should be logged.
 	LogPackets bool
 
@@ -294,6 +298,7 @@ func (c *Config) ToFlags() []string {
 		"--gso=" + strconv.FormatBool(c.HardwareGSO),
 		"--software-gso=" + strconv.FormatBool(c.SoftwareGSO),
 		"--overlayfs-stale-read=" + strconv.FormatBool(c.OverlayfsStaleRead),
+		"--qdisc=" + c.QDisc.String(),
 	}
 	if c.CPUNumFromQuota {
 		f = append(f, "--cpu-num-from-quota")
diff --git a/runsc/boot/network.go b/runsc/boot/network.go
index bee6ee336..0af30456e 100644
--- a/runsc/boot/network.go
+++ b/runsc/boot/network.go
@@ -17,6 +17,7 @@ package boot
 import (
 	"fmt"
 	"net"
+	"runtime"
 	"strings"
 	"syscall"
 
@@ -24,6 +25,7 @@ import (
 	"gvisor.dev/gvisor/pkg/tcpip"
 	"gvisor.dev/gvisor/pkg/tcpip/link/fdbased"
 	"gvisor.dev/gvisor/pkg/tcpip/link/loopback"
+	"gvisor.dev/gvisor/pkg/tcpip/link/qdisc/fifo"
 	"gvisor.dev/gvisor/pkg/tcpip/link/sniffer"
 	"gvisor.dev/gvisor/pkg/tcpip/network/arp"
 	"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
@@ -75,6 +77,44 @@ type DefaultRoute struct {
 	Name  string
 }
 
+// QueueingDiscipline is used to specify the kind of Queueing Discipline to
+// apply for a give FDBasedLink.
+type QueueingDiscipline int
+
+const (
+	// QDiscNone disables any queueing for the underlying FD.
+	QDiscNone QueueingDiscipline = iota
+
+	// QDiscFIFO applies a simple fifo based queue to the underlying
+	// FD.
+	QDiscFIFO
+)
+
+// MakeQueueingDiscipline if possible the equivalent QueuingDiscipline for s
+// else returns an error.
+func MakeQueueingDiscipline(s string) (QueueingDiscipline, error) {
+	switch s {
+	case "none":
+		return QDiscNone, nil
+	case "fifo":
+		return QDiscFIFO, nil
+	default:
+		return 0, fmt.Errorf("unsupported qdisc specified: %q", s)
+	}
+}
+
+// String implements fmt.Stringer.
+func (q QueueingDiscipline) String() string {
+	switch q {
+	case QDiscNone:
+		return "none"
+	case QDiscFIFO:
+		return "fifo"
+	default:
+		panic(fmt.Sprintf("Invalid queueing discipline: %d", q))
+	}
+}
+
 // FDBasedLink configures an fd-based link.
 type FDBasedLink struct {
 	Name               string
@@ -84,6 +124,7 @@ type FDBasedLink struct {
 	GSOMaxSize         uint32
 	SoftwareGSOEnabled bool
 	LinkAddress        net.HardwareAddr
+	QDisc              QueueingDiscipline
 
 	// NumChannels controls how many underlying FD's are to be used to
 	// create this endpoint.
@@ -185,6 +226,8 @@ func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct
 		}
 
 		mac := tcpip.LinkAddress(link.LinkAddress)
+		log.Infof("gso max size is: %d", link.GSOMaxSize)
+
 		linkEP, err := fdbased.New(&fdbased.Options{
 			FDs:                FDs,
 			MTU:                uint32(link.MTU),
@@ -199,6 +242,13 @@ func (n *Network) CreateLinksAndRoutes(args *CreateLinksAndRoutesArgs, _ *struct
 			return err
 		}
 
+		switch link.QDisc {
+		case QDiscNone:
+		case QDiscFIFO:
+			log.Infof("Enabling FIFO QDisc on %q", link.Name)
+			linkEP = fifo.New(linkEP, runtime.GOMAXPROCS(0), 1000)
+		}
+
 		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); err != nil {
 			return err
diff --git a/runsc/main.go b/runsc/main.go
index 8e594c58e..0216e9481 100644
--- a/runsc/main.go
+++ b/runsc/main.go
@@ -72,6 +72,7 @@ var (
 	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.")
 	hardwareGSO        = flag.Bool("gso", true, "enable hardware segmentation offload if it is supported by a network device.")
 	softwareGSO        = flag.Bool("software-gso", true, "enable software segmentation offload when hardware ofload can't be enabled.")
+	qDisc              = flag.String("qdisc", "none", "specifies which queueing discipline to apply by default to the non loopback nics used by the sandbox.")
 	fileAccess         = flag.String("file-access", "exclusive", "specifies which filesystem to use for the root mount: exclusive (default), shared. Volume mounts are always shared.")
 	fsGoferHostUDS     = flag.Bool("fsgofer-host-uds", false, "allow the gofer to mount Unix Domain Sockets.")
 	overlay            = flag.Bool("overlay", false, "wrap filesystem mounts with writable overlay. All modifications are stored in memory inside the sandbox.")
@@ -198,6 +199,11 @@ func main() {
 		cmd.Fatalf("%v", err)
 	}
 
+	queueingDiscipline, err := boot.MakeQueueingDiscipline(*qDisc)
+	if err != nil {
+		cmd.Fatalf("%s", err)
+	}
+
 	// Sets the reference leak check mode. Also set it in config below to
 	// propagate it to child processes.
 	refs.SetLeakMode(refsLeakMode)
@@ -232,7 +238,7 @@ func main() {
 		OverlayfsStaleRead: *overlayfsStaleRead,
 		CPUNumFromQuota:    *cpuNumFromQuota,
 		VFS2:               *vfs2Enabled,
-
+		QDisc:              queueingDiscipline,
 		TestOnlyAllowRunAsCurrentUserWithoutChroot: *testOnlyAllowRunAsCurrentUserWithoutChroot,
 		TestOnlyTestNameEnv:                        *testOnlyTestNameEnv,
 	}
diff --git a/runsc/sandbox/network.go b/runsc/sandbox/network.go
index bc093fba5..209bfdb20 100644
--- a/runsc/sandbox/network.go
+++ b/runsc/sandbox/network.go
@@ -62,7 +62,7 @@ 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.HardwareGSO, conf.SoftwareGSO, conf.NumNetworkChannels); err != nil {
+		if err := createInterfacesAndRoutesFromNS(conn, nsPath, conf.HardwareGSO, conf.SoftwareGSO, conf.NumNetworkChannels, conf.QDisc); err != nil {
 			return fmt.Errorf("creating interfaces from net namespace %q: %v", nsPath, err)
 		}
 	case boot.NetworkHost:
@@ -115,7 +115,7 @@ 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, hardwareGSO bool, softwareGSO bool, numNetworkChannels int) error {
+func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, hardwareGSO bool, softwareGSO bool, numNetworkChannels int, qDisc boot.QueueingDiscipline) error {
 	// Join the network namespace that we will be copying.
 	restore, err := joinNetNS(nsPath)
 	if err != nil {
@@ -201,6 +201,7 @@ func createInterfacesAndRoutesFromNS(conn *urpc.Client, nsPath string, hardwareG
 			MTU:         iface.MTU,
 			Routes:      routes,
 			NumChannels: numNetworkChannels,
+			QDisc:       qDisc,
 		}
 
 		// Get the link for the interface.