Merge pull request #1004 from Honny1/passta-and-pessto

pesto: target address mapping and netavark integration

Author: mheon

common/libnetwork/internal/rootlessnetns/netns_linux.go

@@ -209,6 +209,19 @@ func (n *Netns) setupPasta(nsPath string) error {
 		extraOpts = append(extraOpts, "-c", n.getPath(pestoSocketFile))
 	}
 
+	// For the rootless netns, hard-code IPv6 --map-guest-addr, --address, and
+	// --gateway so pasta uses routable addresses for inbound IPv6 forwarding
+	// (instead of link-local which is not routable across bridges).
+	// The v4 --map-guest-addr must also be listed here because providing any
+	// --map-guest-addr suppresses the default one added by createPastaArgs.
+	// See: https://bugs.passt.top/show_bug.cgi?id=217
+	extraOpts = append(extraOpts,
+		"--map-guest-addr", pasta.MapGuestAddrIpv4,
+		"--map-guest-addr", pasta.MapGuestAddrIpv6,
+		"--address", pasta.GuestAddrIpv6,
+		"--gateway", pasta.GatewayIpv6,
+	)
+
 	pastaOpts := pasta.SetupOptions{
 		Config:       n.config,
 		Netns:        nsPath,

common/libnetwork/netavark/network.go

@@ -73,10 +73,13 @@ type netavarkNetwork struct {
 	rootlessNetns *rootlessnetns.Netns
 
 	// rootlessPortForwarder is the value of config.RootlessPortForwarder from
-	// containers.conf. When set to config.RootlessPortForwarderPasta, HostIP
-	// is stripped from port mappings before passing to netavark because pasta's
-	// splice changes the destination IP.
+	// containers.conf. When set to config.RootlessPortForwarderPasta, pesto
+	// handles port forwarding directly and netavark DNAT rules are skipped.
 	rootlessPortForwarder string
+
+	// config is the containers.conf config, needed for FindHelperBinary
+	// when invoking pesto for dynamic port forwarding.
+	config *config.Config
 }
 
 type InitConfig struct {
@@ -167,6 +170,7 @@ func NewNetworkInterface(conf *InitConfig) (types.ContainerNetwork, error) {
 		syslog:                conf.Syslog,
 		rootlessNetns:         netns,
 		rootlessPortForwarder: conf.Config.Network.RootlessPortForwarder,
+		config:                conf.Config,
 	}
 
 	return n, nil

common/libnetwork/netavark/run.go

@@ -11,6 +11,7 @@ import (
 	"strings"
 
 	"github.com/sirupsen/logrus"
+	"go.podman.io/common/libnetwork/pasta"
 	"go.podman.io/common/libnetwork/types"
 	"go.podman.io/common/pkg/config"
 )
@@ -143,9 +144,107 @@ func (n *netavarkNetwork) Setup(namespacePath string, options types.SetupOptions
 		return nil, fmt.Errorf("unexpected netavark result length, want (%d), got (%d) networks", len(options.Networks), len(result))
 	}
 
+	if n.rootlessPortForwarder == config.RootlessPortForwarderPasta && n.networkRootless && len(options.NetworkOptions.PortMappings) > 0 {
+		// If NetworkStatus was already populated by the caller, pesto setup
+		// already happened on a prior network connect — skip to avoid duplicate rules.
+		if options.NetworkOptions.NetworkStatus != nil {
+			return result, nil
+		}
+		opts := options.NetworkOptions
+		opts.NetworkStatus = result
+		if len(opts.NetworkOrder) == 0 {
+			opts.NetworkOrder = make([]string, 0, len(opts.Networks))
+			for _, net := range opts.Networks {
+				opts.NetworkOrder = append(opts.NetworkOrder, net.Name)
+			}
+		}
+		if err := n.pestoSetup(opts); err != nil {
+			return nil, err
+		}
+	}
+
 	return result, err
 }
 
+// pestoSetup publishes port mappings via pesto with target address mapping.
+// PestoAddPorts is idempotent so this is safe to call on every Setup.
+func (n *netavarkNetwork) pestoSetup(opts types.NetworkOptions) error {
+	if n.rootlessNetns == nil {
+		return nil
+	}
+
+	ipv4, _, ipv6, _ := firstIPsFromStatus(opts.NetworkStatus, opts.NetworkOrder)
+	if ipv4 == "" && ipv6 == "" {
+		return nil
+	}
+	return pasta.PestoAddPorts(n.config, n.PestoSocketPath(), opts.PortMappings, ipv4, ipv6)
+}
+
+// pestoTeardown handles pesto port mapping removal on network disconnect.
+// It derives the active forwarding IPs from the caller-provided NetworkStatus
+// (the podman db state) rather than maintaining a separate state file.
+//   - If NetworkStatus is empty: nothing to do.
+//   - If this is the last network: delete all mappings.
+//   - If the disconnected network supplied the active IP: delete old mappings,
+//     pick a replacement IP from the remaining networks, and re-add.
+//   - Otherwise: no port changes needed.
+func (n *netavarkNetwork) pestoTeardown(opts types.NetworkOptions) error {
+	if n.rootlessNetns == nil {
+		return nil
+	}
+
+	if len(opts.NetworkStatus) == 0 {
+		return nil
+	}
+
+	// Determine current active IPs from the full set of connected networks,
+	// respecting connection-time ordering from NetworkOrder.
+	activeIPv4, activeIPv4Net, activeIPv6, activeIPv6Net := firstIPsFromStatus(opts.NetworkStatus, opts.NetworkOrder)
+
+	// Build the set of networks being disconnected and the remaining status.
+	remaining := make(map[string]types.StatusBlock, len(opts.NetworkStatus))
+	disconnectedNets := make(map[string]struct{}, len(opts.Networks))
+	for _, namedNet := range opts.Networks {
+		disconnectedNets[namedNet.Name] = struct{}{}
+	}
+	for name, sb := range opts.NetworkStatus {
+		if _, removed := disconnectedNets[name]; !removed {
+			remaining[name] = sb
+		}
+	}
+
+	// Last network: remove all port mappings.
+	if len(remaining) == 0 {
+		return pasta.PestoDeletePorts(n.config, n.PestoSocketPath(), opts.PortMappings,
+			activeIPv4, activeIPv6)
+	}
+
+	// Check whether any active IP came from a network being disconnected.
+	_, v4Lost := disconnectedNets[activeIPv4Net]
+	v4Lost = v4Lost && activeIPv4Net != ""
+	_, v6Lost := disconnectedNets[activeIPv6Net]
+	v6Lost = v6Lost && activeIPv6Net != ""
+
+	if !v4Lost && !v6Lost {
+		return nil
+	}
+
+	// Remap: delete old mappings, pick new IPs from remaining networks, re-add.
+	if err := pasta.PestoDeletePorts(n.config, n.PestoSocketPath(), opts.PortMappings,
+		activeIPv4, activeIPv6); err != nil {
+		return fmt.Errorf("deleting port mappings for remap: %w", err)
+	}
+
+	newIPv4, _, newIPv6, _ := firstIPsFromStatus(remaining, opts.NetworkOrder)
+	if newIPv4 != "" || newIPv6 != "" {
+		if err := pasta.PestoAddPorts(n.config, n.PestoSocketPath(), opts.PortMappings,
+			newIPv4, newIPv6); err != nil {
+			return fmt.Errorf("re-adding port mappings after remap: %w", err)
+		}
+	}
+	return nil
+}
+
 // Teardown will teardown the container network namespace.
 func (n *netavarkNetwork) Teardown(namespacePath string, options types.TeardownOptions) error {
 	n.lock.Lock()
@@ -163,6 +262,19 @@ func (n *netavarkNetwork) Teardown(namespacePath string, options types.TeardownO
 		logrus.Error(err)
 	}
 
+	if n.rootlessPortForwarder == config.RootlessPortForwarderPasta && n.networkRootless && len(options.NetworkOptions.PortMappings) > 0 {
+		opts := options.NetworkOptions
+		if len(opts.NetworkOrder) == 0 {
+			opts.NetworkOrder = make([]string, 0, len(opts.Networks))
+			for _, net := range opts.Networks {
+				opts.NetworkOrder = append(opts.NetworkOrder, net.Name)
+			}
+		}
+		if err := n.pestoTeardown(opts); err != nil {
+			logrus.Errorf("pesto: %v", err)
+		}
+	}
+
 	netavarkOpts, needPlugin, err := n.convertNetOpts(options.NetworkOptions)
 	if err != nil {
 		return fmt.Errorf("failed to convert net opts: %w", err)
@@ -206,18 +318,6 @@ func (n *netavarkNetwork) getCommonNetavarkOptions(needPlugin bool) []string {
 }
 
 func (n *netavarkNetwork) convertNetOpts(opts types.NetworkOptions) (*netavarkOptions, bool, error) {
-	// In pasta mode, strip HostIP from port mappings. Pasta handles host-side
-	// address binding; netavark only needs DNAT rules inside the netns without
-	// "ip daddr" constraints (pasta's splice changes the destination IP).
-	if n.rootlessPortForwarder == config.RootlessPortForwarderPasta && n.networkRootless && len(opts.PortMappings) > 0 {
-		stripped := make([]types.PortMapping, len(opts.PortMappings))
-		copy(stripped, opts.PortMappings)
-		for i := range stripped {
-			stripped[i].HostIP = ""
-		}
-		opts.PortMappings = stripped
-	}
-
 	netavarkOptions := netavarkOptions{
 		NetworkOptions: opts,
 		Networks:       make(map[string]*types.Network, len(opts.Networks)),
@@ -266,3 +366,31 @@ func (n *netavarkNetwork) PestoSocketPath() string {
 	}
 	return n.rootlessNetns.PestoSocketPath()
 }
+
+// firstIPsFromStatus picks the first IPv4 and IPv6 container addresses from a
+// set of network StatusBlocks, iterating in the order given by networkOrder
+// (connection time). Returns the IP strings and the owning network name.
+func firstIPsFromStatus(status map[string]types.StatusBlock, networkOrder []string) (ipv4, ipv4Net, ipv6, ipv6Net string) {
+	for _, name := range networkOrder {
+		sb, ok := status[name]
+		if !ok {
+			continue
+		}
+		for _, netInt := range sb.Interfaces {
+			for _, netAddr := range netInt.Subnets {
+				ip := netAddr.IPNet.IP
+				if ip.To4() != nil && ipv4 == "" {
+					ipv4 = ip.String()
+					ipv4Net = name
+				} else if ip.To4() == nil && ipv6 == "" {
+					ipv6 = ip.String()
+					ipv6Net = name
+				}
+			}
+		}
+		if ipv4 != "" && ipv6 != "" {
+			break
+		}
+	}
+	return
+}

common/libnetwork/pasta/pasta_linux.go

@@ -36,6 +36,26 @@ const (
 	// mapGuestAddrIpv4 static ip used as forwarder address inside the netns to reach the host,
 	// given this is a "link local" ip it should be very unlikely that it causes conflicts.
 	mapGuestAddrIpv4 = "169.254.1.2"
+
+	// mapGuestAddrIpv6 static ip used as IPv6 forwarder address inside the netns to reach the host.
+	mapGuestAddrIpv6 = "fc00::2"
+
+	// gatewayIpv6 is the IPv6 default gateway for the guest. pasta uses this as the source
+	// address for inbound IPv6 forwarding. Must differ from mapGuestAddrIpv6 to avoid
+	// --map-guest-addr intercepting reply traffic.
+	// See: https://bugs.passt.top/show_bug.cgi?id=217
+	gatewayIpv6 = "fc00::1"
+
+	// guestAddrIpv6 is assigned to the guest interface so the IPv6 gateway is reachable.
+	guestAddrIpv6 = "fc00::3"
+)
+
+// Exported IPv6 address constants for use by the rootless netns setup.
+const (
+	MapGuestAddrIpv4 = mapGuestAddrIpv4
+	MapGuestAddrIpv6 = mapGuestAddrIpv6
+	GatewayIpv6      = gatewayIpv6
+	GuestAddrIpv6    = guestAddrIpv6
 )
 
 type SetupOptions struct {
@@ -67,39 +87,24 @@ func Setup(opts *SetupOptions) (*SetupResult, error) {
 
 	logrus.Debugf("pasta arguments: %s", strings.Join(cmdArgs, " "))
 
-	for {
-		// pasta forks once ready, and quits once we delete the target namespace
-		out, err := exec.Command(path, cmdArgs...).CombinedOutput()
-		if err != nil {
-			exitErr := &exec.ExitError{}
-			if errors.As(err, &exitErr) {
-				// special backwards compat check, --map-guest-addr was added in pasta version 20240814 so we
-				// cannot hard require it yet. Once we are confident that the update is most distros we can remove it.
-				if exitErr.ExitCode() == 1 &&
-					strings.Contains(string(out), "unrecognized option '"+mapGuestAddrOpt) &&
-					len(mapGuestAddrIPs) == 1 && mapGuestAddrIPs[0] == mapGuestAddrIpv4 {
-					// we did add the default --map-guest-addr option, if users set something different we want
-					// to get to the error below. We have to unset mapGuestAddrIPs here to avoid a infinite loop.
-					mapGuestAddrIPs = nil
-					// Trim off last two args which are --map-guest-addr 169.254.1.2.
-					cmdArgs = cmdArgs[:len(cmdArgs)-2]
-					continue
-				}
-				return nil, fmt.Errorf("pasta failed with exit code %d:\n%s",
-					exitErr.ExitCode(), string(out))
-			}
-			return nil, fmt.Errorf("failed to start pasta: %w", err)
+	// pasta forks once ready, and quits once we delete the target namespace
+	out, err := exec.Command(path, cmdArgs...).CombinedOutput()
+	if err != nil {
+		exitErr := &exec.ExitError{}
+		if errors.As(err, &exitErr) {
+			return nil, fmt.Errorf("pasta failed with exit code %d:\n%s",
+				exitErr.ExitCode(), string(out))
 		}
+		return nil, fmt.Errorf("failed to start pasta: %w", err)
+	}
 
-		if len(out) > 0 {
-			// TODO: This should be warning but as of August 2024 pasta still prints
-			// things with --quiet that we do not care about. In podman CI I still see
-			// "Couldn't get any nameserver address" so until this is fixed we cannot
-			// enable it. For now info is fine and we can bump it up later, it is only a
-			// nice to have.
-			logrus.Infof("pasta logged warnings: %q", strings.TrimSpace(string(out)))
-		}
-		break
+	if len(out) > 0 {
+		// TODO: This should be warning but as of August 2024 pasta still prints
+		// things with --quiet that we do not care about. In podman CI I still see
+		// "Couldn't get any nameserver address" so until this is fixed we cannot
+		// enable it. For now info is fine and we can bump it up later, it is only a
+		// nice to have.
+		logrus.Infof("pasta logged warnings: %q", strings.TrimSpace(string(out)))
 	}
 
 	var ipv4, ipv6 bool
@@ -265,15 +270,12 @@ func createPastaArgs(opts *SetupOptions) ([]string, []string, []string, error) {
 		cmdArgs = append(cmdArgs, "--quiet")
 	}
 
-	cmdArgs = append(cmdArgs, "--netns", opts.Netns)
-
-	// do this as last arg so we can easily trim them off in the error case when we have an older version
 	if len(mapGuestAddrIPs) == 0 {
-		// the user did not request custom --map-guest-addr so add our own so that we can use this
-		// for our own host.containers.internal host entry.
 		cmdArgs = append(cmdArgs, mapGuestAddrOpt, mapGuestAddrIpv4)
 		mapGuestAddrIPs = append(mapGuestAddrIPs, mapGuestAddrIpv4)
 	}
 
+	cmdArgs = append(cmdArgs, "--netns", opts.Netns)
+
 	return cmdArgs, dnsForwardIPs, mapGuestAddrIPs, nil
 }