rootless: support `podman network create` (CNI-in-slirp4netns)

Usage:
```
$ podman network create foo
$ podman run -d --name web --hostname web --network foo nginx:alpine
$ podman run --rm --network foo alpine wget -O - http://web.dns.podman
Connecting to web.dns.podman (10.88.4.6:80)
...
<h1>Welcome to nginx!</h1>
...
```

See contrib/rootless-cni-infra for the design.

Signed-off-by: Akihiro Suda <akihiro.suda.cz@hco.ntt.co.jp>

8 files changed, 367 insertions, 18 deletions

diff --git a/libpod/container_internal.go b/libpod/container_internal.go
index c41d81a2b..c3f07a48b 100644
--- a/libpod/container_internal.go
+++ b/libpod/container_internal.go
@@ -957,8 +957,10 @@ func (c *Container) completeNetworkSetup() error {
 	if err := c.syncContainer(); err != nil {
 		return err
 	}
-	if c.config.NetMode.IsSlirp4netns() {
+	if rootless.IsRootless() {
 		return c.runtime.setupRootlessNetNS(c)
+	} else if c.config.NetMode.IsSlirp4netns() {
+		return c.runtime.setupSlirp4netns(c)
 	}
 	if err := c.runtime.setupNetNS(c); err != nil {
 		return err
diff --git a/libpod/container_internal_linux.go b/libpod/container_internal_linux.go
index f789b0069..605b526a4 100644
--- a/libpod/container_internal_linux.go
+++ b/libpod/container_internal_linux.go
@@ -84,7 +84,11 @@ func (c *Container) prepare() error {
 		// Set up network namespace if not already set up
 		noNetNS := c.state.NetNS == nil
 		if c.config.CreateNetNS && noNetNS && !c.config.PostConfigureNetNS {
-			netNS, networkStatus, createNetNSErr = c.runtime.createNetNS(c)
+			if rootless.IsRootless() && len(c.config.Networks) > 0 {
+				netNS, networkStatus, createNetNSErr = AllocRootlessCNI(context.Background(), c)
+			} else {
+				netNS, networkStatus, createNetNSErr = c.runtime.createNetNS(c)
+			}
 			if createNetNSErr != nil {
 				return
 			}
@@ -98,8 +102,12 @@ func (c *Container) prepare() error {
 		}
 
 		// handle rootless network namespace setup
-		if noNetNS && c.config.NetMode.IsSlirp4netns() && !c.config.PostConfigureNetNS {
-			createNetNSErr = c.runtime.setupRootlessNetNS(c)
+		if noNetNS && !c.config.PostConfigureNetNS {
+			if rootless.IsRootless() {
+				createNetNSErr = c.runtime.setupRootlessNetNS(c)
+			} else if c.config.NetMode.IsSlirp4netns() {
+				createNetNSErr = c.runtime.setupSlirp4netns(c)
+			}
 		}
 	}()
 	// Mount storage if not mounted
diff --git a/libpod/container_validate.go b/libpod/container_validate.go
index d657e3549..b78168cd1 100644
--- a/libpod/container_validate.go
+++ b/libpod/container_validate.go
@@ -2,7 +2,6 @@ package libpod
 
 import (
 	"github.com/containers/podman/v2/libpod/define"
-	"github.com/containers/podman/v2/pkg/rootless"
 	spec "github.com/opencontainers/runtime-spec/specs-go"
 	"github.com/pkg/errors"
 )
@@ -68,16 +67,6 @@ func (c *Container) validate() error {
 		}
 	}
 
-	// Rootless has some requirements, compared to networks.
-	if rootless.IsRootless() {
-		if len(c.config.Networks) > 0 {
-			return errors.Wrapf(define.ErrInvalidArg, "cannot join CNI networks if running rootless")
-		}
-
-		// TODO: Should we make sure network mode is set to Slirp if set
-		// at all?
-	}
-
 	// Can only set static IP or MAC is creating a network namespace.
 	if !c.config.CreateNetNS && (c.config.StaticIP != nil || c.config.StaticMAC != nil) {
 		return errors.Wrapf(define.ErrInvalidArg, "cannot set static IP or MAC address if not creating a network namespace")
diff --git a/libpod/networking_linux.go b/libpod/networking_linux.go
index 6f266e5d6..c0508ce39 100644
--- a/libpod/networking_linux.go
+++ b/libpod/networking_linux.go
@@ -4,6 +4,7 @@ package libpod
 
 import (
 	"bytes"
+	"context"
 	"crypto/rand"
 	"fmt"
 	"io"
@@ -208,6 +209,20 @@ func checkSlirpFlags(path string) (*slirpFeatures, error) {
 
 // Configure the network namespace for a rootless container
 func (r *Runtime) setupRootlessNetNS(ctr *Container) error {
+	if ctr.config.NetMode.IsSlirp4netns() {
+		return r.setupSlirp4netns(ctr)
+	}
+	if len(ctr.config.Networks) > 0 {
+		// set up port forwarder for CNI-in-slirp4netns
+		netnsPath := ctr.state.NetNS.Path()
+		// TODO: support slirp4netns port forwarder as well
+		return r.setupRootlessPortMappingViaRLK(ctr, netnsPath)
+	}
+	return nil
+}
+
+// setupSlirp4netns can be called in rootful as well as in rootless
+func (r *Runtime) setupSlirp4netns(ctr *Container) error {
 	path := r.config.Engine.NetworkCmdPath
 
 	if path == "" {
@@ -711,7 +726,7 @@ func (r *Runtime) teardownNetNS(ctr *Container) error {
 
 	logrus.Debugf("Tearing down network namespace at %s for container %s", ctr.state.NetNS.Path(), ctr.ID())
 
-	// rootless containers do not use the CNI plugin
+	// rootless containers do not use the CNI plugin directly
 	if !rootless.IsRootless() && !ctr.config.NetMode.IsSlirp4netns() {
 		var requestedIP net.IP
 		if ctr.requestedIP != nil {
@@ -738,6 +753,13 @@ func (r *Runtime) teardownNetNS(ctr *Container) error {
 		}
 	}
 
+	// CNI-in-slirp4netns
+	if rootless.IsRootless() && len(ctr.config.Networks) != 0 {
+		if err := DeallocRootlessCNI(context.Background(), ctr); err != nil {
+			return errors.Wrapf(err, "error tearing down CNI-in-slirp4netns for container %s", ctr.ID())
+		}
+	}
+
 	// First unmount the namespace
 	if err := netns.UnmountNS(ctr.state.NetNS); err != nil {
 		return errors.Wrapf(err, "error unmounting network namespace for container %s", ctr.ID())
diff --git a/libpod/networking_unsupported.go b/libpod/networking_unsupported.go
index dd72a3fd8..76bb01424 100644
--- a/libpod/networking_unsupported.go
+++ b/libpod/networking_unsupported.go
@@ -8,6 +8,10 @@ func (r *Runtime) setupRootlessNetNS(ctr *Container) error {
 	return define.ErrNotImplemented
 }
 
+func (r *Runtime) setupSlirp4netns(ctr *Container) error {
+	return define.ErrNotImplemented
+}
+
 func (r *Runtime) setupNetNS(ctr *Container) error {
 	return define.ErrNotImplemented
 }
diff --git a/libpod/oci_conmon_linux.go b/libpod/oci_conmon_linux.go
index f66835771..bb138ca14 100644
--- a/libpod/oci_conmon_linux.go
+++ b/libpod/oci_conmon_linux.go
@@ -1086,7 +1086,7 @@ func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *Co
 	cmd.ExtraFiles = append(cmd.ExtraFiles, childSyncPipe, childStartPipe)
 	cmd.ExtraFiles = append(cmd.ExtraFiles, envFiles...)
 
-	if r.reservePorts && !ctr.config.NetMode.IsSlirp4netns() {
+	if r.reservePorts && !rootless.IsRootless() && !ctr.config.NetMode.IsSlirp4netns() {
 		ports, err := bindPorts(ctr.config.PortMappings)
 		if err != nil {
 			return err
@@ -1098,7 +1098,7 @@ func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *Co
 		cmd.ExtraFiles = append(cmd.ExtraFiles, ports...)
 	}
 
-	if ctr.config.NetMode.IsSlirp4netns() {
+	if ctr.config.NetMode.IsSlirp4netns() || rootless.IsRootless() {
 		if ctr.config.PostConfigureNetNS {
 			havePortMapping := len(ctr.Config().PortMappings) > 0
 			if havePortMapping {
diff --git a/libpod/rootless_cni_linux.go b/libpod/rootless_cni_linux.go
new file mode 100644
index 000000000..76dbfdcae
--- /dev/null
+++ b/libpod/rootless_cni_linux.go
@@ -0,0 +1,320 @@
+// +build linux
+
+package libpod
+
+import (
+	"bytes"
+	"context"
+	"io"
+	"path/filepath"
+	"runtime"
+
+	cnitypes "github.com/containernetworking/cni/pkg/types/current"
+	"github.com/containernetworking/plugins/pkg/ns"
+	"github.com/containers/podman/v2/libpod/define"
+	"github.com/containers/podman/v2/libpod/image"
+	"github.com/containers/podman/v2/pkg/util"
+	"github.com/containers/storage/pkg/lockfile"
+	"github.com/hashicorp/go-multierror"
+	spec "github.com/opencontainers/runtime-spec/specs-go"
+	"github.com/opencontainers/runtime-tools/generate"
+	"github.com/pkg/errors"
+	"github.com/sirupsen/logrus"
+)
+
+var rootlessCNIInfraImage = map[string]string{
+	// Built from ../contrib/rootless-cni-infra
+	// TODO: move to Podman's official quay
+	"amd64": "ghcr.io/akihirosuda/podman-rootless-cni-infra:gd34868a13-amd64",
+}
+
+const (
+	rootlessCNIInfraContainerNamespace = "podman-system"
+	rootlessCNIInfraContainerName      = "rootless-cni-infra"
+)
+
+// AllocRootlessCNI allocates a CNI netns inside the rootless CNI infra container.
+// Locks "rootless-cni-infra.lck".
+//
+// When the infra container is not running, it is created.
+//
+// AllocRootlessCNI does not lock c. c should be already locked.
+func AllocRootlessCNI(ctx context.Context, c *Container) (ns.NetNS, []*cnitypes.Result, error) {
+	if len(c.config.Networks) == 0 {
+		return nil, nil, errors.New("allocRootlessCNI shall not be called when len(c.config.Networks) == 0")
+	}
+	l, err := getRootlessCNIInfraLock(c.runtime)
+	if err != nil {
+		return nil, nil, err
+	}
+	l.Lock()
+	defer l.Unlock()
+	infra, err := ensureRootlessCNIInfraContainerRunning(ctx, c.runtime)
+	if err != nil {
+		return nil, nil, err
+	}
+	k8sPodName := getPodOrContainerName(c) // passed to CNI as K8S_POD_NAME
+	cniResults := make([]*cnitypes.Result, len(c.config.Networks))
+	for i, nw := range c.config.Networks {
+		cniRes, err := rootlessCNIInfraCallAlloc(infra, c.ID(), nw, k8sPodName)
+		if err != nil {
+			return nil, nil, err
+		}
+		cniResults[i] = cniRes
+	}
+	nsObj, err := rootlessCNIInfraGetNS(infra, c.ID())
+	if err != nil {
+		return nil, nil, err
+	}
+	logrus.Debugf("rootless CNI: container %q will join %q", c.ID(), nsObj.Path())
+	return nsObj, cniResults, nil
+}
+
+// DeallocRootlessCNI deallocates a CNI netns inside the rootless CNI infra container.
+// Locks "rootless-cni-infra.lck".
+//
+// When the infra container is no longer needed, it is removed.
+//
+// DeallocRootlessCNI does not lock c. c should be already locked.
+func DeallocRootlessCNI(ctx context.Context, c *Container) error {
+	if len(c.config.Networks) == 0 {
+		return errors.New("deallocRootlessCNI shall not be called when len(c.config.Networks) == 0")
+	}
+	l, err := getRootlessCNIInfraLock(c.runtime)
+	if err != nil {
+		return err
+	}
+	l.Lock()
+	defer l.Unlock()
+	infra, _ := getRootlessCNIInfraContainer(c.runtime)
+	if infra == nil {
+		return nil
+	}
+	var errs *multierror.Error
+	for _, nw := range c.config.Networks {
+		err := rootlessCNIInfraCallDelloc(infra, c.ID(), nw)
+		if err != nil {
+			errs = multierror.Append(errs, err)
+		}
+	}
+	if isIdle, err := rootlessCNIInfraIsIdle(infra); isIdle || err != nil {
+		if err != nil {
+			logrus.Warn(err)
+		}
+		logrus.Debugf("rootless CNI: removing infra container %q", infra.ID())
+		if err := c.runtime.removeContainer(ctx, infra, true, false, true); err != nil {
+			return err
+		}
+		logrus.Debugf("rootless CNI: removed infra container %q", infra.ID())
+	}
+	return errs.ErrorOrNil()
+}
+
+func getRootlessCNIInfraLock(r *Runtime) (lockfile.Locker, error) {
+	fname := filepath.Join(r.config.Engine.TmpDir, "rootless-cni-infra.lck")
+	return lockfile.GetLockfile(fname)
+}
+
+func getPodOrContainerName(c *Container) string {
+	pod, err := c.runtime.GetPod(c.PodID())
+	if err != nil || pod.config.Name == "" {
+		return c.Name()
+	}
+	return pod.config.Name
+}
+
+func rootlessCNIInfraCallAlloc(infra *Container, id, nw, k8sPodName string) (*cnitypes.Result, error) {
+	logrus.Debugf("rootless CNI: alloc %q, %q, %q", id, nw, k8sPodName)
+	var err error
+
+	_, err = rootlessCNIInfraExec(infra, "alloc", id, nw, k8sPodName)
+	if err != nil {
+		return nil, err
+	}
+	cniResStr, err := rootlessCNIInfraExec(infra, "print-cni-result", id, nw)
+	if err != nil {
+		return nil, err
+	}
+	var cniRes cnitypes.Result
+	if err := json.Unmarshal([]byte(cniResStr), &cniRes); err != nil {
+		return nil, errors.Wrapf(err, "unmarshaling as cnitypes.Result: %q", cniResStr)
+	}
+	return &cniRes, nil
+}
+
+func rootlessCNIInfraCallDelloc(infra *Container, id, nw string) error {
+	logrus.Debugf("rootless CNI: dealloc %q, %q", id, nw)
+	_, err := rootlessCNIInfraExec(infra, "dealloc", id, nw)
+	return err
+}
+
+func rootlessCNIInfraIsIdle(infra *Container) (bool, error) {
+	type isIdle struct {
+		Idle bool `json:"idle"`
+	}
+	resStr, err := rootlessCNIInfraExec(infra, "is-idle")
+	if err != nil {
+		return false, err
+	}
+	var res isIdle
+	if err := json.Unmarshal([]byte(resStr), &res); err != nil {
+		return false, errors.Wrapf(err, "unmarshaling as isIdle: %q", resStr)
+	}
+	return res.Idle, nil
+}
+
+func rootlessCNIInfraGetNS(infra *Container, id string) (ns.NetNS, error) {
+	type printNetnsPath struct {
+		Path string `json:"path"`
+	}
+	resStr, err := rootlessCNIInfraExec(infra, "print-netns-path", id)
+	if err != nil {
+		return nil, err
+	}
+	var res printNetnsPath
+	if err := json.Unmarshal([]byte(resStr), &res); err != nil {
+		return nil, errors.Wrapf(err, "unmarshaling as printNetnsPath: %q", resStr)
+	}
+	nsObj, err := ns.GetNS(res.Path)
+	if err != nil {
+		return nil, err
+	}
+	return nsObj, nil
+}
+
+func getRootlessCNIInfraContainer(r *Runtime) (*Container, error) {
+	containers, err := r.GetContainersWithoutLock(func(c *Container) bool {
+		return c.Namespace() == rootlessCNIInfraContainerNamespace &&
+			c.Name() == rootlessCNIInfraContainerName
+	})
+	if err != nil {
+		return nil, err
+	}
+	if len(containers) == 0 {
+		return nil, nil
+	}
+	return containers[0], nil
+}
+
+func ensureRootlessCNIInfraContainerRunning(ctx context.Context, r *Runtime) (*Container, error) {
+	c, err := getRootlessCNIInfraContainer(r)
+	if err != nil {
+		return nil, err
+	}
+	if c == nil {
+		return startRootlessCNIInfraContainer(ctx, r)
+	}
+	st, err := c.ContainerState()
+	if err != nil {
+		return nil, err
+	}
+	if st.State == define.ContainerStateRunning {
+		logrus.Debugf("rootless CNI: infra container %q is already running", c.ID())
+		return c, nil
+	}
+	logrus.Debugf("rootless CNI: infra container %q is %q, being started", c.ID(), st.State)
+	if err := c.initAndStart(ctx); err != nil {
+		return nil, err
+	}
+	logrus.Debugf("rootless CNI: infra container %q is running", c.ID())
+	return c, nil
+}
+
+func startRootlessCNIInfraContainer(ctx context.Context, r *Runtime) (*Container, error) {
+	imageName, ok := rootlessCNIInfraImage[runtime.GOARCH]
+	if !ok {
+		return nil, errors.Errorf("cannot find rootless-podman-network-sandbox image for %s", runtime.GOARCH)
+	}
+	logrus.Debugf("rootless CNI: ensuring image %q to exist", imageName)
+	newImage, err := r.ImageRuntime().New(ctx, imageName, "", "", nil, nil,
+		image.SigningOptions{}, nil, util.PullImageMissing)
+	if err != nil {
+		return nil, err
+	}
+	logrus.Debugf("rootless CNI: image %q is ready", imageName)
+
+	g, err := generate.New("linux")
+	if err != nil {
+		return nil, err
+	}
+	g.SetupPrivileged(true)
+	// Set --pid=host for ease of propagating "/proc/PID/ns/net" string
+	if err := g.RemoveLinuxNamespace(string(spec.PIDNamespace)); err != nil {
+		return nil, err
+	}
+	g.RemoveMount("/proc")
+	procMount := spec.Mount{
+		Destination: "/proc",
+		Type:        "bind",
+		Source:      "/proc",
+		Options:     []string{"rbind", "nosuid", "noexec", "nodev"},
+	}
+	g.AddMount(procMount)
+	// Mount CNI networks
+	etcCNINetD := spec.Mount{
+		Destination: "/etc/cni/net.d",
+		Type:        "bind",
+		Source:      r.config.Network.NetworkConfigDir,
+		Options:     []string{"ro"},
+	}
+	g.AddMount(etcCNINetD)
+	// FIXME: how to propagate ProcessArgs and Envs from Dockerfile?
+	g.SetProcessArgs([]string{"sleep", "infinity"})
+	g.AddProcessEnv("CNI_PATH", "/opt/cni/bin")
+	var options []CtrCreateOption
+	options = append(options, WithRootFSFromImage(newImage.ID(), imageName, imageName))
+	options = append(options, WithCtrNamespace(rootlessCNIInfraContainerNamespace))
+	options = append(options, WithName(rootlessCNIInfraContainerName))
+	options = append(options, WithPrivileged(true))
+	options = append(options, WithSecLabels([]string{"disable"}))
+	options = append(options, WithRestartPolicy("always"))
+	options = append(options, WithNetNS(nil, false, "slirp4netns", nil))
+	c, err := r.NewContainer(ctx, g.Config, options...)
+	if err != nil {
+		return nil, err
+	}
+	logrus.Debugf("rootless CNI infra container %q is created, now being started", c.ID())
+	if err := c.initAndStart(ctx); err != nil {
+		return nil, err
+	}
+	logrus.Debugf("rootless CNI: infra container %q is running", c.ID())
+
+	return c, nil
+}
+
+func rootlessCNIInfraExec(c *Container, args ...string) (string, error) {
+	cmd := "rootless-cni-infra"
+	var (
+		outB    bytes.Buffer
+		errB    bytes.Buffer
+		streams define.AttachStreams
+		config  ExecConfig
+	)
+	streams.OutputStream = &nopWriteCloser{Writer: &outB}
+	streams.ErrorStream = &nopWriteCloser{Writer: &errB}
+	streams.AttachOutput = true
+	streams.AttachError = true
+	config.Command = append([]string{cmd}, args...)
+	config.Privileged = true
+	logrus.Debugf("rootlessCNIInfraExec: c.ID()=%s, config=%+v, streams=%v, begin",
+		c.ID(), config, streams)
+	code, err := c.Exec(&config, &streams, nil)
+	logrus.Debugf("rootlessCNIInfraExec: c.ID()=%s, config=%+v, streams=%v, end (code=%d, err=%v)",
+		c.ID(), config, streams, code, err)
+	if err != nil {
+		return "", err
+	}
+	if code != 0 {
+		return "", errors.Errorf("command %s %v in container %s failed with status %d, stdout=%q, stderr=%q",
+			cmd, args, c.ID(), code, outB.String(), errB.String())
+	}
+	return outB.String(), nil
+}
+
+type nopWriteCloser struct {
+	io.Writer
+}
+
+func (nwc *nopWriteCloser) Close() error {
+	return nil
+}
diff --git a/libpod/runtime_ctr.go b/libpod/runtime_ctr.go
index fa91fe002..6c29e0577 100644
--- a/libpod/runtime_ctr.go
+++ b/libpod/runtime_ctr.go
@@ -770,7 +770,11 @@ func (r *Runtime) LookupContainer(idOrName string) (*Container, error) {
 func (r *Runtime) GetContainers(filters ...ContainerFilter) ([]*Container, error) {
 	r.lock.RLock()
 	defer r.lock.RUnlock()
+	return r.GetContainersWithoutLock(filters...)
+}
 
+// GetContainersWithoutLock is same as GetContainers but without lock
+func (r *Runtime) GetContainersWithoutLock(filters ...ContainerFilter) ([]*Container, error) {
 	if !r.valid {
 		return nil, define.ErrRuntimeStopped
 	}