Manage mount lifecycle and remove cached state

Signed-off-by: Lantao Liu <lantaol@google.com>

pkg/netns/netns.go

@@ -0,0 +1,220 @@
+/*
+Copyright 2018 The Containerd Authors.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+// Copyright 2018 CNI authors
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package netns
+
+import (
+	"crypto/rand"
+	"fmt"
+	"os"
+	"path"
+	"runtime"
+	"sync"
+
+	cnins "github.com/containernetworking/plugins/pkg/ns"
+	"github.com/docker/docker/pkg/symlink"
+	"github.com/pkg/errors"
+	"golang.org/x/sys/unix"
+
+	osinterface "github.com/containerd/cri/pkg/os"
+)
+
+const nsRunDir = "/var/run/netns"
+
+// Some of the following functions are migrated from
+// https://github.com/containernetworking/plugins/blob/master/pkg/testutils/netns_linux.go
+
+// newNS creates a new persistent (bind-mounted) network namespace and returns the
+// path to the network namespace.
+func newNS() (nsPath string, err error) {
+	b := make([]byte, 16)
+	if _, err := rand.Reader.Read(b); err != nil {
+		return "", errors.Wrap(err, "failed to generate random netns name")
+	}
+
+	// Create the directory for mounting network namespaces
+	// This needs to be a shared mountpoint in case it is mounted in to
+	// other namespaces (containers)
+	if err := os.MkdirAll(nsRunDir, 0755); err != nil {
+		return "", err
+	}
+
+	// create an empty file at the mount point
+	nsName := fmt.Sprintf("cni-%x-%x-%x-%x-%x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:])
+	nsPath = path.Join(nsRunDir, nsName)
+	mountPointFd, err := os.Create(nsPath)
+	if err != nil {
+		return "", err
+	}
+	mountPointFd.Close()
+
+	defer func() {
+		// Ensure the mount point is cleaned up on errors
+		if err != nil {
+			os.RemoveAll(nsPath) // nolint: errcheck
+		}
+	}()
+
+	var wg sync.WaitGroup
+	wg.Add(1)
+
+	// do namespace work in a dedicated goroutine, so that we can safely
+	// Lock/Unlock OSThread without upsetting the lock/unlock state of
+	// the caller of this function
+	go (func() {
+		defer wg.Done()
+		runtime.LockOSThread()
+		// Don't unlock. By not unlocking, golang will kill the OS thread when the
+		// goroutine is done (for go1.10+)
+
+		var origNS cnins.NetNS
+		origNS, err = cnins.GetNS(getCurrentThreadNetNSPath())
+		if err != nil {
+			return
+		}
+		defer origNS.Close()
+
+		// create a new netns on the current thread
+		err = unix.Unshare(unix.CLONE_NEWNET)
+		if err != nil {
+			return
+		}
+
+		// Put this thread back to the orig ns, since it might get reused (pre go1.10)
+		defer origNS.Set() // nolint: errcheck
+
+		// bind mount the netns from the current thread (from /proc) onto the
+		// mount point. This causes the namespace to persist, even when there
+		// are no threads in the ns.
+		err = unix.Mount(getCurrentThreadNetNSPath(), nsPath, "none", unix.MS_BIND, "")
+		if err != nil {
+			err = errors.Wrapf(err, "failed to bind mount ns at %s", nsPath)
+		}
+	})()
+	wg.Wait()
+
+	if err != nil {
+		return "", errors.Wrap(err, "failed to create namespace")
+	}
+
+	return nsPath, nil
+}
+
+// unmountNS unmounts the NS held by the netns object. unmountNS is idempotent.
+func unmountNS(path string) error {
+	if _, err := os.Stat(path); err != nil {
+		if os.IsNotExist(err) {
+			return nil
+		}
+		return errors.Wrap(err, "failed to stat netns")
+	}
+	path, err := symlink.FollowSymlinkInScope(path, "/")
+	if err != nil {
+		return errors.Wrap(err, "failed to follow symlink")
+	}
+	if err := osinterface.Unmount(path); err != nil && !os.IsNotExist(err) {
+		return errors.Wrap(err, "failed to umount netns")
+	}
+	if err := os.RemoveAll(path); err != nil {
+		return errors.Wrap(err, "failed to remove netns")
+	}
+	return nil
+}
+
+// getCurrentThreadNetNSPath copied from pkg/ns
+func getCurrentThreadNetNSPath() string {
+	// /proc/self/ns/net returns the namespace of the main thread, not
+	// of whatever thread this goroutine is running on.  Make sure we
+	// use the thread's net namespace since the thread is switching around
+	return fmt.Sprintf("/proc/%d/task/%d/ns/net", os.Getpid(), unix.Gettid())
+}
+
+// NetNS holds network namespace.
+type NetNS struct {
+	path string
+}
+
+// NewNetNS creates a network namespace.
+func NewNetNS() (*NetNS, error) {
+	path, err := newNS()
+	if err != nil {
+		return nil, errors.Wrap(err, "failed to setup netns")
+	}
+	return &NetNS{path: path}, nil
+}
+
+// LoadNetNS loads existing network namespace.
+func LoadNetNS(path string) *NetNS {
+	return &NetNS{path: path}
+}
+
+// Remove removes network namepace. Remove is idempotent, meaning it might
+// be invoked multiple times and provides consistent result.
+func (n *NetNS) Remove() error {
+	return unmountNS(n.path)
+}
+
+// Closed checks whether the network namespace has been closed.
+func (n *NetNS) Closed() (bool, error) {
+	ns, err := cnins.GetNS(n.path)
+	if err != nil {
+		if _, ok := err.(cnins.NSPathNotExistErr); ok {
+			// The network namespace has already been removed.
+			return true, nil
+		}
+		if _, ok := err.(cnins.NSPathNotNSErr); ok {
+			// The network namespace is not mounted, remove it.
+			if err := os.RemoveAll(n.path); err != nil {
+				return false, errors.Wrap(err, "remove netns")
+			}
+			return true, nil
+		}
+		return false, errors.Wrap(err, "get netns fd")
+	}
+	if err := ns.Close(); err != nil {
+		return false, errors.Wrap(err, "close netns fd")
+	}
+	return false, nil
+}
+
+// GetPath returns network namespace path for sandbox container
+func (n *NetNS) GetPath() string {
+	return n.path
+}
+
+// Do runs a function in the network namespace.
+func (n *NetNS) Do(f func(cnins.NetNS) error) error {
+	ns, err := cnins.GetNS(n.path)
+	if err != nil {
+		return errors.Wrap(err, "get netns fd")
+	}
+	defer ns.Close() // nolint: errcheck
+	return ns.Do(f)
+}

pkg/server/helpers.go

@@ -353,7 +353,7 @@ func checkSelinuxLevel(level string) (bool, error) {
 
 	matched, err := regexp.MatchString(`^s\d(-s\d)??(:c\d{1,4}((.c\d{1,4})?,c\d{1,4})*(.c\d{1,4})?(,c\d{1,4}(.c\d{1,4})?)*)?$`, level)
 	if err != nil || !matched {
-		return false, fmt.Errorf("the format of 'level' %q is not correct: %v", level, err)
+		return false, errors.Wrapf(err, "the format of 'level' %q is not correct", level)
 	}
 	return true, nil
 }

pkg/server/restart.go

@@ -34,6 +34,7 @@ import (
 	"golang.org/x/net/context"
 	runtime "k8s.io/kubernetes/pkg/kubelet/apis/cri/runtime/v1alpha2"
 
+	"github.com/containerd/cri/pkg/netns"
 	cio "github.com/containerd/cri/pkg/server/io"
 	containerstore "github.com/containerd/cri/pkg/store/container"
 	sandboxstore "github.com/containerd/cri/pkg/store/sandbox"
@@ -394,14 +395,7 @@ func loadSandbox(ctx context.Context, cntr containerd.Container) (sandboxstore.S
 		// Don't need to load netns for host network sandbox.
 		return sandbox, nil
 	}
-	netNS, err := sandboxstore.LoadNetNS(meta.NetNSPath)
-	if err != nil {
-		if err != sandboxstore.ErrClosedNetNS {
-			return sandbox, errors.Wrapf(err, "failed to load netns %q", meta.NetNSPath)
-		}
-		netNS = nil
-	}
-	sandbox.NetNS = netNS
+	sandbox.NetNS = netns.LoadNetNS(meta.NetNSPath)
 
 	// It doesn't matter whether task is running or not. If it is running, sandbox
 	// status will be `READY`; if it is not running, sandbox status will be `NOT_READY`,

pkg/server/sandbox_portforward.go

@@ -59,10 +59,12 @@ func (c *criService) portForward(id string, port int32, stream io.ReadWriteClose
 	securityContext := s.Config.GetLinux().GetSecurityContext()
 	hostNet := securityContext.GetNamespaceOptions().GetNetwork() == runtime.NamespaceMode_NODE
 	if !hostNet {
-		if s.NetNS == nil || s.NetNS.Closed() {
+		if closed, err := s.NetNS.Closed(); err != nil {
+			return errors.Wrapf(err, "failed to check netwok namespace closed for sandbox %q", id)
+		} else if closed {
 			return errors.Errorf("network namespace for sandbox %q is closed", id)
 		}
-		netNSDo = s.NetNS.GetNs().Do
+		netNSDo = s.NetNS.Do
 		netNSPath = s.NetNS.GetPath()
 	} else {
 		// Run the function directly for host network.

pkg/server/sandbox_remove.go

@@ -52,8 +52,13 @@ func (c *criService) RemovePodSandbox(ctx context.Context, r *runtime.RemovePodS
 	}
 
 	// Return error if sandbox network namespace is not closed yet.
-	if sandbox.NetNS != nil && !sandbox.NetNS.Closed() {
-		return nil, errors.Errorf("sandbox network namespace %q is not fully closed", sandbox.NetNS.GetPath())
+	if sandbox.NetNS != nil {
+		nsPath := sandbox.NetNS.GetPath()
+		if closed, err := sandbox.NetNS.Closed(); err != nil {
+			return nil, errors.Wrapf(err, "failed to check sandbox network namespace %q closed", nsPath)
+		} else if !closed {
+			return nil, errors.Errorf("sandbox network namespace %q is not fully closed", nsPath)
+		}
 	}
 
 	// Remove all containers inside the sandbox.

pkg/server/sandbox_run.go

@@ -41,6 +41,7 @@ import (
 	customopts "github.com/containerd/cri/pkg/containerd/opts"
 	ctrdutil "github.com/containerd/cri/pkg/containerd/util"
 	"github.com/containerd/cri/pkg/log"
+	"github.com/containerd/cri/pkg/netns"
 	sandboxstore "github.com/containerd/cri/pkg/store/sandbox"
 	"github.com/containerd/cri/pkg/util"
 )
@@ -104,7 +105,7 @@ func (c *criService) RunPodSandbox(ctx context.Context, r *runtime.RunPodSandbox
 		// handle. NetNSPath in sandbox metadata and NetNS is non empty only for non host network
 		// namespaces. If the pod is in host network namespace then both are empty and should not
 		// be used.
-		sandbox.NetNS, err = sandboxstore.NewNetNS()
+		sandbox.NetNS, err = netns.NewNetNS()
 		if err != nil {
 			return nil, errors.Wrapf(err, "failed to create network namespace for sandbox %q", id)
 		}

pkg/server/sandbox_status.go

@@ -36,7 +36,10 @@ func (c *criService) PodSandboxStatus(ctx context.Context, r *runtime.PodSandbox
 		return nil, errors.Wrap(err, "an error occurred when try to find sandbox")
 	}
 
-	ip := c.getIP(sandbox)
+	ip, err := c.getIP(sandbox)
+	if err != nil {
+		return nil, errors.Wrap(err, "failed to get sandbox ip")
+	}
 	status := toCRISandboxStatus(sandbox.Metadata, sandbox.Status.Get(), ip)
 	if !r.GetVerbose() {
 		return &runtime.PodSandboxStatusResponse{Status: status}, nil
@@ -54,21 +57,22 @@ func (c *criService) PodSandboxStatus(ctx context.Context, r *runtime.PodSandbox
 	}, nil
 }
 
-func (c *criService) getIP(sandbox sandboxstore.Sandbox) string {
+func (c *criService) getIP(sandbox sandboxstore.Sandbox) (string, error) {
 	config := sandbox.Config
 
 	if config.GetLinux().GetSecurityContext().GetNamespaceOptions().GetNetwork() == runtime.NamespaceMode_NODE {
 		// For sandboxes using the node network we are not
 		// responsible for reporting the IP.
-		return ""
+		return "", nil
 	}
 
-	// The network namespace has been closed.
-	if sandbox.NetNS == nil || sandbox.NetNS.Closed() {
-		return ""
+	if closed, err := sandbox.NetNS.Closed(); err != nil {
+		return "", errors.Wrap(err, "check network namespace closed")
+	} else if closed {
+		return "", nil
 	}
 
-	return sandbox.IP
+	return sandbox.IP, nil
 }
 
 // toCRISandboxStatus converts sandbox metadata into CRI pod sandbox status.
@@ -146,9 +150,13 @@ func toCRISandboxInfo(ctx context.Context, sandbox sandboxstore.Sandbox) (map[st
 		si.Status = "deleted"
 	}
 
-	if sandbox.NetNSPath != "" {
+	if sandbox.NetNS != nil {
 		// Add network closed information if sandbox is not using host network.
-		si.NetNSClosed = (sandbox.NetNS == nil || sandbox.NetNS.Closed())
+		closed, err := sandbox.NetNS.Closed()
+		if err != nil {
+			return nil, errors.Wrap(err, "failed to check network namespace closed")
+		}
+		si.NetNSClosed = closed
 	}
 
 	spec, err := container.Spec(ctx)

pkg/server/sandbox_stop.go

@@ -59,21 +59,20 @@ func (c *criService) StopPodSandbox(ctx context.Context, r *runtime.StopPodSandb
 	}
 
 	// Teardown network for sandbox.
-	if sandbox.NetNSPath != "" {
+	if sandbox.NetNS != nil {
 		netNSPath := sandbox.NetNSPath
-		if sandbox.NetNS == nil || sandbox.NetNS.Closed() {
-			// Use empty netns path if netns is not available. This is defined in:
-			// https://github.com/containernetworking/cni/blob/v0.7.0-alpha1/SPEC.md
+		// Use empty netns path if netns is not available. This is defined in:
+		// https://github.com/containernetworking/cni/blob/v0.7.0-alpha1/SPEC.md
+		if closed, err := sandbox.NetNS.Closed(); err != nil {
+			return nil, errors.Wrap(err, "failed to check network namespace closed")
+		} else if closed {
 			netNSPath = ""
 		}
 		if err := c.teardownPod(id, netNSPath, sandbox.Config); err != nil {
 			return nil, errors.Wrapf(err, "failed to destroy network for sandbox %q", id)
 		}
-		// Close the sandbox network namespace if it was created
-		if sandbox.NetNS != nil {
-			if err = sandbox.NetNS.Remove(); err != nil {
-				return nil, errors.Wrapf(err, "failed to remove network namespace for sandbox %q", id)
-			}
+		if err = sandbox.NetNS.Remove(); err != nil {
+			return nil, errors.Wrapf(err, "failed to remove network namespace for sandbox %q", id)
 		}
 	}
 

pkg/store/sandbox/netns.go

@@ -1,132 +0,0 @@
-/*
-Copyright 2017 The Kubernetes Authors.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-*/
-
-package sandbox
-
-import (
-	"os"
-	"sync"
-
-	cnins "github.com/containernetworking/plugins/pkg/ns"
-	"github.com/docker/docker/pkg/symlink"
-	"github.com/pkg/errors"
-
-	osinterface "github.com/containerd/cri/pkg/os"
-)
-
-// The NetNS library assumes only containerd manages the lifecycle of the
-// network namespace mount. The only case that netns will be unmounted by
-// someone else is node reboot.
-// If this assumption is broken, NetNS won't be aware of the external
-// unmount, and there will be a state mismatch.
-// TODO(random-liu): Don't cache state, always load from the system.
-
-// ErrClosedNetNS is the error returned when network namespace is closed.
-var ErrClosedNetNS = errors.New("network namespace is closed")
-
-// NetNS holds network namespace for sandbox
-type NetNS struct {
-	sync.Mutex
-	ns       cnins.NetNS
-	closed   bool
-	restored bool
-}
-
-// NewNetNS creates a network namespace for the sandbox
-func NewNetNS() (*NetNS, error) {
-	netns, err := cnins.NewNS()
-	if err != nil {
-		return nil, errors.Wrap(err, "failed to setup network namespace")
-	}
-	n := new(NetNS)
-	n.ns = netns
-	return n, nil
-}
-
-// LoadNetNS loads existing network namespace. It returns ErrClosedNetNS
-// if the network namespace has already been closed.
-func LoadNetNS(path string) (*NetNS, error) {
-	ns, err := cnins.GetNS(path)
-	if err != nil {
-		if _, ok := err.(cnins.NSPathNotExistErr); ok {
-			return nil, ErrClosedNetNS
-		}
-		if _, ok := err.(cnins.NSPathNotNSErr); ok {
-			// Do best effort cleanup.
-			os.RemoveAll(path) // nolint: errcheck
-			return nil, ErrClosedNetNS
-		}
-		return nil, errors.Wrap(err, "failed to load network namespace")
-	}
-	return &NetNS{ns: ns, restored: true}, nil
-}
-
-// Remove removes network namepace if it exists and not closed. Remove is idempotent,
-// meaning it might be invoked multiple times and provides consistent result.
-func (n *NetNS) Remove() error {
-	n.Lock()
-	defer n.Unlock()
-	if !n.closed {
-		err := n.ns.Close()
-		if err != nil {
-			return errors.Wrap(err, "failed to close network namespace")
-		}
-		n.closed = true
-	}
-	if n.restored {
-		path := n.ns.Path()
-		// Check netns existence.
-		if _, err := os.Stat(path); err != nil {
-			if os.IsNotExist(err) {
-				return nil
-			}
-			return errors.Wrap(err, "failed to stat netns")
-		}
-		path, err := symlink.FollowSymlinkInScope(path, "/")
-		if err != nil {
-			return errors.Wrap(err, "failed to follow symlink")
-		}
-		if err := osinterface.Unmount(path); err != nil && !os.IsNotExist(err) {
-			return errors.Wrap(err, "failed to umount netns")
-		}
-		if err := os.RemoveAll(path); err != nil {
-			return errors.Wrap(err, "failed to remove netns")
-		}
-		n.restored = false
-	}
-	return nil
-}
-
-// Closed checks whether the network namespace has been closed.
-func (n *NetNS) Closed() bool {
-	n.Lock()
-	defer n.Unlock()
-	return n.closed && !n.restored
-}
-
-// GetPath returns network namespace path for sandbox container
-func (n *NetNS) GetPath() string {
-	n.Lock()
-	defer n.Unlock()
-	return n.ns.Path()
-}
-
-// GetNs returns the network namespace handle
-func (n *NetNS) GetNs() cnins.NetNS {
-	n.Lock()
-	defer n.Unlock()
-	return n.ns
-}

pkg/store/sandbox/sandbox.go

@@ -22,6 +22,7 @@ import (
 	"github.com/containerd/containerd"
 	"github.com/docker/docker/pkg/truncindex"
 
+	"github.com/containerd/cri/pkg/netns"
 	"github.com/containerd/cri/pkg/store"
 )
 
@@ -32,10 +33,12 @@ type Sandbox struct {
 	Metadata
 	// Status stores the status of the sandbox.
 	Status StatusStorage
-	// Container is the containerd sandbox container client
+	// Container is the containerd sandbox container client.
 	Container containerd.Container
-	// CNI network namespace client
-	NetNS *NetNS
+	// CNI network namespace client.
+	// For hostnetwork pod, this is always nil;
+	// For non hostnetwork pod, this should never be nil.
+	NetNS *netns.NetNS
 	// StopCh is used to propagate the stop information of the sandbox.
 	*store.StopCh
 }