moby--moby/libnetwork/osl/namespace_linux.go

package osl

import (
	"fmt"
	"net"
	"os"
	"os/exec"
	"runtime"
	"sync"
	"syscall"
	"time"

	log "github.com/Sirupsen/logrus"
	"github.com/docker/docker/pkg/reexec"
	"github.com/docker/libnetwork/types"
	"github.com/vishvananda/netlink"
	"github.com/vishvananda/netns"
)

const prefix = "/var/run/docker/netns"

var (
	once             sync.Once
	garbagePathMap   = make(map[string]bool)
	gpmLock          sync.Mutex
	gpmWg            sync.WaitGroup
	gpmCleanupPeriod = 60 * time.Second
	gpmChan          = make(chan chan struct{})
	nsOnce           sync.Once
	initNs           netns.NsHandle
)

// The networkNamespace type is the linux implementation of the Sandbox
// interface. It represents a linux network namespace, and moves an interface
// into it when called on method AddInterface or sets the gateway etc.
type networkNamespace struct {
	path         string
	iFaces       []*nwIface
	gw           net.IP
	gwv6         net.IP
	staticRoutes []*types.StaticRoute
	neighbors    []*neigh
	nextIfIndex  int
	sync.Mutex
}

func init() {
	reexec.Register("netns-create", reexecCreateNamespace)
}

func createBasePath() {
	err := os.MkdirAll(prefix, 0644)
	if err != nil {
		panic("Could not create net namespace path directory")
	}

	// Start the garbage collection go routine
	go removeUnusedPaths()
}

func removeUnusedPaths() {
	gpmLock.Lock()
	period := gpmCleanupPeriod
	gpmLock.Unlock()

	ticker := time.NewTicker(period)
	for {
		var (
			gc   chan struct{}
			gcOk bool
		)

		select {
		case <-ticker.C:
		case gc, gcOk = <-gpmChan:
		}

		gpmLock.Lock()
		pathList := make([]string, 0, len(garbagePathMap))
		for path := range garbagePathMap {
			pathList = append(pathList, path)
		}
		garbagePathMap = make(map[string]bool)
		gpmWg.Add(1)
		gpmLock.Unlock()

		for _, path := range pathList {
			os.Remove(path)
		}

		gpmWg.Done()
		if gcOk {
			close(gc)
		}
	}
}

func addToGarbagePaths(path string) {
	gpmLock.Lock()
	garbagePathMap[path] = true
	gpmLock.Unlock()
}

func removeFromGarbagePaths(path string) {
	gpmLock.Lock()
	delete(garbagePathMap, path)
	gpmLock.Unlock()
}

// GC triggers garbage collection of namespace path right away
// and waits for it.
func GC() {
	gpmLock.Lock()
	if len(garbagePathMap) == 0 {
		// No need for GC if map is empty
		gpmLock.Unlock()
		return
	}
	gpmLock.Unlock()

	// if content exists in the garbage paths
	// we can trigger GC to run, providing a
	// channel to be notified on completion
	waitGC := make(chan struct{})
	gpmChan <- waitGC
	// wait for GC completion
	<-waitGC
}

// GenerateKey generates a sandbox key based on the passed
// container id.
func GenerateKey(containerID string) string {
	maxLen := 12
	if len(containerID) < maxLen {
		maxLen = len(containerID)
	}

	return prefix + "/" + containerID[:maxLen]
}

// NewSandbox provides a new sandbox instance created in an os specific way
// provided a key which uniquely identifies the sandbox
func NewSandbox(key string, osCreate bool) (Sandbox, error) {
	err := createNetworkNamespace(key, osCreate)
	if err != nil {
		return nil, err
	}

	return &networkNamespace{path: key}, nil
}

func (n *networkNamespace) InterfaceOptions() IfaceOptionSetter {
	return n
}

func (n *networkNamespace) NeighborOptions() NeighborOptionSetter {
	return n
}

func reexecCreateNamespace() {
	if len(os.Args) < 2 {
		log.Fatal("no namespace path provided")
	}

	if err := syscall.Mount("/proc/self/ns/net", os.Args[1], "bind", syscall.MS_BIND, ""); err != nil {
		log.Fatal(err)
	}

	if err := loopbackUp(); err != nil {
		log.Fatal(err)
	}
}

func createNetworkNamespace(path string, osCreate bool) error {
	runtime.LockOSThread()
	defer runtime.UnlockOSThread()

	origns, err := netns.Get()
	if err != nil {
		return err
	}
	defer origns.Close()

	if err := createNamespaceFile(path); err != nil {
		return err
	}

	cmd := &exec.Cmd{
		Path:   reexec.Self(),
		Args:   append([]string{"netns-create"}, path),
		Stdout: os.Stdout,
		Stderr: os.Stderr,
	}
	if osCreate {
		cmd.SysProcAttr = &syscall.SysProcAttr{}
		cmd.SysProcAttr.Cloneflags = syscall.CLONE_NEWNET
	}
	if err := cmd.Run(); err != nil {
		return fmt.Errorf("namespace creation reexec command failed: %v", err)
	}

	return nil
}

func unmountNamespaceFile(path string) {
	if _, err := os.Stat(path); err == nil {
		syscall.Unmount(path, syscall.MNT_DETACH)
	}
}

func createNamespaceFile(path string) (err error) {
	var f *os.File

	once.Do(createBasePath)
	// Remove it from garbage collection list if present
	removeFromGarbagePaths(path)

	// If the path is there unmount it first
	unmountNamespaceFile(path)

	// wait for garbage collection to complete if it is in progress
	// before trying to create the file.
	gpmWg.Wait()

	if f, err = os.Create(path); err == nil {
		f.Close()
	}

	return err
}

func loopbackUp() error {
	iface, err := netlink.LinkByName("lo")
	if err != nil {
		return err
	}
	return netlink.LinkSetUp(iface)
}

func (n *networkNamespace) InvokeFunc(f func()) error {
	return nsInvoke(n.nsPath(), func(nsFD int) error { return nil }, func(callerFD int) error {
		f()
		return nil
	})
}

func getLink() (string, error) {
	return os.Readlink(fmt.Sprintf("/proc/%d/task/%d/ns/net", os.Getpid(), syscall.Gettid()))
}

func nsInit() {
	var err error

	if initNs, err = netns.Get(); err != nil {
		log.Errorf("could not get initial namespace: %v", err)
	}
}

// InitOSContext initializes OS context while configuring network resources
func InitOSContext() func() {
	runtime.LockOSThread()
	nsOnce.Do(nsInit)
	if err := netns.Set(initNs); err != nil {
		linkInfo, linkErr := getLink()
		if linkErr != nil {
			linkInfo = linkErr.Error()
		}

		log.Errorf("failed to set to initial namespace, %v, initns fd %d: %v",
			linkInfo, initNs, err)
	}

	return runtime.UnlockOSThread
}

func nsInvoke(path string, prefunc func(nsFD int) error, postfunc func(callerFD int) error) error {
	defer InitOSContext()()

	f, err := os.OpenFile(path, os.O_RDONLY, 0)
	if err != nil {
		return fmt.Errorf("failed get network namespace %q: %v", path, err)
	}
	defer f.Close()

	nsFD := f.Fd()

	// Invoked before the namespace switch happens but after the namespace file
	// handle is obtained.
	if err := prefunc(int(nsFD)); err != nil {
		return fmt.Errorf("failed in prefunc: %v", err)
	}

	if err = netns.Set(netns.NsHandle(nsFD)); err != nil {
		return err
	}
	defer netns.Set(initNs)

	// Invoked after the namespace switch.
	return postfunc(int(initNs))
}

func (n *networkNamespace) nsPath() string {
	n.Lock()
	defer n.Unlock()

	return n.path
}

func (n *networkNamespace) Info() Info {
	return n
}

func (n *networkNamespace) Key() string {
	return n.path
}

func (n *networkNamespace) Destroy() error {
	// Assuming no running process is executing in this network namespace,
	// unmounting is sufficient to destroy it.
	if err := syscall.Unmount(n.path, syscall.MNT_DETACH); err != nil {
		return err
	}

	// Stash it into the garbage collection list
	addToGarbagePaths(n.path)
	return nil
}