mount_linux.go

// +build linux

/*
Copyright 2014 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 mount

import (
	"errors"
	"fmt"
	"io/ioutil"
	"os"
	"os/exec"
	"path"
	"path/filepath"
	"strconv"
	"strings"
	"syscall"

	"github.com/golang/glog"
	"golang.org/x/sys/unix"
	"k8s.io/apimachinery/pkg/util/sets"
	utilio "k8s.io/kubernetes/pkg/util/io"
	utilexec "k8s.io/utils/exec"
)

const (
	// How many times to retry for a consistent read of /proc/mounts.
	maxListTries = 3
	// Number of fields per line in /proc/mounts as per the fstab man page.
	expectedNumFieldsPerLine = 6
	// Location of the mount file to use
	procMountsPath = "/proc/mounts"
	// Location of the mountinfo file
	procMountInfoPath = "/proc/self/mountinfo"
	// 'fsck' found errors and corrected them
	fsckErrorsCorrected = 1
	// 'fsck' found errors but exited without correcting them
	fsckErrorsUncorrected = 4

	// place for subpath mounts
	containerSubPathDirectoryName = "volume-subpaths"
	// syscall.Openat flags used to traverse directories not following symlinks
	nofollowFlags = unix.O_RDONLY | unix.O_NOFOLLOW
	// flags for getting file descriptor without following the symlink
	openFDFlags = unix.O_NOFOLLOW | unix.O_PATH
)

// Mounter provides the default implementation of mount.Interface
// for the linux platform.  This implementation assumes that the
// kubelet is running in the host's root mount namespace.
type Mounter struct {
	mounterPath string
	withSystemd bool
}

// New returns a mount.Interface for the current system.
// It provides options to override the default mounter behavior.
// mounterPath allows using an alternative to `/bin/mount` for mounting.
func New(mounterPath string) Interface {
	return &Mounter{
		mounterPath: mounterPath,
		withSystemd: detectSystemd(),
	}
}

// Mount mounts source to target as fstype with given options. 'source' and 'fstype' must
// be an empty string in case it's not required, e.g. for remount, or for auto filesystem
// type, where kernel handles fstype for you. The mount 'options' is a list of options,
// currently come from mount(8), e.g. "ro", "remount", "bind", etc. If no more option is
// required, call Mount with an empty string list or nil.
func (mounter *Mounter) Mount(source string, target string, fstype string, options []string) error {
	// Path to mounter binary if containerized mounter is needed. Otherwise, it is set to empty.
	// All Linux distros are expected to be shipped with a mount utility that a support bind mounts.
	mounterPath := ""
	bind, bindRemountOpts := isBind(options)
	if bind {
		err := mounter.doMount(mounterPath, defaultMountCommand, source, target, fstype, []string{"bind"})
		if err != nil {
			return err
		}
		return mounter.doMount(mounterPath, defaultMountCommand, source, target, fstype, bindRemountOpts)
	}
	// The list of filesystems that require containerized mounter on GCI image cluster
	fsTypesNeedMounter := sets.NewString("nfs", "glusterfs", "ceph", "cifs")
	if fsTypesNeedMounter.Has(fstype) {
		mounterPath = mounter.mounterPath
	}
	return mounter.doMount(mounterPath, defaultMountCommand, source, target, fstype, options)
}

// doMount runs the mount command. mounterPath is the path to mounter binary if containerized mounter is used.
func (m *Mounter) doMount(mounterPath string, mountCmd string, source string, target string, fstype string, options []string) error {
	mountArgs := makeMountArgs(source, target, fstype, options)
	if len(mounterPath) > 0 {
		mountArgs = append([]string{mountCmd}, mountArgs...)
		mountCmd = mounterPath
	}

	if m.withSystemd {
		// Try to run mount via systemd-run --scope. This will escape the
		// service where kubelet runs and any fuse daemons will be started in a
		// specific scope. kubelet service than can be restarted without killing
		// these fuse daemons.
		//
		// Complete command line (when mounterPath is not used):
		// systemd-run --description=... --scope -- mount -t <type> <what> <where>
		//
		// Expected flow:
		// * systemd-run creates a transient scope (=~ cgroup) and executes its
		//   argument (/bin/mount) there.
		// * mount does its job, forks a fuse daemon if necessary and finishes.
		//   (systemd-run --scope finishes at this point, returning mount's exit
		//   code and stdout/stderr - thats one of --scope benefits).
		// * systemd keeps the fuse daemon running in the scope (i.e. in its own
		//   cgroup) until the fuse daemon dies (another --scope benefit).
		//   Kubelet service can be restarted and the fuse daemon survives.
		// * When the fuse daemon dies (e.g. during unmount) systemd removes the
		//   scope automatically.
		//
		// systemd-mount is not used because it's too new for older distros
		// (CentOS 7, Debian Jessie).
		mountCmd, mountArgs = addSystemdScope("systemd-run", target, mountCmd, mountArgs)
	} else {
		// No systemd-run on the host (or we failed to check it), assume kubelet
		// does not run as a systemd service.
		// No code here, mountCmd and mountArgs are already populated.
	}

	glog.V(4).Infof("Mounting cmd (%s) with arguments (%s)", mountCmd, mountArgs)
	command := exec.Command(mountCmd, mountArgs...)
	output, err := command.CombinedOutput()
	if err != nil {
		args := strings.Join(mountArgs, " ")
		glog.Errorf("Mount failed: %v\nMounting command: %s\nMounting arguments: %s\nOutput: %s\n", err, mountCmd, args, string(output))
		return fmt.Errorf("mount failed: %v\nMounting command: %s\nMounting arguments: %s\nOutput: %s\n",
			err, mountCmd, args, string(output))
	}
	return err
}

// GetMountRefs finds all other references to the device referenced
// by mountPath; returns a list of paths.
func GetMountRefs(mounter Interface, mountPath string) ([]string, error) {
	mps, err := mounter.List()
	if err != nil {
		return nil, err
	}
	// Find the device name.
	deviceName := ""
	// If mountPath is symlink, need get its target path.
	slTarget, err := filepath.EvalSymlinks(mountPath)
	if err != nil {
		slTarget = mountPath
	}
	for i := range mps {
		if mps[i].Path == slTarget {
			deviceName = mps[i].Device
			break
		}
	}

	// Find all references to the device.
	var refs []string
	if deviceName == "" {
		glog.Warningf("could not determine device for path: %q", mountPath)
	} else {
		for i := range mps {
			if mps[i].Device == deviceName && mps[i].Path != slTarget {
				refs = append(refs, mps[i].Path)
			}
		}
	}
	return refs, nil
}

// detectSystemd returns true if OS runs with systemd as init. When not sure
// (permission errors, ...), it returns false.
// There may be different ways how to detect systemd, this one makes sure that
// systemd-runs (needed by Mount()) works.
func detectSystemd() bool {
	if _, err := exec.LookPath("systemd-run"); err != nil {
		glog.V(2).Infof("Detected OS without systemd")
		return false
	}
	// Try to run systemd-run --scope /bin/true, that should be enough
	// to make sure that systemd is really running and not just installed,
	// which happens when running in a container with a systemd-based image
	// but with different pid 1.
	cmd := exec.Command("systemd-run", "--description=Kubernetes systemd probe", "--scope", "true")
	output, err := cmd.CombinedOutput()
	if err != nil {
		glog.V(2).Infof("Cannot run systemd-run, assuming non-systemd OS")
		glog.V(4).Infof("systemd-run failed with: %v", err)
		glog.V(4).Infof("systemd-run output: %s", string(output))
		return false
	}
	glog.V(2).Infof("Detected OS with systemd")
	return true
}

// makeMountArgs makes the arguments to the mount(8) command.
func makeMountArgs(source, target, fstype string, options []string) []string {
	// Build mount command as follows:
	//   mount [-t $fstype] [-o $options] [$source] $target
	mountArgs := []string{}
	if len(fstype) > 0 {
		mountArgs = append(mountArgs, "-t", fstype)
	}
	if len(options) > 0 {
		mountArgs = append(mountArgs, "-o", strings.Join(options, ","))
	}
	if len(source) > 0 {
		mountArgs = append(mountArgs, source)
	}
	mountArgs = append(mountArgs, target)

	return mountArgs
}

// addSystemdScope adds "system-run --scope" to given command line
func addSystemdScope(systemdRunPath, mountName, command string, args []string) (string, []string) {
	descriptionArg := fmt.Sprintf("--description=Kubernetes transient mount for %s", mountName)
	systemdRunArgs := []string{descriptionArg, "--scope", "--", command}
	return systemdRunPath, append(systemdRunArgs, args...)
}

// Unmount unmounts the target.
func (mounter *Mounter) Unmount(target string) error {
	glog.V(4).Infof("Unmounting %s", target)
	command := exec.Command("umount", target)
	output, err := command.CombinedOutput()
	if err != nil {
		return fmt.Errorf("Unmount failed: %v\nUnmounting arguments: %s\nOutput: %s\n", err, target, string(output))
	}
	return nil
}

// List returns a list of all mounted filesystems.
func (*Mounter) List() ([]MountPoint, error) {
	return listProcMounts(procMountsPath)
}

func (mounter *Mounter) IsMountPointMatch(mp MountPoint, dir string) bool {
	deletedDir := fmt.Sprintf("%s\\040(deleted)", dir)
	return ((mp.Path == dir) || (mp.Path == deletedDir))
}

func (mounter *Mounter) IsNotMountPoint(dir string) (bool, error) {
	return IsNotMountPoint(mounter, dir)
}

// IsLikelyNotMountPoint determines if a directory is not a mountpoint.
// It is fast but not necessarily ALWAYS correct. If the path is in fact
// a bind mount from one part of a mount to another it will not be detected.
// mkdir /tmp/a /tmp/b; mount --bin /tmp/a /tmp/b; IsLikelyNotMountPoint("/tmp/b")
// will return true. When in fact /tmp/b is a mount point. If this situation
// if of interest to you, don't use this function...
func (mounter *Mounter) IsLikelyNotMountPoint(file string) (bool, error) {
	stat, err := os.Stat(file)
	if err != nil {
		return true, err
	}
	rootStat, err := os.Lstat(filepath.Dir(strings.TrimSuffix(file, "/")))
	if err != nil {
		return true, err
	}
	// If the directory has a different device as parent, then it is a mountpoint.
	if stat.Sys().(*syscall.Stat_t).Dev != rootStat.Sys().(*syscall.Stat_t).Dev {
		return false, nil
	}

	return true, nil
}

// DeviceOpened checks if block device in use by calling Open with O_EXCL flag.
// If pathname is not a device, log and return false with nil error.
// If open returns errno EBUSY, return true with nil error.
// If open returns nil, return false with nil error.
// Otherwise, return false with error
func (mounter *Mounter) DeviceOpened(pathname string) (bool, error) {
	return exclusiveOpenFailsOnDevice(pathname)
}

// PathIsDevice uses FileInfo returned from os.Stat to check if path refers
// to a device.
func (mounter *Mounter) PathIsDevice(pathname string) (bool, error) {
	pathType, err := mounter.GetFileType(pathname)
	isDevice := pathType == FileTypeCharDev || pathType == FileTypeBlockDev
	return isDevice, err
}

func exclusiveOpenFailsOnDevice(pathname string) (bool, error) {
	var isDevice bool
	finfo, err := os.Stat(pathname)
	if os.IsNotExist(err) {
		isDevice = false
	}
	// err in call to os.Stat
	if err != nil {
		return false, fmt.Errorf(
			"PathIsDevice failed for path %q: %v",
			pathname,
			err)
	}
	// path refers to a device
	if finfo.Mode()&os.ModeDevice != 0 {
		isDevice = true
	}

	if !isDevice {
		glog.Errorf("Path %q is not referring to a device.", pathname)
		return false, nil
	}
	fd, errno := unix.Open(pathname, unix.O_RDONLY|unix.O_EXCL, 0)
	// If the device is in use, open will return an invalid fd.
	// When this happens, it is expected that Close will fail and throw an error.
	defer unix.Close(fd)
	if errno == nil {
		// device not in use
		return false, nil
	} else if errno == unix.EBUSY {
		// device is in use
		return true, nil
	}
	// error during call to Open
	return false, errno
}

//GetDeviceNameFromMount: given a mount point, find the device name from its global mount point
func (mounter *Mounter) GetDeviceNameFromMount(mountPath, pluginDir string) (string, error) {
	return getDeviceNameFromMount(mounter, mountPath, pluginDir)
}

// getDeviceNameFromMount find the device name from /proc/mounts in which
// the mount path reference should match the given plugin directory. In case no mount path reference
// matches, returns the volume name taken from its given mountPath
func getDeviceNameFromMount(mounter Interface, mountPath, pluginDir string) (string, error) {
	refs, err := GetMountRefs(mounter, mountPath)
	if err != nil {
		glog.V(4).Infof("GetMountRefs failed for mount path %q: %v", mountPath, err)
		return "", err
	}
	if len(refs) == 0 {
		glog.V(4).Infof("Directory %s is not mounted", mountPath)
		return "", fmt.Errorf("directory %s is not mounted", mountPath)
	}
	basemountPath := path.Join(pluginDir, MountsInGlobalPDPath)
	for _, ref := range refs {
		if strings.HasPrefix(ref, basemountPath) {
			volumeID, err := filepath.Rel(basemountPath, ref)
			if err != nil {
				glog.Errorf("Failed to get volume id from mount %s - %v", mountPath, err)
				return "", err
			}
			return volumeID, nil
		}
	}

	return path.Base(mountPath), nil
}

func listProcMounts(mountFilePath string) ([]MountPoint, error) {
	content, err := utilio.ConsistentRead(mountFilePath, maxListTries)
	if err != nil {
		return nil, err
	}
	return parseProcMounts(content)
}

func parseProcMounts(content []byte) ([]MountPoint, error) {
	out := []MountPoint{}
	lines := strings.Split(string(content), "\n")
	for _, line := range lines {
		if line == "" {
			// the last split() item is empty string following the last \n
			continue
		}
		fields := strings.Fields(line)
		if len(fields) != expectedNumFieldsPerLine {
			return nil, fmt.Errorf("wrong number of fields (expected %d, got %d): %s", expectedNumFieldsPerLine, len(fields), line)
		}

		mp := MountPoint{
			Device: fields[0],
			Path:   fields[1],
			Type:   fields[2],
			Opts:   strings.Split(fields[3], ","),
		}

		freq, err := strconv.Atoi(fields[4])
		if err != nil {
			return nil, err
		}
		mp.Freq = freq

		pass, err := strconv.Atoi(fields[5])
		if err != nil {
			return nil, err
		}
		mp.Pass = pass

		out = append(out, mp)
	}
	return out, nil
}

func (mounter *Mounter) MakeRShared(path string) error {
	return doMakeRShared(path, procMountInfoPath)
}

func (mounter *Mounter) GetFileType(pathname string) (FileType, error) {
	return getFileType(pathname)
}

func (mounter *Mounter) MakeDir(pathname string) error {
	err := os.MkdirAll(pathname, os.FileMode(0755))
	if err != nil {
		if !os.IsExist(err) {
			return err
		}
	}
	return nil
}

func (mounter *Mounter) MakeFile(pathname string) error {
	f, err := os.OpenFile(pathname, os.O_CREATE, os.FileMode(0644))
	defer f.Close()
	if err != nil {
		if !os.IsExist(err) {
			return err
		}
	}
	return nil
}

func (mounter *Mounter) ExistsPath(pathname string) bool {
	_, err := os.Stat(pathname)
	if err != nil {
		return false
	}
	return true
}

// formatAndMount uses unix utils to format and mount the given disk
func (mounter *SafeFormatAndMount) formatAndMount(source string, target string, fstype string, options []string) error {
	readOnly := false
	for _, option := range options {
		if option == "ro" {
			readOnly = true
			break
		}
	}

	options = append(options, "defaults")

	if !readOnly {
		// Run fsck on the disk to fix repairable issues, only do this for volumes requested as rw.
		glog.V(4).Infof("Checking for issues with fsck on disk: %s", source)
		args := []string{"-a", source}
		out, err := mounter.Exec.Run("fsck", args...)
		if err != nil {
			ee, isExitError := err.(utilexec.ExitError)
			switch {
			case err == utilexec.ErrExecutableNotFound:
				glog.Warningf("'fsck' not found on system; continuing mount without running 'fsck'.")
			case isExitError && ee.ExitStatus() == fsckErrorsCorrected:
				glog.Infof("Device %s has errors which were corrected by fsck.", source)
			case isExitError && ee.ExitStatus() == fsckErrorsUncorrected:
				return fmt.Errorf("'fsck' found errors on device %s but could not correct them: %s.", source, string(out))
			case isExitError && ee.ExitStatus() > fsckErrorsUncorrected:
				glog.Infof("`fsck` error %s", string(out))
			}
		}
	}

	// Try to mount the disk
	glog.V(4).Infof("Attempting to mount disk: %s %s %s", fstype, source, target)
	mountErr := mounter.Interface.Mount(source, target, fstype, options)
	if mountErr != nil {
		// Mount failed. This indicates either that the disk is unformatted or
		// it contains an unexpected filesystem.
		existingFormat, err := mounter.GetDiskFormat(source)
		if err != nil {
			return err
		}
		if existingFormat == "" {
			if readOnly {
				// Don't attempt to format if mounting as readonly, return an error to reflect this.
				return errors.New("failed to mount unformatted volume as read only")
			}

			// Disk is unformatted so format it.
			args := []string{source}
			// Use 'ext4' as the default
			if len(fstype) == 0 {
				fstype = "ext4"
			}

			if fstype == "ext4" || fstype == "ext3" {
				args = []string{"-F", source}
			}
			glog.Infof("Disk %q appears to be unformatted, attempting to format as type: %q with options: %v", source, fstype, args)
			_, err := mounter.Exec.Run("mkfs."+fstype, args...)
			if err == nil {
				// the disk has been formatted successfully try to mount it again.
				glog.Infof("Disk successfully formatted (mkfs): %s - %s %s", fstype, source, target)
				return mounter.Interface.Mount(source, target, fstype, options)
			}
			glog.Errorf("format of disk %q failed: type:(%q) target:(%q) options:(%q)error:(%v)", source, fstype, target, options, err)
			return err
		} else {
			// Disk is already formatted and failed to mount
			if len(fstype) == 0 || fstype == existingFormat {
				// This is mount error
				return mountErr
			} else {
				// Block device is formatted with unexpected filesystem, let the user know
				return fmt.Errorf("failed to mount the volume as %q, it already contains %s. Mount error: %v", fstype, existingFormat, mountErr)
			}
		}
	}
	return mountErr
}

// GetDiskFormat uses 'blkid' to see if the given disk is unformated
func (mounter *SafeFormatAndMount) GetDiskFormat(disk string) (string, error) {
	args := []string{"-p", "-s", "TYPE", "-s", "PTTYPE", "-o", "export", disk}
	glog.V(4).Infof("Attempting to determine if disk %q is formatted using blkid with args: (%v)", disk, args)
	dataOut, err := mounter.Exec.Run("blkid", args...)
	output := string(dataOut)
	glog.V(4).Infof("Output: %q, err: %v", output, err)

	if err != nil {
		if exit, ok := err.(utilexec.ExitError); ok {
			if exit.ExitStatus() == 2 {
				// Disk device is unformatted.
				// For `blkid`, if the specified token (TYPE/PTTYPE, etc) was
				// not found, or no (specified) devices could be identified, an
				// exit code of 2 is returned.
				return "", nil
			}
		}
		glog.Errorf("Could not determine if disk %q is formatted (%v)", disk, err)
		return "", err
	}

	var fstype, pttype string

	lines := strings.Split(output, "\n")
	for _, l := range lines {
		if len(l) <= 0 {
			// Ignore empty line.
			continue
		}
		cs := strings.Split(l, "=")
		if len(cs) != 2 {
			return "", fmt.Errorf("blkid returns invalid output: %s", output)
		}
		// TYPE is filesystem type, and PTTYPE is partition table type, according
		// to https://www.kernel.org/pub/linux/utils/util-linux/v2.21/libblkid-docs/.
		if cs[0] == "TYPE" {
			fstype = cs[1]
		} else if cs[0] == "PTTYPE" {
			pttype = cs[1]
		}
	}

	if len(pttype) > 0 {
		glog.V(4).Infof("Disk %s detected partition table type: %s", pttype)
		// Returns a special non-empty string as filesystem type, then kubelet
		// will not format it.
		return "unknown data, probably partitions", nil
	}

	return fstype, nil
}

// isShared returns true, if given path is on a mount point that has shared
// mount propagation.
func isShared(path string, filename string) (bool, error) {
	infos, err := parseMountInfo(filename)
	if err != nil {
		return false, err
	}

	// process /proc/xxx/mountinfo in backward order and find the first mount
	// point that is prefix of 'path' - that's the mount where path resides
	var info *mountInfo
	for i := len(infos) - 1; i >= 0; i-- {
		if strings.HasPrefix(path, infos[i].mountPoint) {
			info = &infos[i]
			break
		}
	}
	if info == nil {
		return false, fmt.Errorf("cannot find mount point for %q", path)
	}

	// parse optional parameters
	for _, opt := range info.optional {
		if strings.HasPrefix(opt, "shared:") {
			return true, nil
		}
	}
	return false, nil
}

type mountInfo struct {
	// Path of the mount point
	mountPoint string
	// list of "optional parameters", mount propagation is one of them
	optional []string
}

// parseMountInfo parses /proc/xxx/mountinfo.
func parseMountInfo(filename string) ([]mountInfo, error) {
	content, err := utilio.ConsistentRead(filename, maxListTries)
	if err != nil {
		return []mountInfo{}, err
	}
	contentStr := string(content)
	infos := []mountInfo{}

	for _, line := range strings.Split(contentStr, "\n") {
		if line == "" {
			// the last split() item is empty string following the last \n
			continue
		}
		// See `man proc` for authoritative description of format of the file.
		fields := strings.Fields(line)
		if len(fields) < 7 {
			return nil, fmt.Errorf("wrong number of fields in (expected %d, got %d): %s", 8, len(fields), line)
		}
		info := mountInfo{
			mountPoint: fields[4],
			optional:   []string{},
		}
		// All fields until "-" are "optional fields".
		for i := 6; i < len(fields) && fields[i] != "-"; i++ {
			info.optional = append(info.optional, fields[i])
		}
		infos = append(infos, info)
	}
	return infos, nil
}

// doMakeRShared is common implementation of MakeRShared on Linux. It checks if
// path is shared and bind-mounts it as rshared if needed. mountCmd and
// mountArgs are expected to contain mount-like command, doMakeRShared will add
// '--bind <path> <path>' and '--make-rshared <path>' to mountArgs.
func doMakeRShared(path string, mountInfoFilename string) error {
	shared, err := isShared(path, mountInfoFilename)
	if err != nil {
		return err
	}
	if shared {
		glog.V(4).Infof("Directory %s is already on a shared mount", path)
		return nil
	}

	glog.V(2).Infof("Bind-mounting %q with shared mount propagation", path)
	// mount --bind /var/lib/kubelet /var/lib/kubelet
	if err := syscall.Mount(path, path, "" /*fstype*/, syscall.MS_BIND, "" /*data*/); err != nil {
		return fmt.Errorf("failed to bind-mount %s: %v", path, err)
	}

	// mount --make-rshared /var/lib/kubelet
	if err := syscall.Mount(path, path, "" /*fstype*/, syscall.MS_SHARED|syscall.MS_REC, "" /*data*/); err != nil {
		return fmt.Errorf("failed to make %s rshared: %v", path, err)
	}

	return nil
}

func (mounter *Mounter) PrepareSafeSubpath(subPath Subpath) (newHostPath string, cleanupAction func(), err error) {
	newHostPath, err = doBindSubPath(mounter, subPath, os.Getpid())
	// There is no action when the container starts. Bind-mount will be cleaned
	// when container stops by CleanSubPaths.
	cleanupAction = nil
	return newHostPath, cleanupAction, err
}

// This implementation is shared between Linux and NsEnterMounter
// kubeletPid is PID of kubelet in the PID namespace where bind-mount is done,
// i.e. pid on the *host* if kubelet runs in a container.
func doBindSubPath(mounter Interface, subpath Subpath, kubeletPid int) (hostPath string, err error) {
	// Check early for symlink. This is just a pre-check to avoid bind-mount
	// before the final check.
	evalSubPath, err := filepath.EvalSymlinks(subpath.Path)
	if err != nil {
		return "", fmt.Errorf("evalSymlinks %q failed: %v", subpath.Path, err)
	}
	glog.V(5).Infof("doBindSubPath %q, full subpath %q for volumepath %q", subpath.Path, evalSubPath, subpath.VolumePath)

	evalSubPath = filepath.Clean(evalSubPath)
	if !pathWithinBase(evalSubPath, subpath.VolumePath) {
		return "", fmt.Errorf("subpath %q not within volume path %q", evalSubPath, subpath.VolumePath)
	}

	// Prepare directory for bind mounts
	// containerName is DNS label, i.e. safe as a directory name.
	bindDir := filepath.Join(subpath.PodDir, containerSubPathDirectoryName, subpath.VolumeName, subpath.ContainerName)
	err = os.MkdirAll(bindDir, 0750)
	if err != nil && !os.IsExist(err) {
		return "", fmt.Errorf("error creating directory %s: %s", bindDir, err)
	}
	bindPathTarget := filepath.Join(bindDir, strconv.Itoa(subpath.VolumeMountIndex))

	success := false
	defer func() {
		// Cleanup subpath on error
		if !success {
			glog.V(4).Infof("doBindSubPath() failed for %q, cleaning up subpath", bindPathTarget)
			if cleanErr := cleanSubPath(mounter, subpath); cleanErr != nil {
				glog.Errorf("Failed to clean subpath %q: %v", bindPathTarget, cleanErr)
			}
		}
	}()

	// Check it's not already bind-mounted
	notMount, err := IsNotMountPoint(mounter, bindPathTarget)
	if err != nil {
		if !os.IsNotExist(err) {
			return "", fmt.Errorf("error checking path %s for mount: %s", bindPathTarget, err)
		}
		// Ignore ErrorNotExist: the file/directory will be created below if it does not exist yet.
		notMount = true
	}
	if !notMount {
		// It's already mounted
		glog.V(5).Infof("Skipping bind-mounting subpath %s: already mounted", bindPathTarget)
		success = true
		return bindPathTarget, nil
	}

	// Create target of the bind mount. A directory for directories, empty file
	// for everything else.
	t, err := os.Lstat(subpath.Path)
	if err != nil {
		return "", fmt.Errorf("lstat %s failed: %s", subpath.Path, err)
	}
	if t.Mode()&os.ModeDir > 0 {
		if err = os.Mkdir(bindPathTarget, 0750); err != nil && !os.IsExist(err) {
			return "", fmt.Errorf("error creating directory %s: %s", bindPathTarget, err)
		}
	} else {
		// "/bin/touch <bindDir>".
		// A file is enough for all possible targets (symlink, device, pipe,
		// socket, ...), bind-mounting them into a file correctly changes type
		// of the target file.
		if err = ioutil.WriteFile(bindPathTarget, []byte{}, 0640); err != nil {
			return "", fmt.Errorf("error creating file %s: %s", bindPathTarget, err)
		}
	}

	// Safe open subpath and get the fd
	fd, err := doSafeOpen(evalSubPath, subpath.VolumePath)
	if err != nil {
		return "", fmt.Errorf("error opening subpath %v: %v", evalSubPath, err)
	}
	defer syscall.Close(fd)

	mountSource := fmt.Sprintf("/proc/%d/fd/%v", kubeletPid, fd)

	// Do the bind mount
	options := []string{"bind"}
	if subpath.ReadOnly {
		options = append(options, "ro")
	}

	glog.V(5).Infof("bind mounting %q at %q", mountSource, bindPathTarget)
	if err = mounter.Mount(mountSource, bindPathTarget, "" /*fstype*/, options); err != nil {
		return "", fmt.Errorf("error mounting %s: %s", subpath.Path, err)
	}

	success = true
	glog.V(3).Infof("Bound SubPath %s into %s", subpath.Path, bindPathTarget)
	return bindPathTarget, nil
}

func (mounter *Mounter) CleanSubPaths(podDir string, volumeName string) error {
	return doCleanSubPaths(mounter, podDir, volumeName)
}

// This implementation is shared between Linux and NsEnterMounter
func doCleanSubPaths(mounter Interface, podDir string, volumeName string) error {
	// scan /var/lib/kubelet/pods/<uid>/volume-subpaths/<volume>/*
	subPathDir := filepath.Join(podDir, containerSubPathDirectoryName, volumeName)
	glog.V(4).Infof("Cleaning up subpath mounts for %s", subPathDir)

	containerDirs, err := ioutil.ReadDir(subPathDir)
	if err != nil {
		if os.IsNotExist(err) {
			return nil
		}
		return fmt.Errorf("error reading %s: %s", subPathDir, err)
	}

	for _, containerDir := range containerDirs {
		if !containerDir.IsDir() {
			glog.V(4).Infof("Container file is not a directory: %s", containerDir.Name())
			continue
		}
		glog.V(4).Infof("Cleaning up subpath mounts for container %s", containerDir.Name())

		// scan /var/lib/kubelet/pods/<uid>/volume-subpaths/<volume>/<container name>/*
		fullContainerDirPath := filepath.Join(subPathDir, containerDir.Name())
		subPaths, err := ioutil.ReadDir(fullContainerDirPath)
		if err != nil {
			return fmt.Errorf("error reading %s: %s", fullContainerDirPath, err)
		}
		for _, subPath := range subPaths {
			if err = doCleanSubPath(mounter, fullContainerDirPath, subPath.Name()); err != nil {
				return err
			}
		}
		// Whole container has been processed, remove its directory.
		if err := os.Remove(fullContainerDirPath); err != nil {
			return fmt.Errorf("error deleting %s: %s", fullContainerDirPath, err)
		}
		glog.V(5).Infof("Removed %s", fullContainerDirPath)
	}
	// Whole pod volume subpaths have been cleaned up, remove its subpath directory.
	if err := os.Remove(subPathDir); err != nil {
		return fmt.Errorf("error deleting %s: %s", subPathDir, err)
	}
	glog.V(5).Infof("Removed %s", subPathDir)

	// Remove entire subpath directory if it's the last one
	podSubPathDir := filepath.Join(podDir, containerSubPathDirectoryName)
	if err := os.Remove(podSubPathDir); err != nil && !os.IsExist(err) {
		return fmt.Errorf("error deleting %s: %s", podSubPathDir, err)
	}
	glog.V(5).Infof("Removed %s", podSubPathDir)
	return nil
}

// doCleanSubPath tears down the single subpath bind mount
func doCleanSubPath(mounter Interface, fullContainerDirPath, subPathIndex string) error {
	// process /var/lib/kubelet/pods/<uid>/volume-subpaths/<volume>/<container name>/<subPathName>
	glog.V(4).Infof("Cleaning up subpath mounts for subpath %v", subPathIndex)
	fullSubPath := filepath.Join(fullContainerDirPath, subPathIndex)
	notMnt, err := IsNotMountPoint(mounter, fullSubPath)
	if err != nil {
		return fmt.Errorf("error checking %s for mount: %s", fullSubPath, err)
	}
	// Unmount it
	if !notMnt {
		if err = mounter.Unmount(fullSubPath); err != nil {
			return fmt.Errorf("error unmounting %s: %s", fullSubPath, err)
		}
		glog.V(5).Infof("Unmounted %s", fullSubPath)
	}
	// Remove it *non*-recursively, just in case there were some hiccups.
	if err = os.Remove(fullSubPath); err != nil {
		return fmt.Errorf("error deleting %s: %s", fullSubPath, err)
	}
	glog.V(5).Infof("Removed %s", fullSubPath)
	return nil
}

// cleanSubPath will teardown the subpath bind mount and any remove any directories if empty
func cleanSubPath(mounter Interface, subpath Subpath) error {
	containerDir := filepath.Join(subpath.PodDir, containerSubPathDirectoryName, subpath.VolumeName, subpath.ContainerName)

	// Clean subdir bindmount
	if err := doCleanSubPath(mounter, containerDir, strconv.Itoa(subpath.VolumeMountIndex)); err != nil && !os.IsNotExist(err) {
		return err
	}

	// Recusively remove directories if empty
	if err := removeEmptyDirs(subpath.PodDir, containerDir); err != nil {
		return err
	}

	return nil
}

// removeEmptyDirs works backwards from endDir to baseDir and removes each directory
// if it is empty.  It stops once it encounters a directory that has content
func removeEmptyDirs(baseDir, endDir string) error {
	if !pathWithinBase(endDir, baseDir) {
		return fmt.Errorf("endDir %q is not within baseDir %q", endDir, baseDir)
	}

	for curDir := endDir; curDir != baseDir; curDir = filepath.Dir(curDir) {
		s, err := os.Stat(curDir)
		if err != nil {
			if os.IsNotExist(err) {
				glog.V(5).Infof("curDir %q doesn't exist, skipping", curDir)
				continue
			}
			return fmt.Errorf("error stat %q: %v", curDir, err)
		}
		if !s.IsDir() {
			return fmt.Errorf("path %q not a directory", curDir)
		}

		err = os.Remove(curDir)
		if os.IsExist(err) {
			glog.V(5).Infof("Directory %q not empty, not removing", curDir)
			break
		} else if err != nil {
			return fmt.Errorf("error removing directory %q: %v", curDir, err)
		}
		glog.V(5).Infof("Removed directory %q", curDir)
	}
	return nil
}

func (mounter *Mounter) SafeMakeDir(pathname string, base string, perm os.FileMode) error {
	return doSafeMakeDir(pathname, base, perm)
}

// This implementation is shared between Linux and NsEnterMounter
func doSafeMakeDir(pathname string, base string, perm os.FileMode) error {
	glog.V(4).Infof("Creating directory %q within base %q", pathname, base)

	if !pathWithinBase(pathname, base) {
		return fmt.Errorf("path %s is outside of allowed base %s", pathname, base)
	}

	// Quick check if the directory already exists
	s, err := os.Stat(pathname)
	if err == nil {
		// Path exists
		if s.IsDir() {
			// The directory already exists. It can be outside of the parent,
			// but there is no race-proof check.
			glog.V(4).Infof("Directory %s already exists", pathname)
			return nil
		}
		return &os.PathError{Op: "mkdir", Path: pathname, Err: syscall.ENOTDIR}
	}

	// Find all existing directories
	existingPath, toCreate, err := findExistingPrefix(base, pathname)
	if err != nil {
		return fmt.Errorf("error opening directory %s: %s", pathname, err)
	}
	// Ensure the existing directory is inside allowed base
	fullExistingPath, err := filepath.EvalSymlinks(existingPath)
	if err != nil {
		return fmt.Errorf("error opening directory %s: %s", existingPath, err)
	}
	if !pathWithinBase(fullExistingPath, base) {
		return fmt.Errorf("path %s is outside of allowed base %s", fullExistingPath, err)
	}

	glog.V(4).Infof("%q already exists, %q to create", fullExistingPath, filepath.Join(toCreate...))
	parentFD, err := doSafeOpen(fullExistingPath, base)
	if err != nil {
		return fmt.Errorf("cannot open directory %s: %s", existingPath, err)
	}
	childFD := -1
	defer func() {
		if parentFD != -1 {
			if err = syscall.Close(parentFD); err != nil {
				glog.V(4).Infof("Closing FD %v failed for safemkdir(%v): %v", parentFD, pathname, err)
			}
		}
		if childFD != -1 {
			if err = syscall.Close(childFD); err != nil {
				glog.V(4).Infof("Closing FD %v failed for safemkdir(%v): %v", childFD, pathname, err)
			}
		}
	}()

	currentPath := fullExistingPath
	// create the directories one by one, making sure nobody can change
	// created directory into symlink.
	for _, dir := range toCreate {
		currentPath = filepath.Join(currentPath, dir)
		glog.V(4).Infof("Creating %s", dir)
		err = syscall.Mkdirat(parentFD, currentPath, uint32(perm))
		if err != nil {
			return fmt.Errorf("cannot create directory %s: %s", currentPath, err)
		}
		// Dive into the created directory
		childFD, err := syscall.Openat(parentFD, dir, nofollowFlags, 0)
		if err != nil {
			return fmt.Errorf("cannot open %s: %s", currentPath, err)
		}
		// We can be sure that childFD is safe to use. It could be changed
		// by user after Mkdirat() and before Openat(), however:
		// - it could not be changed to symlink - we use nofollowFlags
		// - it could be changed to a file (or device, pipe, socket, ...)
		//   but either subsequent Mkdirat() fails or we mount this file
		//   to user's container. Security is no violated in both cases
		//   and user either gets error or the file that it can already access.

		if err = syscall.Close(parentFD); err != nil {
			glog.V(4).Infof("Closing FD %v failed for safemkdir(%v): %v", parentFD, pathname, err)
		}
		parentFD = childFD
		childFD = -1
	}

	// Everything was created. mkdirat(..., perm) above was affected by current
	// umask and we must apply the right permissions to the last directory
	// (that's the one that will be available to the container as subpath)
	// so user can read/write it. This is the behavior of previous code.
	// TODO: chmod all created directories, not just the last one.
	// parentFD is the last created directory.

	// Translate perm (os.FileMode) to uint32 that fchmod() expects
	kernelPerm := uint32(perm & os.ModePerm)
	if perm&os.ModeSetgid > 0 {
		kernelPerm |= syscall.S_ISGID
	}
	if perm&os.ModeSetuid > 0 {
		kernelPerm |= syscall.S_ISUID
	}
	if perm&os.ModeSticky > 0 {
		kernelPerm |= syscall.S_ISVTX
	}
	if err = syscall.Fchmod(parentFD, kernelPerm); err != nil {
		return fmt.Errorf("chmod %q failed: %s", currentPath, err)
	}
	return nil
}

// findExistingPrefix finds prefix of pathname that exists. In addition, it
// returns list of remaining directories that don't exist yet.
func findExistingPrefix(base, pathname string) (string, []string, error) {
	rel, err := filepath.Rel(base, pathname)
	if err != nil {
		return base, nil, err
	}
	dirs := strings.Split(rel, string(filepath.Separator))

	// Do OpenAt in a loop to find the first non-existing dir. Resolve symlinks.
	// This should be faster than looping through all dirs and calling os.Stat()
	// on each of them, as the symlinks are resolved only once with OpenAt().
	currentPath := base
	fd, err := syscall.Open(currentPath, syscall.O_RDONLY, 0)
	if err != nil {
		return pathname, nil, fmt.Errorf("error opening %s: %s", currentPath, err)
	}
	defer func() {
		if err = syscall.Close(fd); err != nil {
			glog.V(4).Infof("Closing FD %v failed for findExistingPrefix(%v): %v", fd, pathname, err)
		}
	}()
	for i, dir := range dirs {
		// Using O_PATH here will prevent hangs in case user replaces directory with
		// fifo
		childFD, err := syscall.Openat(fd, dir, unix.O_PATH, 0)
		if err != nil {
			if os.IsNotExist(err) {
				return currentPath, dirs[i:], nil
			}
			return base, nil, err
		}
		if err = syscall.Close(fd); err != nil {
			glog.V(4).Infof("Closing FD %v failed for findExistingPrefix(%v): %v", fd, pathname, err)
		}
		fd = childFD
		currentPath = filepath.Join(currentPath, dir)
	}
	return pathname, []string{}, nil
}

// This implementation is shared between Linux and NsEnterMounter
// Open path and return its fd.
// Symlinks are disallowed (pathname must already resolve symlinks),
// and the path must be within the base directory.
func doSafeOpen(pathname string, base string) (int, error) {
	// Calculate segments to follow
	subpath, err := filepath.Rel(base, pathname)
	if err != nil {
		return -1, err
	}
	segments := strings.Split(subpath, string(filepath.Separator))

	// Assumption: base is the only directory that we have under control.
	// Base dir is not allowed to be a symlink.
	parentFD, err := syscall.Open(base, nofollowFlags, 0)
	if err != nil {
		return -1, fmt.Errorf("cannot open directory %s: %s", base, err)
	}
	defer func() {
		if parentFD != -1 {
			if err = syscall.Close(parentFD); err != nil {
				glog.V(4).Infof("Closing FD %v failed for safeopen(%v): %v", parentFD, pathname, err)
			}
		}
	}()

	childFD := -1
	defer func() {
		if childFD != -1 {
			if err = syscall.Close(childFD); err != nil {
				glog.V(4).Infof("Closing FD %v failed for safeopen(%v): %v", childFD, pathname, err)
			}
		}
	}()

	currentPath := base

	// Follow the segments one by one using openat() to make
	// sure the user cannot change already existing directories into symlinks.
	for _, seg := range segments {
		currentPath = filepath.Join(currentPath, seg)
		if !pathWithinBase(currentPath, base) {
			return -1, fmt.Errorf("path %s is outside of allowed base %s", currentPath, base)
		}

		glog.V(5).Infof("Opening path %s", currentPath)
		childFD, err = syscall.Openat(parentFD, seg, openFDFlags, 0)
		if err != nil {
			return -1, fmt.Errorf("cannot open %s: %s", currentPath, err)
		}

		var deviceStat unix.Stat_t
		err := unix.Fstat(childFD, &deviceStat)
		if err != nil {
			return -1, fmt.Errorf("Error running fstat on %s with %v", currentPath, err)
		}
		fileFmt := deviceStat.Mode & syscall.S_IFMT
		if fileFmt == syscall.S_IFLNK {
			return -1, fmt.Errorf("Unexpected symlink found %s", currentPath)
		}

		// Close parentFD
		if err = syscall.Close(parentFD); err != nil {
			return -1, fmt.Errorf("closing fd for %q failed: %v", filepath.Dir(currentPath), err)
		}
		// Set child to new parent
		parentFD = childFD
		childFD = -1
	}

	// We made it to the end, return this fd, don't close it
	finalFD := parentFD
	parentFD = -1

	return finalFD, nil
}