gvisor/gofer: Use pivot_root instead of chroot
PiperOrigin-RevId: 231864273 Change-Id: I8545b72b615f5c2945df374b801b80be64ec3e13
This commit is contained in:
parent
88b4ce8cac
commit
4e695adcd0
|
@ -36,6 +36,29 @@ func mountInChroot(chroot, src, dst, typ string, flags uint32) error {
|
|||
return nil
|
||||
}
|
||||
|
||||
func pivotRoot(root string) error {
|
||||
if err := os.Chdir(root); err != nil {
|
||||
return fmt.Errorf("error changing working directory: %v", err)
|
||||
}
|
||||
// pivot_root(new_root, put_old) moves the root filesystem (old_root)
|
||||
// of the calling process to the directory put_old and makes new_root
|
||||
// the new root filesystem of the calling process.
|
||||
//
|
||||
// pivot_root(".", ".") makes a mount of the working directory the new
|
||||
// root filesystem, so it will be moved in "/" and then the old_root
|
||||
// will be moved to "/" too. The parent mount of the old_root will be
|
||||
// new_root, so after umounting the old_root, we will see only
|
||||
// the new_root in "/".
|
||||
if err := syscall.PivotRoot(".", "."); err != nil {
|
||||
return fmt.Errorf("error changing root filesystem: %v", err)
|
||||
}
|
||||
|
||||
if err := syscall.Unmount(".", syscall.MNT_DETACH); err != nil {
|
||||
return fmt.Errorf("error umounting the old root file system: %v", err)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
// setUpChroot creates an empty directory with runsc mounted at /runsc and proc
|
||||
// mounted at /proc.
|
||||
func setUpChroot(pidns bool) error {
|
||||
|
@ -66,29 +89,9 @@ func setUpChroot(pidns bool) error {
|
|||
}
|
||||
}
|
||||
|
||||
if err := os.Chdir(chroot); err != nil {
|
||||
return fmt.Errorf("error changing working directory: %v", err)
|
||||
}
|
||||
|
||||
if err := syscall.Mount("", chroot, "", syscall.MS_REMOUNT|syscall.MS_RDONLY|syscall.MS_BIND, ""); err != nil {
|
||||
return fmt.Errorf("error remounting chroot in read-only: %v", err)
|
||||
}
|
||||
// pivot_root(new_root, put_old) moves the root filesystem (old_root)
|
||||
// of the calling process to the directory put_old and makes new_root
|
||||
// the new root filesystem of the calling process.
|
||||
//
|
||||
// pivot_root(".", ".") makes a mount of the working directory the new
|
||||
// root filesystem, so it will be moved in "/" and then the old_root
|
||||
// will be moved to "/" too. The parent mount of the old_root will be
|
||||
// new_root, so after umounting the old_root, we will see only
|
||||
// the new_root in "/".
|
||||
if err := syscall.PivotRoot(".", "."); err != nil {
|
||||
return fmt.Errorf("error changing root filesystem: %v", err)
|
||||
}
|
||||
|
||||
if err := syscall.Unmount(".", syscall.MNT_DETACH); err != nil {
|
||||
return fmt.Errorf("error umounting the old root file system: %v", err)
|
||||
}
|
||||
|
||||
return nil
|
||||
return pivotRoot(chroot)
|
||||
}
|
||||
|
|
|
@ -26,6 +26,7 @@ import (
|
|||
"gvisor.googlesource.com/gvisor/pkg/log"
|
||||
"gvisor.googlesource.com/gvisor/pkg/p9"
|
||||
"gvisor.googlesource.com/gvisor/pkg/unet"
|
||||
"gvisor.googlesource.com/gvisor/runsc/boot"
|
||||
"gvisor.googlesource.com/gvisor/runsc/fsgofer"
|
||||
"gvisor.googlesource.com/gvisor/runsc/fsgofer/filter"
|
||||
"gvisor.googlesource.com/gvisor/runsc/specutils"
|
||||
|
@ -54,8 +55,10 @@ type Gofer struct {
|
|||
bundleDir string
|
||||
ioFDs intFlags
|
||||
applyCaps bool
|
||||
setUpRoot bool
|
||||
|
||||
panicOnWrite bool
|
||||
specFD int
|
||||
}
|
||||
|
||||
// Name implements subcommands.Command.
|
||||
|
@ -79,43 +82,83 @@ func (g *Gofer) SetFlags(f *flag.FlagSet) {
|
|||
f.Var(&g.ioFDs, "io-fds", "list of FDs to connect 9P servers. They must follow this order: root first, then mounts as defined in the spec")
|
||||
f.BoolVar(&g.applyCaps, "apply-caps", true, "if true, apply capabilities to restrict what the Gofer process can do")
|
||||
f.BoolVar(&g.panicOnWrite, "panic-on-write", false, "if true, panics on attempts to write to RO mounts. RW mounts are unnaffected")
|
||||
f.BoolVar(&g.setUpRoot, "setup-root", true, "if true, set up an empty root for the process")
|
||||
f.IntVar(&g.specFD, "spec-fd", -1, "required fd with the container spec")
|
||||
}
|
||||
|
||||
// Execute implements subcommands.Command.
|
||||
func (g *Gofer) Execute(_ context.Context, f *flag.FlagSet, args ...interface{}) subcommands.ExitStatus {
|
||||
if g.bundleDir == "" || len(g.ioFDs) < 1 {
|
||||
if g.bundleDir == "" || len(g.ioFDs) < 1 || g.specFD < 0 {
|
||||
f.Usage()
|
||||
return subcommands.ExitUsageError
|
||||
}
|
||||
|
||||
specFile := os.NewFile(uintptr(g.specFD), "spec file")
|
||||
defer specFile.Close()
|
||||
spec, err := specutils.ReadSpecFromFile(g.bundleDir, specFile)
|
||||
if err != nil {
|
||||
Fatalf("reading spec: %v", err)
|
||||
}
|
||||
|
||||
// Find what path is going to be served by this gofer.
|
||||
root := spec.Root.Path
|
||||
|
||||
conf := args[0].(*boot.Config)
|
||||
|
||||
if g.setUpRoot && !conf.TestOnlyAllowRunAsCurrentUserWithoutChroot {
|
||||
// Convert all shared mounts into slave to be sure that nothing will be
|
||||
// propagated outside of our namespace.
|
||||
if err := syscall.Mount("", "/", "", syscall.MS_SLAVE|syscall.MS_REC, ""); err != nil {
|
||||
Fatalf("error converting mounts: %v", err)
|
||||
}
|
||||
|
||||
// FIXME: runsc can't be re-executed without
|
||||
// /proc, so we create a tmpfs mount, mount ./proc and ./root
|
||||
// there, then move this mount to the root and after
|
||||
// setCapsAndCallSelf, runsc will chroot into /root.
|
||||
//
|
||||
// We need a directory to construct a new root and we know that
|
||||
// runsc can't start without /proc, so we can use it for this.
|
||||
flags := uintptr(syscall.MS_NOSUID | syscall.MS_NODEV | syscall.MS_NOEXEC)
|
||||
if err := syscall.Mount("runsc-root", "/proc", "tmpfs", flags, ""); err != nil {
|
||||
Fatalf("error mounting tmpfs: %v", err)
|
||||
}
|
||||
os.Mkdir("/proc/proc", 0755)
|
||||
os.Mkdir("/proc/root", 0755)
|
||||
if err := syscall.Mount("runsc-proc", "/proc/proc", "proc", flags|syscall.MS_RDONLY, ""); err != nil {
|
||||
Fatalf("error mounting proc: %v", err)
|
||||
}
|
||||
if err := syscall.Mount(root, "/proc/root", "", syscall.MS_BIND|syscall.MS_REC, ""); err != nil {
|
||||
Fatalf("error mounting root: %v", err)
|
||||
}
|
||||
if err := pivotRoot("/proc"); err != nil {
|
||||
Fatalf("faild to change the root file system: %v", err)
|
||||
}
|
||||
if err := os.Chdir("/"); err != nil {
|
||||
Fatalf("failed to change working directory")
|
||||
}
|
||||
}
|
||||
|
||||
if g.applyCaps {
|
||||
// Disable caps when calling myself again.
|
||||
// Note: minimal argument handling for the default case to keep it simple.
|
||||
args := os.Args
|
||||
args = append(args, "--apply-caps=false")
|
||||
args = append(args, "--apply-caps=false", "--setup-root=false")
|
||||
if err := setCapsAndCallSelf(args, goferCaps); err != nil {
|
||||
Fatalf("Unable to apply caps: %v", err)
|
||||
}
|
||||
panic("unreachable")
|
||||
}
|
||||
|
||||
specFile, err := specutils.OpenCleanSpec(g.bundleDir)
|
||||
if err != nil {
|
||||
Fatalf("opening spec: %v", err)
|
||||
}
|
||||
spec, err := specutils.ReadSpecFromFile(g.bundleDir, specFile)
|
||||
specFile.Close()
|
||||
if err != nil {
|
||||
Fatalf("reading spec: %v", err)
|
||||
}
|
||||
specutils.LogSpec(spec)
|
||||
|
||||
// fsgofer should run with a umask of 0, because we want to preserve file
|
||||
// modes exactly as sent by the sandbox, which will have applied its own umask.
|
||||
syscall.Umask(0)
|
||||
|
||||
// Find what path is going to be served by this gofer.
|
||||
root := spec.Root.Path
|
||||
if !conf.TestOnlyAllowRunAsCurrentUserWithoutChroot {
|
||||
root = "/root"
|
||||
}
|
||||
if err := syscall.Chroot(root); err != nil {
|
||||
Fatalf("failed to chroot to %q: %v", root, err)
|
||||
}
|
||||
|
|
|
@ -787,11 +787,50 @@ func (c *Container) waitForStopped() error {
|
|||
func (c *Container) createGoferProcess(spec *specs.Spec, conf *boot.Config, bundleDir string) ([]*os.File, error) {
|
||||
// Start with the general config flags.
|
||||
args := conf.ToFlags()
|
||||
|
||||
var goferEnds []*os.File
|
||||
|
||||
// nextFD is the next available file descriptor for the gofer process.
|
||||
// It starts at 3 because 0-2 are used by stdin/stdout/stderr.
|
||||
nextFD := 3
|
||||
|
||||
if conf.LogFilename != "" {
|
||||
logFile, err := os.OpenFile(conf.LogFilename, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("opening log file %q: %v", conf.LogFilename, err)
|
||||
}
|
||||
defer logFile.Close()
|
||||
goferEnds = append(goferEnds, logFile)
|
||||
args = append(args, "--log-fd="+strconv.Itoa(nextFD))
|
||||
nextFD++
|
||||
}
|
||||
|
||||
if conf.DebugLog != "" {
|
||||
debugLogFile, err := specutils.DebugLogFile(conf.DebugLog, "gofer")
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("opening debug log file in %q: %v", conf.DebugLog, err)
|
||||
}
|
||||
defer debugLogFile.Close()
|
||||
goferEnds = append(goferEnds, debugLogFile)
|
||||
args = append(args, "--debug-log-fd="+strconv.Itoa(nextFD))
|
||||
nextFD++
|
||||
}
|
||||
|
||||
args = append(args, "gofer", "--bundle", bundleDir)
|
||||
if conf.Overlay {
|
||||
args = append(args, "--panic-on-write=true")
|
||||
}
|
||||
|
||||
// Open the spec file to donate to the sandbox.
|
||||
specFile, err := specutils.OpenCleanSpec(bundleDir)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("opening spec file: %v", err)
|
||||
}
|
||||
defer specFile.Close()
|
||||
goferEnds = append(goferEnds, specFile)
|
||||
args = append(args, "--spec-fd="+strconv.Itoa(nextFD))
|
||||
nextFD++
|
||||
|
||||
// Add root mount and then add any other additional mounts.
|
||||
mountCount := 1
|
||||
|
||||
|
@ -802,12 +841,8 @@ func (c *Container) createGoferProcess(spec *specs.Spec, conf *boot.Config, bund
|
|||
}
|
||||
}
|
||||
sandEnds := make([]*os.File, 0, mountCount)
|
||||
goferEnds := make([]*os.File, 0, mountCount)
|
||||
|
||||
// nextFD is the next available file descriptor for the gofer process.
|
||||
// It starts at 3 because 0-2 are used by stdin/stdout/stderr.
|
||||
nextFD := 3
|
||||
for ; nextFD-3 < mountCount; nextFD++ {
|
||||
for i := 0; i < mountCount; i++ {
|
||||
fds, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_CLOEXEC, 0)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
|
@ -819,6 +854,7 @@ func (c *Container) createGoferProcess(spec *specs.Spec, conf *boot.Config, bund
|
|||
goferEnds = append(goferEnds, goferEnd)
|
||||
|
||||
args = append(args, fmt.Sprintf("--io-fds=%d", nextFD))
|
||||
nextFD++
|
||||
}
|
||||
|
||||
binPath := specutils.ExePath
|
||||
|
|
|
@ -179,8 +179,8 @@ func main() {
|
|||
|
||||
// Quick sanity check to make sure no other commands get passed
|
||||
// a log fd (they should use log dir instead).
|
||||
if subcommand != "boot" {
|
||||
cmd.Fatalf("flag --debug-log-fd should only be passed to 'boot' command, but was passed to %q", subcommand)
|
||||
if subcommand != "boot" && subcommand != "gofer" {
|
||||
cmd.Fatalf("flag --debug-log-fd should only be passed to 'boot' and 'gofer' command, but was passed to %q", subcommand)
|
||||
}
|
||||
|
||||
// If we are the boot process, then we own our stdio FDs and
|
||||
|
|
|
@ -118,10 +118,7 @@ func TestChrootGofer(t *testing.T) {
|
|||
|
||||
// This where the root directory is mapped on the host and that's where the
|
||||
// gofer must have chroot'd to.
|
||||
root, err := d.RootDirInHost()
|
||||
if err != nil {
|
||||
t.Fatalf("Docker.RootDirInHost(): %v", err)
|
||||
}
|
||||
root := "/root"
|
||||
|
||||
for _, child := range children {
|
||||
childPID, err := strconv.Atoi(child)
|
||||
|
|
|
@ -297,15 +297,6 @@ func (d *Docker) SandboxPid() (int, error) {
|
|||
return pid, nil
|
||||
}
|
||||
|
||||
// RootDirInHost returns where the root directory is mapped on the host.
|
||||
func (d *Docker) RootDirInHost() (string, error) {
|
||||
out, err := do("inspect", "-f={{.GraphDriver.Data.MergedDir}}", d.Name)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("error retrieving pid: %v", err)
|
||||
}
|
||||
return strings.TrimSuffix(string(out), "\n"), nil
|
||||
}
|
||||
|
||||
// ID returns the container ID.
|
||||
func (d *Docker) ID() (string, error) {
|
||||
out, err := do("inspect", "-f={{.Id}}", d.Name)
|
||||
|
|
Loading…
Reference in New Issue