libcontainer: Don't use UsetCloseOnExec, it is racy

We can't keep file descriptors without close-on-exec except with
syscall.ForkLock held, as otherwise they could leak by accident into
other children from forks in other threads.

Instead we just use Cmd.ExtraFiles which handles all this for us.

This fixes https://github.com/dotcloud/docker/issues/4493

Docker-DCO-1.1-Signed-off-by: Alexander Larsson <alexl@redhat.com> (github: alexlarsson)

execdriver/native/driver.go

@@ -239,7 +239,7 @@ type dockerCommandFactory struct {
 // createCommand will return an exec.Cmd with the Cloneflags set to the proper namespaces
 // defined on the container's configuration and use the current binary as the init with the
 // args provided
-func (d *dockerCommandFactory) Create(container *libcontainer.Container, console string, syncFd uintptr, args []string) *exec.Cmd {
+func (d *dockerCommandFactory) Create(container *libcontainer.Container, console string, syncFile *os.File, args []string) *exec.Cmd {
 	// we need to join the rootfs because nsinit will setup the rootfs and chroot
 	initPath := filepath.Join(d.c.Rootfs, d.c.InitPath)
 
@@ -248,7 +248,7 @@ func (d *dockerCommandFactory) Create(container *libcontainer.Container, console
 		initPath,
 		"-driver", DriverName,
 		"-console", console,
-		"-pipe", fmt.Sprint(syncFd),
+		"-pipe", "3",
 		"-root", filepath.Join(d.driver.root, d.c.ID),
 		"--",
 	}, args...)
@@ -256,6 +256,7 @@ func (d *dockerCommandFactory) Create(container *libcontainer.Container, console
 	// set this to nil so that when we set the clone flags anything else is reset
 	d.c.SysProcAttr = nil
 	system.SetCloneFlags(&d.c.Cmd, uintptr(nsinit.GetNamespaceFlags(container.Namespaces)))
+	d.c.ExtraFiles = []*os.File{syncFile}
 
 	d.c.Env = container.Env
 	d.c.Dir = d.c.Rootfs