Merge pull request from GHSA-xr7r-f8xq-vfvv

fix GHSA-xr7r-f8xq-vfvv and harden fd leaks

libcontainer/cgroups/file.go

@@ -66,16 +66,16 @@ var (
 	// TestMode is set to true by unit tests that need "fake" cgroupfs.
 	TestMode bool
 
-	cgroupFd     int = -1
-	prepOnce     sync.Once
-	prepErr      error
-	resolveFlags uint64
+	cgroupRootHandle *os.File
+	prepOnce         sync.Once
+	prepErr          error
+	resolveFlags     uint64
 )
 
 func prepareOpenat2() error {
 	prepOnce.Do(func() {
 		fd, err := unix.Openat2(-1, cgroupfsDir, &unix.OpenHow{
-			Flags: unix.O_DIRECTORY | unix.O_PATH,
+			Flags: unix.O_DIRECTORY | unix.O_PATH | unix.O_CLOEXEC,
 		})
 		if err != nil {
 			prepErr = &os.PathError{Op: "openat2", Path: cgroupfsDir, Err: err}
@@ -86,15 +86,16 @@ func prepareOpenat2() error {
 			}
 			return
 		}
+		file := os.NewFile(uintptr(fd), cgroupfsDir)
+
 		var st unix.Statfs_t
-		if err = unix.Fstatfs(fd, &st); err != nil {
+		if err := unix.Fstatfs(int(file.Fd()), &st); err != nil {
 			prepErr = &os.PathError{Op: "statfs", Path: cgroupfsDir, Err: err}
 			logrus.Warnf("falling back to securejoin: %s", prepErr)
 			return
 		}
 
-		cgroupFd = fd
-
+		cgroupRootHandle = file
 		resolveFlags = unix.RESOLVE_BENEATH | unix.RESOLVE_NO_MAGICLINKS
 		if st.Type == unix.CGROUP2_SUPER_MAGIC {
 			// cgroupv2 has a single mountpoint and no "cpu,cpuacct" symlinks
@@ -121,7 +122,7 @@ func openFile(dir, file string, flags int) (*os.File, error) {
 		return openFallback(path, flags, mode)
 	}
 
-	fd, err := unix.Openat2(cgroupFd, relPath,
+	fd, err := unix.Openat2(int(cgroupRootHandle.Fd()), relPath,
 		&unix.OpenHow{
 			Resolve: resolveFlags,
 			Flags:   uint64(flags) | unix.O_CLOEXEC,
@@ -129,21 +130,21 @@ func openFile(dir, file string, flags int) (*os.File, error) {
 		})
 	if err != nil {
 		err = &os.PathError{Op: "openat2", Path: path, Err: err}
-		// Check if cgroupFd is still opened to cgroupfsDir
+		// Check if cgroupRootHandle is still opened to cgroupfsDir
 		// (happens when this package is incorrectly used
 		// across the chroot/pivot_root/mntns boundary, or
 		// when /sys/fs/cgroup is remounted).
 		//
 		// TODO: if such usage will ever be common, amend this
-		// to reopen cgroupFd and retry openat2.
-		fdPath, closer := utils.ProcThreadSelf("fd/" + strconv.Itoa(cgroupFd))
+		// to reopen cgroupRootHandle and retry openat2.
+		fdPath, closer := utils.ProcThreadSelf("fd/" + strconv.Itoa(int(cgroupRootHandle.Fd())))
 		defer closer()
 		fdDest, _ := os.Readlink(fdPath)
 		if fdDest != cgroupfsDir {
-			// Wrap the error so it is clear that cgroupFd
+			// Wrap the error so it is clear that cgroupRootHandle
 			// is opened to an unexpected/wrong directory.
-			err = fmt.Errorf("cgroupFd %d unexpectedly opened to %s != %s: %w",
-				cgroupFd, fdDest, cgroupfsDir, err)
+			err = fmt.Errorf("cgroupRootHandle %d unexpectedly opened to %s != %s: %w",
+				cgroupRootHandle.Fd(), fdDest, cgroupfsDir, err)
 		}
 		return nil, err
 	}

libcontainer/container_linux.go

@@ -332,6 +332,15 @@ func (c *Container) start(process *Process) (retErr error) {
 		}()
 	}
 
+	// Before starting "runc init", mark all non-stdio open files as O_CLOEXEC
+	// to make sure we don't leak any files into "runc init". Any files to be
+	// passed to "runc init" through ExtraFiles will get dup2'd by the Go
+	// runtime and thus their O_CLOEXEC flag will be cleared. This is some
+	// additional protection against attacks like CVE-2024-21626, by making
+	// sure we never leak files to "runc init" we didn't intend to.
+	if err := utils.CloseExecFrom(3); err != nil {
+		return fmt.Errorf("unable to mark non-stdio fds as cloexec: %w", err)
+	}
 	if err := parent.start(); err != nil {
 		return fmt.Errorf("unable to start container process: %w", err)
 	}

libcontainer/init_linux.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"net"
 	"os"
+	"path/filepath"
 	"runtime"
 	"runtime/debug"
 	"strconv"
@@ -89,7 +90,7 @@ func Init() {
 	}
 	// Normally, StartInitialization() never returns, meaning
 	// if we are here, it had failed.
-	os.Exit(1)
+	os.Exit(255)
 }
 
 // Normally, this function does not return. If it returns, with or without an
@@ -274,6 +275,32 @@ func populateProcessEnvironment(env []string) error {
 	return nil
 }
 
+// verifyCwd ensures that the current directory is actually inside the mount
+// namespace root of the current process.
+func verifyCwd() error {
+	// getcwd(2) on Linux detects if cwd is outside of the rootfs of the
+	// current mount namespace root, and in that case prefixes "(unreachable)"
+	// to the returned string. glibc's getcwd(3) and Go's Getwd() both detect
+	// when this happens and return ENOENT rather than returning a non-absolute
+	// path. In both cases we can therefore easily detect if we have an invalid
+	// cwd by checking the return value of getcwd(3). See getcwd(3) for more
+	// details, and CVE-2024-21626 for the security issue that motivated this
+	// check.
+	//
+	// We have to use unix.Getwd() here because os.Getwd() has a workaround for
+	// $PWD which involves doing stat(.), which can fail if the current
+	// directory is inaccessible to the container process.
+	if wd, err := unix.Getwd(); errors.Is(err, unix.ENOENT) {
+		return errors.New("current working directory is outside of container mount namespace root -- possible container breakout detected")
+	} else if err != nil {
+		return fmt.Errorf("failed to verify if current working directory is safe: %w", err)
+	} else if !filepath.IsAbs(wd) {
+		// We shouldn't ever hit this, but check just in case.
+		return fmt.Errorf("current working directory is not absolute -- possible container breakout detected: cwd is %q", wd)
+	}
+	return nil
+}
+
 // finalizeNamespace drops the caps, sets the correct user
 // and working dir, and closes any leaked file descriptors
 // before executing the command inside the namespace
@@ -332,6 +359,10 @@ func finalizeNamespace(config *initConfig) error {
 			return fmt.Errorf("chdir to cwd (%q) set in config.json failed: %w", config.Cwd, err)
 		}
 	}
+	// Make sure our final working directory is inside the container.
+	if err := verifyCwd(); err != nil {
+		return err
+	}
 	if err := system.ClearKeepCaps(); err != nil {
 		return fmt.Errorf("unable to clear keep caps: %w", err)
 	}

libcontainer/integration/seccomp_test.go

@@ -13,7 +13,7 @@ import (
 	libseccomp "github.com/seccomp/libseccomp-golang"
 )
 
-func TestSeccompDenyGetcwdWithErrno(t *testing.T) {
+func TestSeccompDenySyslogWithErrno(t *testing.T) {
 	if testing.Short() {
 		return
 	}
@@ -25,7 +25,7 @@ func TestSeccompDenyGetcwdWithErrno(t *testing.T) {
 		DefaultAction: configs.Allow,
 		Syscalls: []*configs.Syscall{
 			{
-				Name:     "getcwd",
+				Name:     "syslog",
 				Action:   configs.Errno,
 				ErrnoRet: &errnoRet,
 			},
@@ -39,7 +39,7 @@ func TestSeccompDenyGetcwdWithErrno(t *testing.T) {
 	buffers := newStdBuffers()
 	pwd := &libcontainer.Process{
 		Cwd:    "/",
-		Args:   []string{"pwd"},
+		Args:   []string{"dmesg"},
 		Env:    standardEnvironment,
 		Stdin:  buffers.Stdin,
 		Stdout: buffers.Stdout,
@@ -65,17 +65,17 @@ func TestSeccompDenyGetcwdWithErrno(t *testing.T) {
 	}
 
 	if exitCode == 0 {
-		t.Fatalf("Getcwd should fail with negative exit code, instead got %d!", exitCode)
+		t.Fatalf("dmesg should fail with negative exit code, instead got %d!", exitCode)
 	}
 
-	expected := "pwd: getcwd: No such process"
+	expected := "dmesg: klogctl: No such process"
 	actual := strings.Trim(buffers.Stderr.String(), "\n")
 	if actual != expected {
 		t.Fatalf("Expected output %s but got %s\n", expected, actual)
 	}
 }
 
-func TestSeccompDenyGetcwd(t *testing.T) {
+func TestSeccompDenySyslog(t *testing.T) {
 	if testing.Short() {
 		return
 	}
@@ -85,7 +85,7 @@ func TestSeccompDenyGetcwd(t *testing.T) {
 		DefaultAction: configs.Allow,
 		Syscalls: []*configs.Syscall{
 			{
-				Name:   "getcwd",
+				Name:   "syslog",
 				Action: configs.Errno,
 			},
 		},
@@ -98,7 +98,7 @@ func TestSeccompDenyGetcwd(t *testing.T) {
 	buffers := newStdBuffers()
 	pwd := &libcontainer.Process{
 		Cwd:    "/",
-		Args:   []string{"pwd"},
+		Args:   []string{"dmesg"},
 		Env:    standardEnvironment,
 		Stdin:  buffers.Stdin,
 		Stdout: buffers.Stdout,
@@ -124,10 +124,10 @@ func TestSeccompDenyGetcwd(t *testing.T) {
 	}
 
 	if exitCode == 0 {
-		t.Fatalf("Getcwd should fail with negative exit code, instead got %d!", exitCode)
+		t.Fatalf("dmesg should fail with negative exit code, instead got %d!", exitCode)
 	}
 
-	expected := "pwd: getcwd: Operation not permitted"
+	expected := "dmesg: klogctl: Operation not permitted"
 	actual := strings.Trim(buffers.Stderr.String(), "\n")
 	if actual != expected {
 		t.Fatalf("Expected output %s but got %s\n", expected, actual)

libcontainer/setns_init_linux.go

@@ -14,6 +14,7 @@ import (
 	"github.com/opencontainers/runc/libcontainer/keys"
 	"github.com/opencontainers/runc/libcontainer/seccomp"
 	"github.com/opencontainers/runc/libcontainer/system"
+	"github.com/opencontainers/runc/libcontainer/utils"
 )
 
 // linuxSetnsInit performs the container's initialization for running a new process
@@ -138,5 +139,23 @@ func (l *linuxSetnsInit) Init() error {
 		l.config.Args[0] = name
 		return system.Fexecve(l.dmzExe.Fd(), l.config.Args, os.Environ())
 	}
+	// Close all file descriptors we are not passing to the container. This is
+	// necessary because the execve target could use internal runc fds as the
+	// execve path, potentially giving access to binary files from the host
+	// (which can then be opened by container processes, leading to container
+	// escapes). Note that because this operation will close any open file
+	// descriptors that are referenced by (*os.File) handles from underneath
+	// the Go runtime, we must not do any file operations after this point
+	// (otherwise the (*os.File) finaliser could close the wrong file). See
+	// CVE-2024-21626 for more information as to why this protection is
+	// necessary.
+	//
+	// This is not needed for runc-dmz, because the extra execve(2) step means
+	// that all O_CLOEXEC file descriptors have already been closed and thus
+	// the second execve(2) from runc-dmz cannot access internal file
+	// descriptors from runc.
+	if err := utils.UnsafeCloseFrom(l.config.PassedFilesCount + 3); err != nil {
+		return err
+	}
 	return system.Exec(name, l.config.Args, os.Environ())
 }

libcontainer/standard_init_linux.go

@@ -281,5 +281,23 @@ func (l *linuxStandardInit) Init() error {
 		l.config.Args[0] = name
 		return system.Fexecve(l.dmzExe.Fd(), l.config.Args, os.Environ())
 	}
+	// Close all file descriptors we are not passing to the container. This is
+	// necessary because the execve target could use internal runc fds as the
+	// execve path, potentially giving access to binary files from the host
+	// (which can then be opened by container processes, leading to container
+	// escapes). Note that because this operation will close any open file
+	// descriptors that are referenced by (*os.File) handles from underneath
+	// the Go runtime, we must not do any file operations after this point
+	// (otherwise the (*os.File) finaliser could close the wrong file). See
+	// CVE-2024-21626 for more information as to why this protection is
+	// necessary.
+	//
+	// This is not needed for runc-dmz, because the extra execve(2) step means
+	// that all O_CLOEXEC file descriptors have already been closed and thus
+	// the second execve(2) from runc-dmz cannot access internal file
+	// descriptors from runc.
+	if err := utils.UnsafeCloseFrom(l.config.PassedFilesCount + 3); err != nil {
+		return err
+	}
 	return system.Exec(name, l.config.Args, os.Environ())
 }

libcontainer/utils/utils_unix.go

@@ -11,6 +11,7 @@ import (
 	"runtime"
 	"strconv"
 	"sync"
+	_ "unsafe" // for go:linkname
 
 	securejoin "github.com/cyphar/filepath-securejoin"
 	"github.com/sirupsen/logrus"
@@ -53,14 +54,11 @@ func haveCloseRangeCloexec() bool {
 	return haveCloseRangeCloexecBool
 }
 
-// CloseExecFrom applies O_CLOEXEC to all file descriptors currently open for
-// the process (except for those below the given fd value).
-func CloseExecFrom(minFd int) error {
-	if haveCloseRangeCloexec() {
-		err := unix.CloseRange(uint(minFd), math.MaxUint, unix.CLOSE_RANGE_CLOEXEC)
-		return os.NewSyscallError("close_range", err)
-	}
+type fdFunc func(fd int)
 
+// fdRangeFrom calls the passed fdFunc for each file descriptor that is open in
+// the current process.
+func fdRangeFrom(minFd int, fn fdFunc) error {
 	procSelfFd, closer := ProcThreadSelf("fd")
 	defer closer()
 
@@ -88,15 +86,67 @@ func CloseExecFrom(minFd int) error {
 		if fd < minFd {
 			continue
 		}
-		// Intentionally ignore errors from unix.CloseOnExec -- the cases where
-		// this might fail are basically file descriptors that have already
-		// been closed (including and especially the one that was created when
-		// os.ReadDir did the "opendir" syscall).
-		unix.CloseOnExec(fd)
+		// Ignore the file descriptor we used for readdir, as it will be closed
+		// when we return.
+		if uintptr(fd) == fdDir.Fd() {
+			continue
+		}
+		// Run the closure.
+		fn(fd)
 	}
 	return nil
 }
 
+// CloseExecFrom sets the O_CLOEXEC flag on all file descriptors greater or
+// equal to minFd in the current process.
+func CloseExecFrom(minFd int) error {
+	// Use close_range(CLOSE_RANGE_CLOEXEC) if possible.
+	if haveCloseRangeCloexec() {
+		err := unix.CloseRange(uint(minFd), math.MaxUint, unix.CLOSE_RANGE_CLOEXEC)
+		return os.NewSyscallError("close_range", err)
+	}
+	// Otherwise, fall back to the standard loop.
+	return fdRangeFrom(minFd, unix.CloseOnExec)
+}
+
+//go:linkname runtime_IsPollDescriptor internal/poll.IsPollDescriptor
+
+// In order to make sure we do not close the internal epoll descriptors the Go
+// runtime uses, we need to ensure that we skip descriptors that match
+// "internal/poll".IsPollDescriptor. Yes, this is a Go runtime internal thing,
+// unfortunately there's no other way to be sure we're only keeping the file
+// descriptors the Go runtime needs. Hopefully nothing blows up doing this...
+func runtime_IsPollDescriptor(fd uintptr) bool //nolint:revive
+
+// UnsafeCloseFrom closes all file descriptors greater or equal to minFd in the
+// current process, except for those critical to Go's runtime (such as the
+// netpoll management descriptors).
+//
+// NOTE: That this function is incredibly dangerous to use in most Go code, as
+// closing file descriptors from underneath *os.File handles can lead to very
+// bad behaviour (the closed file descriptor can be re-used and then any
+// *os.File operations would apply to the wrong file). This function is only
+// intended to be called from the last stage of runc init.
+func UnsafeCloseFrom(minFd int) error {
+	// We cannot use close_range(2) even if it is available, because we must
+	// not close some file descriptors.
+	return fdRangeFrom(minFd, func(fd int) {
+		if runtime_IsPollDescriptor(uintptr(fd)) {
+			// These are the Go runtimes internal netpoll file descriptors.
+			// These file descriptors are operated on deep in the Go scheduler,
+			// and closing those files from underneath Go can result in panics.
+			// There is no issue with keeping them because they are not
+			// executable and are not useful to an attacker anyway. Also we
+			// don't have any choice.
+			return
+		}
+		// There's nothing we can do about errors from close(2), and the
+		// only likely error to be seen is EBADF which indicates the fd was
+		// already closed (in which case, we got what we wanted).
+		_ = unix.Close(fd)
+	})
+}
+
 // NewSockPair returns a new SOCK_STREAM unix socket pair.
 func NewSockPair(name string) (parent, child *os.File, err error) {
 	fds, err := unix.Socketpair(unix.AF_LOCAL, unix.SOCK_STREAM|unix.SOCK_CLOEXEC, 0)