diff --git a/vendor/github.com/opencontainers/runc/libcontainer/cgroups/file.go b/vendor/github.com/opencontainers/runc/libcontainer/cgroups/file.go index 0cdaf74784..4da7f365ce 100644 --- a/vendor/github.com/opencontainers/runc/libcontainer/cgroups/file.go +++ b/vendor/github.com/opencontainers/runc/libcontainer/cgroups/file.go @@ -76,16 +76,16 @@ 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} @@ -96,14 +96,16 @@ 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 { @@ -131,7 +133,7 @@ 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, @@ -139,20 +141,20 @@ 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. - fdStr := strconv.Itoa(cgroupFd) + // to reopen cgroupRootHandle and retry openat2. + fdStr := strconv.Itoa(int(cgroupRootHandle.Fd())) fdDest, _ := os.Readlink("/proc/self/fd/" + fdStr) 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 %s unexpectedly opened to %s != %s: %w", - fdStr, fdDest, cgroupfsDir, err) + err = fmt.Errorf("cgroupRootHandle %d unexpectedly opened to %s != %s: %w", + cgroupRootHandle.Fd(), fdDest, cgroupfsDir, err) } return nil, err } diff --git a/vendor/github.com/opencontainers/runc/libcontainer/cgroups/fs/paths.go b/vendor/github.com/opencontainers/runc/libcontainer/cgroups/fs/paths.go index 1092331b25..2cb970a3d5 100644 --- a/vendor/github.com/opencontainers/runc/libcontainer/cgroups/fs/paths.go +++ b/vendor/github.com/opencontainers/runc/libcontainer/cgroups/fs/paths.go @@ -83,6 +83,7 @@ func tryDefaultCgroupRoot() string { if err != nil { return "" } + defer dir.Close() names, err := dir.Readdirnames(1) if err != nil { return "" diff --git a/vendor/github.com/opencontainers/runc/libcontainer/utils/utils_unix.go b/vendor/github.com/opencontainers/runc/libcontainer/utils/utils_unix.go index 220d0b4393..bf3237a291 100644 --- a/vendor/github.com/opencontainers/runc/libcontainer/utils/utils_unix.go +++ b/vendor/github.com/opencontainers/runc/libcontainer/utils/utils_unix.go @@ -7,6 +7,7 @@ import ( "fmt" "os" "strconv" + _ "unsafe" // for go:linkname "golang.org/x/sys/unix" ) @@ -23,9 +24,11 @@ func EnsureProcHandle(fh *os.File) error { return nil } -// 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 { +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 { fdDir, err := os.Open("/proc/self/fd") if err != nil { return err @@ -50,15 +53,60 @@ 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 { + 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 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 unix socket pair func NewSockPair(name string) (parent *os.File, child *os.File, err error) { fds, err := unix.Socketpair(unix.AF_LOCAL, unix.SOCK_STREAM|unix.SOCK_CLOEXEC, 0)