| 1 | //go:build linux |
| 2 | |
| 3 | package instancefs |
| 4 | |
| 5 | import ( |
| 6 | "bufio" |
| 7 | "crypto/rand" |
| 8 | "encoding/hex" |
| 9 | "errors" |
| 10 | "fmt" |
| 11 | "io" |
| 12 | "os" |
| 13 | "path/filepath" |
| 14 | "strconv" |
| 15 | "strings" |
| 16 | |
| 17 | "golang.org/x/sys/unix" |
| 18 | ) |
| 19 | |
| 20 | const resolveFlags = unix.RESOLVE_BENEATH | unix.RESOLVE_NO_SYMLINKS | unix.RESOLVE_NO_MAGICLINKS |
| 21 | |
| 22 | // openDir opens dir itself as a directory fd, refusing a symlink at ANY |
| 23 | // component of dir, not only its final one: the instance directory's |
| 24 | // parent is the agent's, but a same-uid process could have swapped an |
| 25 | // intermediate component for a symlink before the mount option landed, |
| 26 | // and openat2's RESOLVE_NO_SYMLINKS catches that the way a plain |
| 27 | // O_NOFOLLOW open (which only refuses a symlink at the leaf) would not. |
| 28 | // RESOLVE_BENEATH is deliberately absent here: dir is an absolute path |
| 29 | // handed in by the caller, not a name resolved beneath some other fd, so |
| 30 | // there is no "beneath" for it to stay under. |
| 31 | func openDir(dir string) (int, error) { |
| 32 | how := unix.OpenHow{ |
| 33 | Flags: unix.O_PATH | unix.O_DIRECTORY | unix.O_CLOEXEC, |
| 34 | Resolve: unix.RESOLVE_NO_SYMLINKS | unix.RESOLVE_NO_MAGICLINKS, |
| 35 | } |
| 36 | fd, err := unix.Openat2(unix.AT_FDCWD, dir, &how) |
| 37 | if err != nil { |
| 38 | if errors.Is(err, unix.ELOOP) { |
| 39 | return -1, fmt.Errorf("%w: %s", ErrSymlink, dir) |
| 40 | } |
| 41 | return -1, &os.PathError{Op: "openat2", Path: dir, Err: err} |
| 42 | } |
| 43 | return fd, nil |
| 44 | } |
| 45 | |
| 46 | // SelfCheck is the one call a process makes before it trusts the rest of |
| 47 | // this package: it opens dir exactly as openDir does and reports whether |
| 48 | // the kernel let the call through at all. It exists because a seccomp |
| 49 | // filter is invisible from inside: the agent would learn of one from the |
| 50 | // first nested open that fails, which is to say from a layout migration |
| 51 | // half-way through its first host. ENOSYS is ErrOpenat2Unavailable, named |
| 52 | // so a gate can say what to fix; any other error is dir's own and is |
| 53 | // returned as it is. The probe unit of design §8.5 runs under the |
| 54 | // instance templates' text and so cannot see the edge unit's; this is the |
| 55 | // edge's own check, run by the agent in its own unit. |
| 56 | func SelfCheck(dir string) error { |
| 57 | fd, err := openDir(dir) |
| 58 | if err != nil { |
| 59 | if errors.Is(err, unix.ENOSYS) { |
| 60 | return fmt.Errorf("%w: %s", ErrOpenat2Unavailable, dir) |
| 61 | } |
| 62 | return err |
| 63 | } |
| 64 | unix.Close(fd) |
| 65 | return nil |
| 66 | } |
| 67 | |
| 68 | // OpenBeneath opens name under dir with openat2, following no symlink at |
| 69 | // any component and never leaving dir. Every open, whatever its flags, |
| 70 | // carries O_NONBLOCK on the wire and is then checked with fstat: a FIFO |
| 71 | // with no reader on the other end blocks a write-open indefinitely, and |
| 72 | // one with no writer blocks a read-open indefinitely, and a tenant can |
| 73 | // plant one anywhere under an instance directory the agent reads OR |
| 74 | // writes by contract (its run directory, its env files, its admin token |
| 75 | // file). O_NONBLOCK changes the two blocking cases without changing the |
| 76 | // non-blocking ones: opening a regular file (the common case, and every |
| 77 | // O_CREAT|O_EXCL caller's own fresh file) behaves identically either way, |
| 78 | // a read-open of a FIFO with no writer now returns at once instead of |
| 79 | // blocking, and a write-open of a FIFO with no reader answers ENXIO at |
| 80 | // once instead of blocking. Whatever DOES come back with a valid fd (a |
| 81 | // FIFO opened for read, or for write against an existing reader) is then |
| 82 | // refused by the fstat check unless it is S_IFREG, so nothing non-regular |
| 83 | // can be read or written through this function at all. |
| 84 | func OpenBeneath(dir, name string, flag int, perm os.FileMode) (*os.File, error) { |
| 85 | return openBeneathAs(dir, name, flag, perm, unix.S_IFREG) |
| 86 | } |
| 87 | |
| 88 | // OpenDirBeneath opens name under dir as a directory, read-only, exactly as |
| 89 | // OpenBeneath opens a file: openat2 with no symlink at any component of dir |
| 90 | // or name, O_NONBLOCK, and then fstat, which accepts only a directory. A |
| 91 | // name of "." is dir itself. The descriptor is the directory that was |
| 92 | // vetted: a caller that works from it (an *at call, or OpenRootBeneath) |
| 93 | // never resolves the path again, so a link swapped in after the open is |
| 94 | // never followed. |
| 95 | func OpenDirBeneath(dir, name string) (*os.File, error) { |
| 96 | return openBeneathAs(dir, name, os.O_RDONLY|unix.O_DIRECTORY, 0, unix.S_IFDIR) |
| 97 | } |
| 98 | |
| 99 | // OpenRootBeneath is OpenDirBeneath as an *os.Root, for code built on |
| 100 | // os.Root (hostfiles.RestoreRoot). os.OpenRoot takes only a path and opens |
| 101 | // it without O_NONBLOCK, so it would follow a link at the name and block |
| 102 | // on a FIFO there. This opens the directory with OpenDirBeneath and hands |
| 103 | // os.OpenRoot that descriptor's /proc/self/fd entry, a magic link the |
| 104 | // kernel resolves to the very directory the descriptor holds: no |
| 105 | // component is resolved twice. Without /proc it is an error, never a |
| 106 | // fallback to the path. |
| 107 | func OpenRootBeneath(dir, name string) (*os.Root, error) { |
| 108 | f, err := OpenDirBeneath(dir, name) |
| 109 | if err != nil { |
| 110 | return nil, err |
| 111 | } |
| 112 | defer f.Close() |
| 113 | betweenOpenAndRoot() |
| 114 | r, err := os.OpenRoot(procSelfFD + strconv.Itoa(int(f.Fd()))) |
| 115 | if err != nil { |
| 116 | return nil, fmt.Errorf("instancefs: open %s as a root: %w", filepath.Join(dir, name), err) |
| 117 | } |
| 118 | return r, nil |
| 119 | } |
| 120 | |
| 121 | // betweenOpenAndRoot is a test seam: where a path swapped after the open |
| 122 | // would race a root opened by path. |
| 123 | var betweenOpenAndRoot = func() {} |
| 124 | |
| 125 | // procSelfFD is where OpenRootBeneath finds its descriptor's magic link. A |
| 126 | // test seam: a test points it at a missing directory to pin that no /proc |
| 127 | // is an error, never a fallback to the path. |
| 128 | var procSelfFD = "/proc/self/fd/" |
| 129 | |
| 130 | // OpenBeneathOrAbsent opens name under dir read-only as OpenBeneath does, |
| 131 | // and is the one place "absent" is decided on ground someone else can |
| 132 | // write. It opens dir once and keeps that descriptor: absent is true only |
| 133 | // when the lookup of name beneath that very descriptor answers ENOENT. A |
| 134 | // dir that cannot be opened (missing, a symlink, not a directory) is an |
| 135 | // error wrapped in ErrDirUnopened, never absence: closing a check's |
| 136 | // descriptor and opening the path again would read a directory renamed |
| 137 | // away in between as a missing file. |
| 138 | func OpenBeneathOrAbsent(dir, name string) (*os.File, bool, error) { |
| 139 | rel, err := cleanRel(name) |
| 140 | if err != nil { |
| 141 | return nil, false, err |
| 142 | } |
| 143 | dfd, err := openDir(dir) |
| 144 | if err != nil { |
| 145 | return nil, false, fmt.Errorf("%w: %w", ErrDirUnopened, err) |
| 146 | } |
| 147 | defer unix.Close(dfd) |
| 148 | betweenDirAndLookup() |
| 149 | f, err := openAt(dfd, dir, rel, os.O_RDONLY, 0, unix.S_IFREG) |
| 150 | if errors.Is(err, unix.ENOENT) { |
| 151 | return nil, true, nil |
| 152 | } |
| 153 | if err != nil { |
| 154 | return nil, false, err |
| 155 | } |
| 156 | return f, false, nil |
| 157 | } |
| 158 | |
| 159 | // betweenDirAndLookup is a test seam: after OpenBeneathOrAbsent has opened |
| 160 | // dir and before it looks name up beneath it. |
| 161 | var betweenDirAndLookup = func() {} |
| 162 | |
| 163 | // openBeneathAs is OpenBeneath's open, accepting only the file type want |
| 164 | // (S_IFREG or S_IFDIR) from fstat. |
| 165 | func openBeneathAs(dir, name string, flag int, perm os.FileMode, want uint32) (*os.File, error) { |
| 166 | rel, err := cleanRel(name) |
| 167 | if err != nil { |
| 168 | return nil, err |
| 169 | } |
| 170 | dfd, err := openDir(dir) |
| 171 | if err != nil { |
| 172 | return nil, err |
| 173 | } |
| 174 | defer unix.Close(dfd) |
| 175 | return openAt(dfd, dir, rel, flag, perm, want) |
| 176 | } |
| 177 | |
| 178 | // openAt is the open every function here shares: rel beneath the |
| 179 | // directory descriptor dfd (which names dir, for errors), openat2 with no |
| 180 | // symlink at any component, O_NONBLOCK, and then fstat, accepting only the |
| 181 | // file type want. |
| 182 | func openAt(dfd int, dir, rel string, flag int, perm os.FileMode, want uint32) (*os.File, error) { |
| 183 | how := unix.OpenHow{ |
| 184 | Flags: uint64(flag) | unix.O_NONBLOCK | unix.O_CLOEXEC, |
| 185 | Mode: uint64(perm.Perm()), |
| 186 | Resolve: resolveFlags, |
| 187 | } |
| 188 | fd, err := unix.Openat2(dfd, rel, &how) |
| 189 | if err != nil { |
| 190 | switch { |
| 191 | case errors.Is(err, unix.ELOOP): |
| 192 | return nil, fmt.Errorf("%w: %s", ErrSymlink, filepath.Join(dir, rel)) |
| 193 | case errors.Is(err, unix.EXDEV): |
| 194 | return nil, fmt.Errorf("%w: %s", ErrEscapes, filepath.Join(dir, rel)) |
| 195 | } |
| 196 | return nil, &os.PathError{Op: "openat2", Path: filepath.Join(dir, rel), Err: err} |
| 197 | } |
| 198 | path := filepath.Join(dir, rel) |
| 199 | var st unix.Stat_t |
| 200 | if err := unix.Fstat(fd, &st); err != nil { |
| 201 | unix.Close(fd) |
| 202 | return nil, &os.PathError{Op: "fstat", Path: path, Err: err} |
| 203 | } |
| 204 | if st.Mode&unix.S_IFMT != want { |
| 205 | unix.Close(fd) |
| 206 | if want == unix.S_IFDIR { |
| 207 | return nil, fmt.Errorf("%w: %s", ErrNotDir, path) |
| 208 | } |
| 209 | return nil, fmt.Errorf("%w: %s", ErrNotRegular, path) |
| 210 | } |
| 211 | return os.NewFile(uintptr(fd), path), nil |
| 212 | } |
| 213 | |
| 214 | // WriteAtomic writes data to a fresh O_EXCL temp file beside name and |
| 215 | // renames it over name. rename(2) replaces the directory entry and never |
| 216 | // follows a symlink at the destination, so a planted link is displaced, |
| 217 | // not written through. The temp name is random, so no same-uid process |
| 218 | // can pre-plant it (the O_EXCL refuses it if one does). |
| 219 | func WriteAtomic(dir, name string, data []byte, perm os.FileMode) error { |
| 220 | rel, err := cleanRel(name) |
| 221 | if err != nil { |
| 222 | return err |
| 223 | } |
| 224 | if strings.Contains(rel, "/") { |
| 225 | return fmt.Errorf("%w: WriteAtomic takes a bare file name, got %q", ErrEscapes, name) |
| 226 | } |
| 227 | dfd, err := openDir(dir) |
| 228 | if err != nil { |
| 229 | return err |
| 230 | } |
| 231 | defer unix.Close(dfd) |
| 232 | var rnd [8]byte |
| 233 | if _, err := rand.Read(rnd[:]); err != nil { |
| 234 | return err |
| 235 | } |
| 236 | tmp := "." + rel + ".tmp-" + hex.EncodeToString(rnd[:]) |
| 237 | how := unix.OpenHow{ |
| 238 | Flags: unix.O_WRONLY | unix.O_CREAT | unix.O_EXCL | unix.O_CLOEXEC, |
| 239 | Mode: uint64(perm.Perm()), |
| 240 | Resolve: resolveFlags, |
| 241 | } |
| 242 | fd, err := unix.Openat2(dfd, tmp, &how) |
| 243 | if err != nil { |
| 244 | return &os.PathError{Op: "openat2", Path: filepath.Join(dir, tmp), Err: err} |
| 245 | } |
| 246 | f := os.NewFile(uintptr(fd), filepath.Join(dir, tmp)) |
| 247 | cleanup := func() { f.Close(); _ = unix.Unlinkat(dfd, tmp, 0) } |
| 248 | if _, err := f.Write(data); err != nil { |
| 249 | cleanup() |
| 250 | return err |
| 251 | } |
| 252 | if err := f.Sync(); err != nil { |
| 253 | cleanup() |
| 254 | return err |
| 255 | } |
| 256 | if err := f.Close(); err != nil { |
| 257 | _ = unix.Unlinkat(dfd, tmp, 0) |
| 258 | return err |
| 259 | } |
| 260 | if err := unix.Renameat(dfd, tmp, dfd, rel); err != nil { |
| 261 | _ = unix.Unlinkat(dfd, tmp, 0) |
| 262 | return &os.PathError{Op: "renameat", Path: filepath.Join(dir, rel), Err: err} |
| 263 | } |
| 264 | return nil |
| 265 | } |
| 266 | |
| 267 | // MountOptionOf reports whether mountpoint is a mountpoint whose per-mount |
| 268 | // options include opt, by reading /proc/self/mountinfo. A path that is not |
| 269 | // a mountpoint is (false, nil). |
| 270 | func MountOptionOf(mountpoint, opt string) (bool, error) { |
| 271 | f, err := os.Open("/proc/self/mountinfo") |
| 272 | if err != nil { |
| 273 | return false, err |
| 274 | } |
| 275 | defer f.Close() |
| 276 | return mountOptionOf(f, mountpoint, opt) |
| 277 | } |
| 278 | |
| 279 | // MountNoSymfollow reports whether mountpoint carries nosymfollow. It |
| 280 | // predates MountOptionOf and stays as a thin wrapper because most callers |
| 281 | // only ever ask this one question and naming it directly reads better at |
| 282 | // the call site than spelling the option out every time. |
| 283 | func MountNoSymfollow(mountpoint string) (bool, error) { |
| 284 | return MountOptionOf(mountpoint, "nosymfollow") |
| 285 | } |
| 286 | |
| 287 | // mountOptionOf is MountOptionOf's body, taking its mountinfo source as an |
| 288 | // io.Reader so a test can feed it a fixture with stacked mount lines |
| 289 | // instead of the live /proc/self/mountinfo. |
| 290 | // |
| 291 | // Stacked mounts at one mountpoint (a bind mount, then a later remount) |
| 292 | // list oldest first in mountinfo, and the LAST line for a mountpoint is |
| 293 | // the one actually in effect. Every line is therefore scanned to the end |
| 294 | // and the last matching line's verdict remembered, rather than returning |
| 295 | // on the first match: an earlier stacked mount carrying opt must not |
| 296 | // outvote a later one that dropped it, or vice versa. |
| 297 | func mountOptionOf(r io.Reader, mountpoint, opt string) (bool, error) { |
| 298 | want := filepath.Clean(mountpoint) |
| 299 | sc := bufio.NewScanner(r) |
| 300 | verdict := false |
| 301 | for sc.Scan() { |
| 302 | line := sc.Text() |
| 303 | mp, ok := mountinfoLineMountpoint(line) |
| 304 | if !ok || mp != want { |
| 305 | continue |
| 306 | } |
| 307 | verdict = mountinfoLineHas(line, want, opt) |
| 308 | } |
| 309 | return verdict, sc.Err() |
| 310 | } |
| 311 | |
| 312 | // mountinfoLineMountpoint returns field 5 of a mountinfo line (the |
| 313 | // mountpoint, octal-escaped), unescaped, and whether the line has enough |
| 314 | // fields to have one at all. |
| 315 | func mountinfoLineMountpoint(line string) (string, bool) { |
| 316 | fields := strings.Fields(line) |
| 317 | if len(fields) < 6 { |
| 318 | return "", false |
| 319 | } |
| 320 | return unescapeMountinfo(fields[4]), true |
| 321 | } |
| 322 | |
| 323 | // mountinfoLineHas parses one mountinfo line: field 5 is the mountpoint |
| 324 | // (octal-escaped), field 6 the per-mount options. |
| 325 | func mountinfoLineHas(line, mountpoint, option string) bool { |
| 326 | fields := strings.Fields(line) |
| 327 | if len(fields) < 6 { |
| 328 | return false |
| 329 | } |
| 330 | if unescapeMountinfo(fields[4]) != mountpoint { |
| 331 | return false |
| 332 | } |
| 333 | for _, o := range strings.Split(fields[5], ",") { |
| 334 | if o == option { |
| 335 | return true |
| 336 | } |
| 337 | } |
| 338 | return false |
| 339 | } |
| 340 | |
| 341 | // unescapeMountinfo undoes the \040-style escapes mountinfo applies to |
| 342 | // spaces, tabs, newlines and backslashes in paths. |
| 343 | func unescapeMountinfo(s string) string { |
| 344 | if !strings.Contains(s, `\`) { |
| 345 | return s |
| 346 | } |
| 347 | var b strings.Builder |
| 348 | for i := 0; i < len(s); i++ { |
| 349 | if s[i] == '\\' && i+3 < len(s) { |
| 350 | var v int |
| 351 | if _, err := fmt.Sscanf(s[i+1:i+4], "%o", &v); err == nil { |
| 352 | b.WriteByte(byte(v)) |
| 353 | i += 3 |
| 354 | continue |
| 355 | } |
| 356 | } |
| 357 | b.WriteByte(s[i]) |
| 358 | } |
| 359 | return b.String() |
| 360 | } |
| 361 | |
amadan