| 1 | //go:build linux |
| 2 | |
| 3 | package operator |
| 4 | |
| 5 | import ( |
| 6 | "context" |
| 7 | "database/sql" |
| 8 | "database/sql/driver" |
| 9 | "io" |
| 10 | "net/url" |
| 11 | "os" |
| 12 | "path/filepath" |
| 13 | "strconv" |
| 14 | "strings" |
| 15 | "sync" |
| 16 | "time" |
| 17 | |
| 18 | "golang.org/x/sys/unix" |
| 19 | "modernc.org/sqlite" |
| 20 | ) |
| 21 | |
| 22 | type fileIdentity struct{ device, inode uint64 } |
| 23 | type descriptor struct { |
| 24 | identity fileIdentity |
| 25 | flags int |
| 26 | links uint64 |
| 27 | } |
| 28 | type boundDatabase struct { |
| 29 | db *sql.DB |
| 30 | connector *boundConnector |
| 31 | } |
| 32 | type boundConnector struct { |
| 33 | mu sync.Mutex |
| 34 | path string |
| 35 | pin *os.File |
| 36 | identity fileIdentity |
| 37 | deadline time.Time |
| 38 | sqlite *sqlite.Driver |
| 39 | active bool |
| 40 | hook func(string) |
| 41 | } |
| 42 | type boundConnection struct { |
| 43 | driver.Conn |
| 44 | owner *boundConnector |
| 45 | mainFD int |
| 46 | siblings map[string]fileIdentity |
| 47 | closed bool |
| 48 | } |
| 49 | |
| 50 | // newBoundDatabase proves file identity only. The unmounted command adapter must |
| 51 | // first establish the separately verified launch context and signed authority. |
| 52 | func newBoundDatabase(path string, device, inode uint64, deadline time.Time) (*boundDatabase, error) { |
| 53 | if device == 0 || inode == 0 || !deadline.After(time.Now()) { |
| 54 | return nil, ErrRefused |
| 55 | } |
| 56 | pin, id, e := pinExisting(path) |
| 57 | if e != nil { |
| 58 | return nil, e |
| 59 | } |
| 60 | if id != (fileIdentity{device, inode}) { |
| 61 | pin.Close() |
| 62 | return nil, ErrRefused |
| 63 | } |
| 64 | c := &boundConnector{path: path, pin: pin, identity: id, deadline: deadline, sqlite: &sqlite.Driver{}} |
| 65 | db := sql.OpenDB(c) |
| 66 | db.SetMaxOpenConns(1) |
| 67 | db.SetMaxIdleConns(1) |
| 68 | return &boundDatabase{db, c}, nil |
| 69 | } |
| 70 | func (b *boundDatabase) close() error { |
| 71 | e := b.db.Close() |
| 72 | b.connector.mu.Lock() |
| 73 | defer b.connector.mu.Unlock() |
| 74 | if b.connector.active { |
| 75 | return ErrUnavailable |
| 76 | } |
| 77 | if x := b.connector.pin.Close(); e == nil { |
| 78 | e = x |
| 79 | } |
| 80 | return e |
| 81 | } |
| 82 | func pinExisting(path string) (*os.File, fileIdentity, error) { |
| 83 | if !filepath.IsAbs(path) || filepath.Clean(path) != path || path == "/" || len(path) > 1024 || strings.ContainsRune(path, 0) { |
| 84 | return nil, fileIdentity{}, ErrRefused |
| 85 | } |
| 86 | fd, e := unix.Open("/", unix.O_RDONLY|unix.O_DIRECTORY|unix.O_CLOEXEC, 0) |
| 87 | if e != nil { |
| 88 | return nil, fileIdentity{}, ErrUnavailable |
| 89 | } |
| 90 | parts := strings.Split(strings.TrimPrefix(path, "/"), "/") |
| 91 | for i, part := range parts { |
| 92 | flags := unix.O_RDONLY | unix.O_NOFOLLOW | unix.O_CLOEXEC | unix.O_NONBLOCK |
| 93 | if i < len(parts)-1 { |
| 94 | flags |= unix.O_DIRECTORY |
| 95 | } |
| 96 | next, err := unix.Openat(fd, part, flags, 0) |
| 97 | unix.Close(fd) |
| 98 | if err != nil { |
| 99 | return nil, fileIdentity{}, ErrRefused |
| 100 | } |
| 101 | fd = next |
| 102 | } |
| 103 | var s unix.Stat_t |
| 104 | if unix.Fstat(fd, &s) != nil { |
| 105 | unix.Close(fd) |
| 106 | return nil, fileIdentity{}, ErrUnavailable |
| 107 | } |
| 108 | if s.Mode&unix.S_IFMT != unix.S_IFREG || s.Nlink != 1 || s.Dev == 0 || s.Ino == 0 { |
| 109 | unix.Close(fd) |
| 110 | return nil, fileIdentity{}, ErrRefused |
| 111 | } |
| 112 | return os.NewFile(uintptr(fd), path), fileIdentity{uint64(s.Dev), s.Ino}, nil |
| 113 | } |
| 114 | |
| 115 | // statExisting does not open the regular file: closing a second descriptor for |
| 116 | // a SQLite inode would release this process's POSIX locks on that inode. |
| 117 | func statExisting(path string) (fileIdentity, error) { |
| 118 | if !filepath.IsAbs(path) || filepath.Clean(path) != path || path == "/" || len(path) > 1024 || strings.ContainsRune(path, 0) { |
| 119 | return fileIdentity{}, ErrRefused |
| 120 | } |
| 121 | fd, e := unix.Open("/", unix.O_RDONLY|unix.O_DIRECTORY|unix.O_CLOEXEC, 0) |
| 122 | if e != nil { |
| 123 | return fileIdentity{}, ErrUnavailable |
| 124 | } |
| 125 | parts := strings.Split(strings.TrimPrefix(path, "/"), "/") |
| 126 | for _, part := range parts[:len(parts)-1] { |
| 127 | next, err := unix.Openat(fd, part, unix.O_RDONLY|unix.O_DIRECTORY|unix.O_NOFOLLOW|unix.O_CLOEXEC, 0) |
| 128 | unix.Close(fd) |
| 129 | if err != nil { |
| 130 | return fileIdentity{}, ErrRefused |
| 131 | } |
| 132 | fd = next |
| 133 | } |
| 134 | defer unix.Close(fd) |
| 135 | var s unix.Stat_t |
| 136 | if unix.Fstatat(fd, parts[len(parts)-1], &s, unix.AT_SYMLINK_NOFOLLOW) != nil { |
| 137 | return fileIdentity{}, ErrRefused |
| 138 | } |
| 139 | if s.Mode&unix.S_IFMT != unix.S_IFREG || s.Nlink != 1 || s.Dev == 0 || s.Ino == 0 { |
| 140 | return fileIdentity{}, ErrRefused |
| 141 | } |
| 142 | return fileIdentity{uint64(s.Dev), s.Ino}, nil |
| 143 | } |
| 144 | func descriptors() (map[int]descriptor, error) { |
| 145 | directory, e := os.Open("/proc/self/fd") |
| 146 | if e != nil { |
| 147 | return nil, ErrUnavailable |
| 148 | } |
| 149 | defer directory.Close() |
| 150 | entries, e := directory.Readdirnames(-1) |
| 151 | if e != nil { |
| 152 | return nil, ErrUnavailable |
| 153 | } |
| 154 | out := map[int]descriptor{} |
| 155 | for _, entry := range entries { |
| 156 | fd, e := strconv.Atoi(entry) |
| 157 | if e != nil { |
| 158 | return nil, ErrRefused |
| 159 | } |
| 160 | var s unix.Stat_t |
| 161 | if fd == int(directory.Fd()) { |
| 162 | continue |
| 163 | } |
| 164 | if e = unix.Fstat(fd, &s); e != nil { |
| 165 | return nil, ErrRefused |
| 166 | } |
| 167 | if s.Mode&unix.S_IFMT != unix.S_IFREG { |
| 168 | continue |
| 169 | } |
| 170 | flags, e := unix.FcntlInt(uintptr(fd), unix.F_GETFL, 0) |
| 171 | if e != nil { |
| 172 | return nil, ErrUnavailable |
| 173 | } |
| 174 | out[fd] = descriptor{fileIdentity{uint64(s.Dev), s.Ino}, flags, uint64(s.Nlink)} |
| 175 | } |
| 176 | return out, nil |
| 177 | } |
| 178 | func writable(flags int) bool { return flags&unix.O_ACCMODE != unix.O_RDONLY } |
| 179 | func (c *boundConnector) Driver() driver.Driver { return c.sqlite } |
| 180 | func (c *boundConnector) checkpoint(stage string) { |
| 181 | if c.hook != nil { |
| 182 | c.hook(stage) |
| 183 | } |
| 184 | } |
| 185 | func (c *boundConnector) pathMatches() error { |
| 186 | if !time.Now().Before(c.deadline) { |
| 187 | return ErrRefused |
| 188 | } |
| 189 | var stat unix.Stat_t |
| 190 | if unix.Fstat(int(c.pin.Fd()), &stat) != nil || (fileIdentity{uint64(stat.Dev), stat.Ino}) != c.identity || stat.Nlink != 1 { |
| 191 | return ErrRefused |
| 192 | } |
| 193 | id, e := statExisting(c.path) |
| 194 | if e != nil { |
| 195 | return e |
| 196 | } |
| 197 | if id != c.identity { |
| 198 | return ErrRefused |
| 199 | } |
| 200 | return nil |
| 201 | } |
| 202 | func (c *boundConnector) Connect(ctx context.Context) (driver.Conn, error) { |
| 203 | c.mu.Lock() |
| 204 | defer c.mu.Unlock() |
| 205 | if c.active || ctx.Err() != nil || c.pathMatches() != nil { |
| 206 | return nil, ErrRefused |
| 207 | } |
| 208 | c.checkpoint("before-inventory") |
| 209 | baseline, e := descriptors() |
| 210 | if e != nil { |
| 211 | return nil, e |
| 212 | } |
| 213 | for _, d := range baseline { |
| 214 | if writable(d.flags) { |
| 215 | return nil, ErrRefused |
| 216 | } |
| 217 | } |
| 218 | c.checkpoint("before-open") |
| 219 | if c.pathMatches() != nil { |
| 220 | return nil, ErrRefused |
| 221 | } |
| 222 | c.checkpoint("before-driver") |
| 223 | u := url.URL{Scheme: "file", Path: c.path} |
| 224 | q := u.Query() |
| 225 | q.Set("mode", "rw") |
| 226 | u.RawQuery = q.Encode() |
| 227 | native, e := c.sqlite.Open(u.String()) |
| 228 | if e != nil { |
| 229 | return nil, ErrUnavailable |
| 230 | } |
| 231 | keep := false |
| 232 | defer func() { |
| 233 | if !keep { |
| 234 | native.Close() |
| 235 | } |
| 236 | }() |
| 237 | c.checkpoint("after-open") |
| 238 | current, e := descriptors() |
| 239 | if e != nil { |
| 240 | return nil, e |
| 241 | } |
| 242 | mainFD := -1 |
| 243 | for fd, d := range current { |
| 244 | if !writable(d.flags) { |
| 245 | continue |
| 246 | } |
| 247 | if _, old := baseline[fd]; old { |
| 248 | return nil, ErrRefused |
| 249 | } |
| 250 | if d.identity == c.identity { |
| 251 | if d.flags&unix.O_ACCMODE != unix.O_RDWR || d.links != 1 || mainFD != -1 { |
| 252 | return nil, ErrRefused |
| 253 | } |
| 254 | mainFD = fd |
| 255 | } |
| 256 | } |
| 257 | if mainFD < 0 { |
| 258 | return nil, ErrRefused |
| 259 | } |
| 260 | bound := &boundConnection{Conn: native, owner: c, mainFD: mainFD, siblings: map[string]fileIdentity{}} |
| 261 | if e = bound.check(); e != nil { |
| 262 | return nil, e |
| 263 | } |
| 264 | query, ok := native.(driver.QueryerContext) |
| 265 | if !ok { |
| 266 | return nil, ErrRefused |
| 267 | } |
| 268 | rows, e := query.QueryContext(ctx, "PRAGMA database_list", nil) |
| 269 | if e != nil { |
| 270 | return nil, ErrUnavailable |
| 271 | } |
| 272 | values := make([]driver.Value, 3) |
| 273 | if rows.Next(values) != nil || values[1] != "main" || values[2] != c.path { |
| 274 | rows.Close() |
| 275 | return nil, ErrRefused |
| 276 | } |
| 277 | if e = rows.Next(values); e != io.EOF { |
| 278 | rows.Close() |
| 279 | return nil, ErrRefused |
| 280 | } |
| 281 | if rows.Close() != nil { |
| 282 | return nil, ErrUnavailable |
| 283 | } |
| 284 | c.checkpoint("after-metadata") |
| 285 | if e = bound.check(); e != nil { |
| 286 | return nil, e |
| 287 | } |
| 288 | c.active = true |
| 289 | keep = true |
| 290 | return bound, nil |
| 291 | } |
| 292 | func (c *boundConnection) check() error { |
| 293 | if c.closed || c.owner.pathMatches() != nil { |
| 294 | return ErrRefused |
| 295 | } |
| 296 | all, e := descriptors() |
| 297 | if e != nil { |
| 298 | return e |
| 299 | } |
| 300 | main, ok := all[c.mainFD] |
| 301 | if !ok || main.identity != c.owner.identity || main.flags&unix.O_ACCMODE != unix.O_RDWR || main.links != 1 { |
| 302 | return ErrRefused |
| 303 | } |
| 304 | count := 0 |
| 305 | seenSiblings := map[string]bool{} |
| 306 | for fd, d := range all { |
| 307 | if !writable(d.flags) { |
| 308 | continue |
| 309 | } |
| 310 | if fd == c.mainFD { |
| 311 | count++ |
| 312 | continue |
| 313 | } |
| 314 | if d.flags&unix.O_ACCMODE != unix.O_RDWR || d.links != 1 { |
| 315 | return ErrRefused |
| 316 | } |
| 317 | matched := false |
| 318 | for _, suffix := range []string{"-wal", "-shm"} { |
| 319 | id, e := statExisting(c.owner.path + suffix) |
| 320 | if e != nil { |
| 321 | continue |
| 322 | } |
| 323 | if id == d.identity { |
| 324 | if seenSiblings[suffix] { |
| 325 | return ErrRefused |
| 326 | } |
| 327 | seenSiblings[suffix] = true |
| 328 | if old, exists := c.siblings[suffix]; exists && old != id { |
| 329 | return ErrRefused |
| 330 | } |
| 331 | c.siblings[suffix] = id |
| 332 | matched = true |
| 333 | break |
| 334 | } |
| 335 | } |
| 336 | if !matched { |
| 337 | return ErrRefused |
| 338 | } |
| 339 | } |
| 340 | if count != 1 { |
| 341 | return ErrRefused |
| 342 | } |
| 343 | return nil |
| 344 | } |
| 345 | func (c *boundConnection) Close() error { |
| 346 | c.owner.mu.Lock() |
| 347 | defer c.owner.mu.Unlock() |
| 348 | if c.closed { |
| 349 | return nil |
| 350 | } |
| 351 | c.closed = true |
| 352 | e := c.Conn.Close() |
| 353 | c.owner.active = false |
| 354 | return e |
| 355 | } |
| 356 | func (c *boundConnection) Prepare(string) (driver.Stmt, error) { return nil, ErrRefused } |
| 357 | func (c *boundConnection) PrepareContext(context.Context, string) (driver.Stmt, error) { |
| 358 | return nil, ErrRefused |
| 359 | } |
| 360 | func (c *boundConnection) Begin() (driver.Tx, error) { |
| 361 | return c.BeginTx(context.Background(), driver.TxOptions{}) |
| 362 | } |
| 363 | func (c *boundConnection) BeginTx(ctx context.Context, opts driver.TxOptions) (driver.Tx, error) { |
| 364 | if c.check() != nil { |
| 365 | return nil, ErrRefused |
| 366 | } |
| 367 | x, ok := c.Conn.(driver.ConnBeginTx) |
| 368 | if !ok { |
| 369 | return nil, ErrRefused |
| 370 | } |
| 371 | tx, e := x.BeginTx(ctx, opts) |
| 372 | if e != nil { |
| 373 | return nil, e |
| 374 | } |
| 375 | if c.check() != nil { |
| 376 | tx.Rollback() |
| 377 | return nil, ErrUnavailable |
| 378 | } |
| 379 | return &boundTransaction{tx, c}, nil |
| 380 | } |
| 381 | func (c *boundConnection) ExecContext(ctx context.Context, query string, args []driver.NamedValue) (driver.Result, error) { |
| 382 | if query != "ROLLBACK" && c.check() != nil { |
| 383 | return nil, ErrRefused |
| 384 | } |
| 385 | x, ok := c.Conn.(driver.ExecerContext) |
| 386 | if !ok { |
| 387 | return nil, ErrRefused |
| 388 | } |
| 389 | result, e := x.ExecContext(ctx, query, args) |
| 390 | if e != nil { |
| 391 | return nil, e |
| 392 | } |
| 393 | c.owner.checkpoint("after-exec") |
| 394 | if query != "ROLLBACK" && c.check() != nil { |
| 395 | return nil, ErrUnavailable |
| 396 | } |
| 397 | return result, nil |
| 398 | } |
| 399 | func (c *boundConnection) QueryContext(ctx context.Context, query string, args []driver.NamedValue) (driver.Rows, error) { |
| 400 | if c.check() != nil { |
| 401 | return nil, ErrRefused |
| 402 | } |
| 403 | x, ok := c.Conn.(driver.QueryerContext) |
| 404 | if !ok { |
| 405 | return nil, ErrRefused |
| 406 | } |
| 407 | rows, e := x.QueryContext(ctx, query, args) |
| 408 | if e != nil { |
| 409 | return nil, e |
| 410 | } |
| 411 | if c.check() != nil { |
| 412 | rows.Close() |
| 413 | return nil, ErrUnavailable |
| 414 | } |
| 415 | return &boundRows{rows, c}, nil |
| 416 | } |
| 417 | func (c *boundConnection) Ping(ctx context.Context) error { |
| 418 | if c.check() != nil { |
| 419 | return ErrRefused |
| 420 | } |
| 421 | p, ok := c.Conn.(driver.Pinger) |
| 422 | if !ok { |
| 423 | return ErrRefused |
| 424 | } |
| 425 | if e := p.Ping(ctx); e != nil { |
| 426 | return e |
| 427 | } |
| 428 | if c.check() != nil { |
| 429 | return ErrUnavailable |
| 430 | } |
| 431 | return nil |
| 432 | } |
| 433 | func (c *boundConnection) ResetSession(ctx context.Context) error { |
| 434 | if c.check() != nil { |
| 435 | return ErrRefused |
| 436 | } |
| 437 | if r, ok := c.Conn.(driver.SessionResetter); ok { |
| 438 | return r.ResetSession(ctx) |
| 439 | } |
| 440 | return nil |
| 441 | } |
| 442 | func (c *boundConnection) IsValid() bool { return c.check() == nil } |
| 443 | |
| 444 | type boundRows struct { |
| 445 | driver.Rows |
| 446 | c *boundConnection |
| 447 | } |
| 448 | |
| 449 | func (r *boundRows) Next(values []driver.Value) error { |
| 450 | if r.c.check() != nil { |
| 451 | return ErrRefused |
| 452 | } |
| 453 | e := r.Rows.Next(values) |
| 454 | if e != nil { |
| 455 | return e |
| 456 | } |
| 457 | if r.c.check() != nil { |
| 458 | return ErrUnavailable |
| 459 | } |
| 460 | return nil |
| 461 | } |
| 462 | |
| 463 | type boundTransaction struct { |
| 464 | driver.Tx |
| 465 | c *boundConnection |
| 466 | } |
| 467 | |
| 468 | func (t *boundTransaction) Commit() error { |
| 469 | if t.c.check() != nil { |
| 470 | return ErrRefused |
| 471 | } |
| 472 | e := t.Tx.Commit() |
| 473 | t.c.owner.checkpoint("after-commit") |
| 474 | if e != nil { |
| 475 | return e |
| 476 | } |
| 477 | if t.c.check() != nil { |
| 478 | return ErrUnavailable |
| 479 | } |
| 480 | return nil |
| 481 | } |
| 482 | |