File
Blob: src/workerd/util/sqlite.c++
| 1 | // Copyright (c) 2023 Cloudflare, Inc. |
| 2 | // Licensed under the Apache 2.0 license found in the LICENSE file or at: |
| 3 | // https://opensource.org/licenses/Apache-2.0 |
| 4 | |
| 5 | #include "sqlite.h" |
| 6 | |
| 7 | #include <workerd/util/autogate.h> |
| 8 | #include <workerd/util/sentry.h> |
| 9 | |
| 10 | #include <kj/debug.h> |
| 11 | #include <kj/refcount.h> |
| 12 | #include <kj/string-tree.h> |
| 13 | |
| 14 | #if _WIN32 |
| 15 | #include <windows.h> |
| 16 | |
| 17 | #include <kj/win32-api-version.h> |
| 18 | #include <kj/windows-sanity.h> |
| 19 | #else |
| 20 | #include <unistd.h> |
| 21 | #endif |
| 22 | |
| 23 | #include <fcntl.h> |
| 24 | #include <sqlite3.h> |
| 25 | #include <sys/stat.h> |
| 26 | |
| 27 | #include <kj/map.h> |
| 28 | #include <kj/mutex.h> |
| 29 | #include <kj/vector.h> |
| 30 | |
| 31 | #include <atomic> |
| 32 | |
| 33 | #if _WIN32 |
| 34 | #define strncasecmp _strnicmp |
| 35 | #define strcasecmp _stricmp |
| 36 | #endif |
| 37 | |
| 38 | namespace workerd { |
| 39 | |
| 40 | namespace { |
| 41 | |
| 42 | // SQLite has a function like this in its internals, but it's not exposed to library consumers. |
| 43 | // |
| 44 | // These error codes come from https://www.sqlite.org/rescode.html#primary_result_code_list. |
| 45 | kj::String namedErrorCode(int errorCode) { |
| 46 | #define LITERAL(name) \ |
| 47 | case name: \ |
| 48 | return kj::str(#name); |
| 49 | switch (errorCode) { |
| 50 | LITERAL(SQLITE_OK) |
| 51 | LITERAL(SQLITE_ERROR) |
| 52 | LITERAL(SQLITE_INTERNAL) |
| 53 | LITERAL(SQLITE_PERM) |
| 54 | LITERAL(SQLITE_ABORT) |
| 55 | LITERAL(SQLITE_BUSY) |
| 56 | LITERAL(SQLITE_LOCKED) |
| 57 | LITERAL(SQLITE_NOMEM) |
| 58 | LITERAL(SQLITE_READONLY) |
| 59 | LITERAL(SQLITE_INTERRUPT) |
| 60 | LITERAL(SQLITE_IOERR) |
| 61 | LITERAL(SQLITE_CORRUPT) |
| 62 | LITERAL(SQLITE_NOTFOUND) |
| 63 | LITERAL(SQLITE_FULL) |
| 64 | LITERAL(SQLITE_CANTOPEN) |
| 65 | LITERAL(SQLITE_PROTOCOL) |
| 66 | LITERAL(SQLITE_EMPTY) |
| 67 | LITERAL(SQLITE_SCHEMA) |
| 68 | LITERAL(SQLITE_TOOBIG) |
| 69 | LITERAL(SQLITE_CONSTRAINT) |
| 70 | LITERAL(SQLITE_MISMATCH) |
| 71 | LITERAL(SQLITE_MISUSE) |
| 72 | LITERAL(SQLITE_NOLFS) |
| 73 | LITERAL(SQLITE_AUTH) |
| 74 | LITERAL(SQLITE_FORMAT) |
| 75 | LITERAL(SQLITE_RANGE) |
| 76 | LITERAL(SQLITE_NOTADB) |
| 77 | LITERAL(SQLITE_NOTICE) |
| 78 | LITERAL(SQLITE_WARNING) |
| 79 | LITERAL(SQLITE_ROW) |
| 80 | LITERAL(SQLITE_DONE) |
| 81 | default: |
| 82 | return kj::str("SQLITE_UNKNOWN_ERROR_CODE(", errorCode, ")"); |
| 83 | } |
| 84 | #undef LITERAL |
| 85 | } |
| 86 | |
| 87 | // Maps extended error codes to their symbolic names. |
| 88 | // See https://www.sqlite.org/rescode.html#extended_result_code_list. |
| 89 | kj::Maybe<kj::String> namedExtendedErrorCode(int extendedErrorCode) { |
| 90 | #define LITERAL(name) \ |
| 91 | case name: \ |
| 92 | return kj::str(#name); |
| 93 | switch (extendedErrorCode) { |
| 94 | LITERAL(SQLITE_ABORT_ROLLBACK) |
| 95 | LITERAL(SQLITE_AUTH_USER) |
| 96 | LITERAL(SQLITE_BUSY_RECOVERY) |
| 97 | LITERAL(SQLITE_BUSY_SNAPSHOT) |
| 98 | LITERAL(SQLITE_BUSY_TIMEOUT) |
| 99 | LITERAL(SQLITE_CANTOPEN_CONVPATH) |
| 100 | LITERAL(SQLITE_CANTOPEN_DIRTYWAL) |
| 101 | LITERAL(SQLITE_CANTOPEN_FULLPATH) |
| 102 | LITERAL(SQLITE_CANTOPEN_ISDIR) |
| 103 | LITERAL(SQLITE_CANTOPEN_NOTEMPDIR) |
| 104 | LITERAL(SQLITE_CANTOPEN_SYMLINK) |
| 105 | LITERAL(SQLITE_CONSTRAINT_CHECK) |
| 106 | LITERAL(SQLITE_CONSTRAINT_COMMITHOOK) |
| 107 | LITERAL(SQLITE_CONSTRAINT_DATATYPE) |
| 108 | LITERAL(SQLITE_CONSTRAINT_FOREIGNKEY) |
| 109 | LITERAL(SQLITE_CONSTRAINT_FUNCTION) |
| 110 | LITERAL(SQLITE_CONSTRAINT_NOTNULL) |
| 111 | LITERAL(SQLITE_CONSTRAINT_PINNED) |
| 112 | LITERAL(SQLITE_CONSTRAINT_PRIMARYKEY) |
| 113 | LITERAL(SQLITE_CONSTRAINT_ROWID) |
| 114 | LITERAL(SQLITE_CONSTRAINT_TRIGGER) |
| 115 | LITERAL(SQLITE_CONSTRAINT_UNIQUE) |
| 116 | LITERAL(SQLITE_CONSTRAINT_VTAB) |
| 117 | LITERAL(SQLITE_CORRUPT_INDEX) |
| 118 | LITERAL(SQLITE_CORRUPT_SEQUENCE) |
| 119 | LITERAL(SQLITE_CORRUPT_VTAB) |
| 120 | LITERAL(SQLITE_ERROR_MISSING_COLLSEQ) |
| 121 | LITERAL(SQLITE_ERROR_RETRY) |
| 122 | LITERAL(SQLITE_ERROR_SNAPSHOT) |
| 123 | LITERAL(SQLITE_IOERR_ACCESS) |
| 124 | LITERAL(SQLITE_IOERR_AUTH) |
| 125 | LITERAL(SQLITE_IOERR_BEGIN_ATOMIC) |
| 126 | LITERAL(SQLITE_IOERR_BLOCKED) |
| 127 | LITERAL(SQLITE_IOERR_CHECKRESERVEDLOCK) |
| 128 | LITERAL(SQLITE_IOERR_CLOSE) |
| 129 | LITERAL(SQLITE_IOERR_COMMIT_ATOMIC) |
| 130 | LITERAL(SQLITE_IOERR_CONVPATH) |
| 131 | LITERAL(SQLITE_IOERR_CORRUPTFS) |
| 132 | LITERAL(SQLITE_IOERR_DATA) |
| 133 | LITERAL(SQLITE_IOERR_DELETE) |
| 134 | LITERAL(SQLITE_IOERR_DELETE_NOENT) |
| 135 | LITERAL(SQLITE_IOERR_DIR_CLOSE) |
| 136 | LITERAL(SQLITE_IOERR_DIR_FSYNC) |
| 137 | LITERAL(SQLITE_IOERR_FSTAT) |
| 138 | LITERAL(SQLITE_IOERR_FSYNC) |
| 139 | LITERAL(SQLITE_IOERR_GETTEMPPATH) |
| 140 | LITERAL(SQLITE_IOERR_LOCK) |
| 141 | LITERAL(SQLITE_IOERR_MMAP) |
| 142 | LITERAL(SQLITE_IOERR_NOMEM) |
| 143 | LITERAL(SQLITE_IOERR_RDLOCK) |
| 144 | LITERAL(SQLITE_IOERR_READ) |
| 145 | LITERAL(SQLITE_IOERR_ROLLBACK_ATOMIC) |
| 146 | LITERAL(SQLITE_IOERR_SEEK) |
| 147 | LITERAL(SQLITE_IOERR_SHMLOCK) |
| 148 | LITERAL(SQLITE_IOERR_SHMMAP) |
| 149 | LITERAL(SQLITE_IOERR_SHMOPEN) |
| 150 | LITERAL(SQLITE_IOERR_SHMSIZE) |
| 151 | LITERAL(SQLITE_IOERR_SHORT_READ) |
| 152 | LITERAL(SQLITE_IOERR_TRUNCATE) |
| 153 | LITERAL(SQLITE_IOERR_UNLOCK) |
| 154 | LITERAL(SQLITE_IOERR_VNODE) |
| 155 | LITERAL(SQLITE_IOERR_WRITE) |
| 156 | LITERAL(SQLITE_LOCKED_SHAREDCACHE) |
| 157 | LITERAL(SQLITE_LOCKED_VTAB) |
| 158 | LITERAL(SQLITE_NOTICE_RECOVER_ROLLBACK) |
| 159 | LITERAL(SQLITE_NOTICE_RECOVER_WAL) |
| 160 | LITERAL(SQLITE_OK_LOAD_PERMANENTLY) |
| 161 | LITERAL(SQLITE_READONLY_CANTINIT) |
| 162 | LITERAL(SQLITE_READONLY_CANTLOCK) |
| 163 | LITERAL(SQLITE_READONLY_DBMOVED) |
| 164 | LITERAL(SQLITE_READONLY_DIRECTORY) |
| 165 | LITERAL(SQLITE_READONLY_RECOVERY) |
| 166 | LITERAL(SQLITE_READONLY_ROLLBACK) |
| 167 | LITERAL(SQLITE_WARNING_AUTOINDEX) |
| 168 | default: |
| 169 | return kj::none; |
| 170 | } |
| 171 | #undef LITERAL |
| 172 | } |
| 173 | |
| 174 | constexpr size_t RA_MAX_METRICS_QUERY_SIZE = 1024; |
| 175 | |
| 176 | kj::String dbErrorMessage(int errorCode, sqlite3* db) { |
| 177 | kj::StringTree msg = kj::strTree(sqlite3_errmsg(db)); |
| 178 | if (int offset = sqlite3_error_offset(db); offset != -1) { |
| 179 | msg = kj::strTree(kj::mv(msg), " at offset ", offset); |
| 180 | } |
| 181 | msg = kj::strTree(kj::mv(msg), ": ", namedErrorCode(errorCode)); |
| 182 | int extendedCode = sqlite3_extended_errcode(db); |
| 183 | if (extendedCode != errorCode) { |
| 184 | KJ_IF_SOME(extendedName, namedExtendedErrorCode(extendedCode)) { |
| 185 | msg = kj::strTree(kj::mv(msg), " (extended: ", extendedName, ")"); |
| 186 | } |
| 187 | } |
| 188 | return msg.flatten(); |
| 189 | } |
| 190 | |
| 191 | // If a VFS call throws an exception, and vfsErrorListener is non-null, the exception will |
| 192 | // be placed there, otherwise it will be logged. This is used to implement pass-through of KJ |
| 193 | // exceptions through SQLite. |
| 194 | static thread_local kj::Maybe<kj::Exception>* vfsErrorListener = nullptr; |
| 195 | |
| 196 | // Report that in a sqlite VFS callback, an exception was caught, and SQLITE_IOERROR is being |
| 197 | // returned to SQLite. |
| 198 | // |
| 199 | // The exception must be caught using `catch (kj::Exception& e)`, NOT using `catch (...)` followed |
| 200 | // by kj::getCaughtExceptionAsKj(). This is because the latter truncates the stack trace to show |
| 201 | // only the frames between the throw and the catch. We actually want to retain the full trace |
| 202 | // through SQLite. |
| 203 | void reportVfsErrorCaught(kj::Exception&& e) { |
| 204 | if (vfsErrorListener != nullptr) { |
| 205 | // Only capture the first error; assume subsequent errors are side effects. |
| 206 | if (*vfsErrorListener == kj::none) { |
| 207 | *vfsErrorListener = kj::mv(e); |
| 208 | } |
| 209 | } else { |
| 210 | LOG_EXCEPTION("sqliteVfsError", e); |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | // Implements SQLITE_CALL_SCOPE. |
| 215 | class SqliteCallScope { |
| 216 | public: |
| 217 | SqliteCallScope() { |
| 218 | KJ_DASSERT(vfsErrorListener == nullptr); |
| 219 | vfsErrorListener = &error; |
| 220 | } |
| 221 | ~SqliteCallScope() { |
| 222 | vfsErrorListener = nullptr; |
| 223 | } |
| 224 | |
| 225 | void rethrowVfsError() { |
| 226 | KJ_IF_SOME(e, error) { |
| 227 | // Slight hack: The exception already has a stack trace attached which should include the |
| 228 | // current stack, but `kj::throwFatalException()` would re-append the current stack trace |
| 229 | // to the exception. We can avoid that by calling |
| 230 | // kj::getExceptionCallback().onFatalException() directly, which is what |
| 231 | // `throwFatalException()` does after extending the stack. |
| 232 | kj::getExceptionCallback().onFatalException(kj::mv(e)); |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | kj::Maybe<const kj::Exception&> getException() { |
| 237 | return error; |
| 238 | } |
| 239 | |
| 240 | // Hack to allow block syntax with for(); see SQLITE_CALL_SCOPE. |
| 241 | bool done = false; |
| 242 | |
| 243 | private: |
| 244 | kj::Maybe<kj::Exception> error; |
| 245 | }; |
| 246 | |
| 247 | } // namespace |
| 248 | |
| 249 | // Like KJ_REQUIRE() but give the Regulator a chance to report the error. `errorMessage` is either |
| 250 | // the return value of sqlite3_errmsg() or a string literal containing a similarly |
| 251 | // application-approriate error message. A reference called `regulator` must be in-scope. |
| 252 | // sqliteErrorCode is a kj::Maybe<int> and represents the error code from sqlite. |
| 253 | #define SQLITE_REQUIRE(condition, sqliteErrorCode, errorMessage, ...) \ |
| 254 | if (!(condition)) { \ |
| 255 | regulator.onError(sqliteErrorCode, errorMessage); \ |
| 256 | KJ_FAIL_REQUIRE("SENTRY_DO SQLite failed", errorMessage, ##__VA_ARGS__); \ |
| 257 | } |
| 258 | |
| 259 | // Make a SQLite call and check the returned error code. Use this version when the call is not |
| 260 | // associated with an open DB connection. |
| 261 | #define SQLITE_CALL_NODB(code, ...) \ |
| 262 | do { \ |
| 263 | int _ec = code; \ |
| 264 | KJ_ASSERT( \ |
| 265 | _ec == SQLITE_OK, kj::str(sqlite3_errstr(_ec), ": ", namedErrorCode(_ec)), ##__VA_ARGS__); \ |
| 266 | } while (false) |
| 267 | |
| 268 | // This version requires the scope to contain a variable named `db` which is of type sqlite3*, or |
| 269 | // can convert to it. |
| 270 | #define SQLITE_CALL(code, ...) \ |
| 271 | do { \ |
| 272 | SqliteCallScope sqliteCallScope; \ |
| 273 | int _ec = code; \ |
| 274 | /* SQLITE_MISUSE doesn't put error info on the database object, so check it separately */ \ |
| 275 | KJ_ASSERT(_ec != SQLITE_MISUSE, "SQLite misused: " #code, ##__VA_ARGS__); \ |
| 276 | handleCriticalError(_ec, dbErrorMessage(_ec, db), sqliteCallScope.getException()); \ |
| 277 | if (_ec == SQLITE_IOERR) sqliteCallScope.rethrowVfsError(); \ |
| 278 | SQLITE_REQUIRE(_ec == SQLITE_OK, _ec, dbErrorMessage(_ec, db), ##__VA_ARGS__); \ |
| 279 | } while (false) |
| 280 | |
| 281 | // Version of `SQLITE_CALL` that can be called after inspecting the error code, in case some codes |
| 282 | // aren't really errors. |
| 283 | // |
| 284 | // Temporarily marking SQLITE_BUSY as NOSENTRY to reduce sentry volume while debugging issue. |
| 285 | // TODO(soon): reenable SQLITE_BUSY sentry logging. |
| 286 | #define SQLITE_CALL_FAILED(code, error, ...) \ |
| 287 | do { \ |
| 288 | KJ_ASSERT(error != SQLITE_MISUSE, "SQLite misused: " code, ##__VA_ARGS__); \ |
| 289 | handleCriticalError(error, dbErrorMessage(error, db), sqliteCallScope.getException()); \ |
| 290 | if (error == SQLITE_IOERR) sqliteCallScope.rethrowVfsError(); \ |
| 291 | SQLITE_REQUIRE(error != SQLITE_BUSY, error, kj::str("NOSENTRY ", dbErrorMessage(error, db)), \ |
| 292 | ##__VA_ARGS__); \ |
| 293 | SQLITE_REQUIRE(error == SQLITE_OK, error, dbErrorMessage(error, db), ##__VA_ARGS__); \ |
| 294 | } while (false); |
| 295 | |
| 296 | // When using SQLITE_CALL_FAILED(), you must place the actual sqlite call and the |
| 297 | // SQLITE_CALL_FAILED() invocation within a SQLITE_CALL_SCOPE block, in order to set up VFS error |
| 298 | // capture. Example: |
| 299 | // |
| 300 | // SQLITE_CALL_SCOPE { |
| 301 | // int errorCode = sqlite3_do_something(); |
| 302 | // if (errorCode != SQLITE_OK) { |
| 303 | // SQLITE_CALL_FAILED("sqlite3_do_something()", errorCode, "failed to do something"); |
| 304 | // } |
| 305 | // } |
| 306 | // |
| 307 | // Note that if you use SQLITE_CALL(), this is handled automatically. |
| 308 | #define SQLITE_CALL_SCOPE \ |
| 309 | for (SqliteCallScope sqliteCallScope; !sqliteCallScope.done; sqliteCallScope.done = true) |
| 310 | |
| 311 | namespace { |
| 312 | |
| 313 | void disposeSqlite(sqlite3_stmt* stmt) { |
| 314 | sqlite3_finalize(stmt); |
| 315 | |
| 316 | // Note that any returned error code is actually the last error to occur while executing the |
| 317 | // statement. This does not really mean that finalization failed, and the error in question |
| 318 | // should have been checked and reported earlier. So, we ignore it here. |
| 319 | } |
| 320 | |
| 321 | template <typename T> |
| 322 | class SqliteDisposer: public kj::Disposer { |
| 323 | public: |
| 324 | void disposeImpl(void* pointer) const override { |
| 325 | disposeSqlite(reinterpret_cast<T*>(pointer)); |
| 326 | } |
| 327 | }; |
| 328 | |
| 329 | template <typename T> |
| 330 | kj::Own<T> ownSqlite(T* obj) { |
| 331 | static const SqliteDisposer<T> disposer; |
| 332 | return kj::Own<T>(obj, disposer); |
| 333 | } |
| 334 | |
| 335 | #if _WIN32 |
| 336 | // https://github.com/capnproto/capnproto/blob/master/c%2B%2B/src/kj/filesystem-disk-win32.c%2B%2B#L255-L269 |
| 337 | static kj::Path getPathFromWin32Handle(HANDLE handle) { |
| 338 | DWORD tryLen = MAX_PATH; |
| 339 | for (;;) { |
| 340 | auto temp = kj::heapArray<wchar_t>(tryLen + 1); |
| 341 | DWORD len = GetFinalPathNameByHandleW(handle, temp.begin(), tryLen, 0); |
| 342 | if (len == 0) { |
| 343 | KJ_FAIL_WIN32("GetFinalPathNameByHandleW", GetLastError()); |
| 344 | } |
| 345 | if (len < temp.size()) { |
| 346 | return kj::Path::parseWin32Api(temp.first(len)); |
| 347 | } |
| 348 | // Try again with new length. |
| 349 | tryLen = len; |
| 350 | } |
| 351 | } |
| 352 | #endif |
| 353 | |
| 354 | kj::Maybe<kj::StringPtr> toMaybeString(const char* cstr) { |
| 355 | if (cstr == nullptr) { |
| 356 | return kj::none; |
| 357 | } else { |
| 358 | return kj::StringPtr(cstr); |
| 359 | } |
| 360 | } |
| 361 | |
| 362 | // We allowlist these SQLite functions. |
| 363 | static constexpr kj::StringPtr ALLOWED_SQLITE_FUNCTIONS[] = { |
| 364 | // https://www.sqlite.org/lang_corefunc.html |
| 365 | "abs"_kj, |
| 366 | "changes"_kj, |
| 367 | "char"_kj, |
| 368 | "coalesce"_kj, |
| 369 | "concat"_kj, |
| 370 | "concat_ws"_kj, |
| 371 | "format"_kj, |
| 372 | "glob"_kj, |
| 373 | "hex"_kj, |
| 374 | "ifnull"_kj, |
| 375 | "iif"_kj, |
| 376 | "instr"_kj, |
| 377 | "last_insert_rowid"_kj, |
| 378 | "length"_kj, |
| 379 | "like"_kj, |
| 380 | "likelihood"_kj, |
| 381 | "likely"_kj, |
| 382 | "load_extension"_kj, |
| 383 | "lower"_kj, |
| 384 | "ltrim"_kj, |
| 385 | "max_scalar"_kj, |
| 386 | "min_scalar"_kj, |
| 387 | "nullif"_kj, |
| 388 | "octet_length"_kj, |
| 389 | "printf"_kj, |
| 390 | "quote"_kj, |
| 391 | "random"_kj, |
| 392 | "randomblob"_kj, |
| 393 | "replace"_kj, |
| 394 | "round"_kj, |
| 395 | "rtrim"_kj, |
| 396 | "sign"_kj, |
| 397 | "soundex"_kj, |
| 398 | // These functions query SQLite internals and build details in a way we'd prefer not to reveal. |
| 399 | // "sqlite_compileoption_get"_kj, |
| 400 | // "sqlite_compileoption_used"_kj, |
| 401 | // "sqlite_offset"_kj, |
| 402 | // "sqlite_source_id"_kj, |
| 403 | // "sqlite_version"_kj, |
| 404 | "substr"_kj, |
| 405 | "substring"_kj, |
| 406 | "total_changes"_kj, |
| 407 | "trim"_kj, |
| 408 | "typeof"_kj, |
| 409 | "unhex"_kj, |
| 410 | "unicode"_kj, |
| 411 | "unlikely"_kj, |
| 412 | "upper"_kj, |
| 413 | "zeroblob"_kj, |
| 414 | |
| 415 | // https://www.sqlite.org/lang_datefunc.html |
| 416 | "date"_kj, |
| 417 | "time"_kj, |
| 418 | "datetime"_kj, |
| 419 | "julianday"_kj, |
| 420 | "unixepoch"_kj, |
| 421 | "strftime"_kj, |
| 422 | "timediff"_kj, |
| 423 | "current_date"_kj, |
| 424 | "current_time"_kj, |
| 425 | "current_timestamp"_kj, |
| 426 | |
| 427 | // https://www.sqlite.org/lang_aggfunc.html |
| 428 | "avg"_kj, |
| 429 | "count"_kj, |
| 430 | "group_concat"_kj, |
| 431 | "max"_kj, |
| 432 | "min"_kj, |
| 433 | "string_agg"_kj, |
| 434 | "sum"_kj, |
| 435 | "total"_kj, |
| 436 | |
| 437 | // https://www.sqlite.org/windowfunctions.html#biwinfunc |
| 438 | "row_number"_kj, |
| 439 | "rank"_kj, |
| 440 | "dense_rank"_kj, |
| 441 | "percent_rank"_kj, |
| 442 | "cume_dist"_kj, |
| 443 | "ntile"_kj, |
| 444 | "lag"_kj, |
| 445 | "lead"_kj, |
| 446 | "first_value"_kj, |
| 447 | "last_value"_kj, |
| 448 | "nth_value"_kj, |
| 449 | |
| 450 | // https://www.sqlite.org/lang_mathfunc.html |
| 451 | "acos"_kj, |
| 452 | "acosh"_kj, |
| 453 | "asin"_kj, |
| 454 | "asinh"_kj, |
| 455 | "atan"_kj, |
| 456 | "atan2"_kj, |
| 457 | "atanh"_kj, |
| 458 | "ceil"_kj, |
| 459 | "cos"_kj, |
| 460 | "cosh"_kj, |
| 461 | "degrees"_kj, |
| 462 | "exp"_kj, |
| 463 | "floor"_kj, |
| 464 | "ln"_kj, |
| 465 | "log"_kj, |
| 466 | "log2"_kj, |
| 467 | "mod"_kj, |
| 468 | "pi"_kj, |
| 469 | "pow"_kj, |
| 470 | "radians"_kj, |
| 471 | "sin"_kj, |
| 472 | "sinh"_kj, |
| 473 | "sqrt"_kj, |
| 474 | "tan"_kj, |
| 475 | "tanh"_kj, |
| 476 | "trunc"_kj, |
| 477 | |
| 478 | // https://www.sqlite.org/json1.html |
| 479 | "json"_kj, |
| 480 | "jsonb"_kj, |
| 481 | "json_array"_kj, |
| 482 | "jsonb_array"_kj, |
| 483 | "json_array_length"_kj, |
| 484 | "json_extract"_kj, |
| 485 | "jsonb_extract"_kj, |
| 486 | "->"_kj, |
| 487 | "->>"_kj, |
| 488 | "json_insert"_kj, |
| 489 | "jsonb_insert"_kj, |
| 490 | "json_object"_kj, |
| 491 | "jsonb_object"_kj, |
| 492 | "json_patch"_kj, |
| 493 | "jsonb_patch"_kj, |
| 494 | "json_remove"_kj, |
| 495 | "jsonb_remove"_kj, |
| 496 | "json_replace"_kj, |
| 497 | "jsonb_replace"_kj, |
| 498 | "json_set"_kj, |
| 499 | "jsonb_set"_kj, |
| 500 | "json_type"_kj, |
| 501 | "json_valid"_kj, |
| 502 | "json_quote"_kj, |
| 503 | "json_group_array"_kj, |
| 504 | "jsonb_group_array"_kj, |
| 505 | "json_group_object"_kj, |
| 506 | "jsonb_group_object"_kj, |
| 507 | "json_each"_kj, |
| 508 | "json_tree"_kj, |
| 509 | |
| 510 | // https://www.sqlite.org/fts5.html |
| 511 | "match"_kj, |
| 512 | "highlight"_kj, |
| 513 | "bm25"_kj, |
| 514 | "snippet"_kj, |
| 515 | |
| 516 | // https://www.sqlite.org/lang_altertable.html |
| 517 | // Functions declared in https://sqlite.org/src/file?name=src/alter.c&ci=trunk |
| 518 | "sqlite_rename_column"_kj, |
| 519 | "sqlite_rename_table"_kj, |
| 520 | "sqlite_rename_test"_kj, |
| 521 | "sqlite_drop_column"_kj, |
| 522 | "sqlite_rename_quotefix"_kj, |
| 523 | }; |
| 524 | |
| 525 | enum class PragmaSignature { |
| 526 | NO_ARG, |
| 527 | BOOLEAN, |
| 528 | OBJECT_NAME, |
| 529 | OPTIONAL_OBJECT_NAME, |
| 530 | NULL_OR_NUMBER, |
| 531 | NULL_NUMBER_OR_OBJECT_NAME |
| 532 | }; |
| 533 | struct PragmaInfo { |
| 534 | kj::StringPtr name; |
| 535 | PragmaSignature signature; |
| 536 | }; |
| 537 | |
| 538 | // We allowlist these SQLite pragmas (for read only, never with arguments). |
| 539 | // https://www.sqlite.org/pragma.html |
| 540 | static constexpr PragmaInfo ALLOWED_PRAGMAS[] = {{"data_version"_kj, PragmaSignature::NO_ARG}, |
| 541 | |
| 542 | // We allowlist some SQLite pragmas for changing internal state |
| 543 | |
| 544 | // Toggle constraints on/off |
| 545 | {"case_sensitive_like"_kj, PragmaSignature::BOOLEAN}, |
| 546 | {"foreign_keys"_kj, PragmaSignature::BOOLEAN}, |
| 547 | {"defer_foreign_keys"_kj, PragmaSignature::BOOLEAN}, |
| 548 | {"ignore_check_constraints"_kj, PragmaSignature::BOOLEAN}, |
| 549 | {"legacy_alter_table"_kj, PragmaSignature::BOOLEAN}, |
| 550 | {"recursive_triggers"_kj, PragmaSignature::BOOLEAN}, |
| 551 | {"reverse_unordered_selects"_kj, PragmaSignature::BOOLEAN}, |
| 552 | |
| 553 | // Takes an argument of table name or index name, returns info about it. |
| 554 | {"foreign_key_check"_kj, PragmaSignature::OPTIONAL_OBJECT_NAME}, |
| 555 | {"foreign_key_list"_kj, PragmaSignature::OBJECT_NAME}, |
| 556 | {"index_info"_kj, PragmaSignature::OBJECT_NAME}, {"index_list"_kj, PragmaSignature::OBJECT_NAME}, |
| 557 | {"index_xinfo"_kj, PragmaSignature::OBJECT_NAME}, |
| 558 | |
| 559 | // Takes an argument of table name/index name OR a max number of results, or nothing |
| 560 | {"quick_check"_kj, PragmaSignature::NULL_NUMBER_OR_OBJECT_NAME}, |
| 561 | |
| 562 | // Takes a number representing a bit mask or nothing to use the default mask. |
| 563 | {"optimize"_kj, PragmaSignature::NULL_OR_NUMBER}}; |
| 564 | |
| 565 | } // namespace |
| 566 | |
| 567 | // ======================================================================================= |
| 568 | |
| 569 | SqliteObserver SqliteObserver::DEFAULT = SqliteObserver{}; |
| 570 | |
| 571 | SqliteDatabase::SqliteDatabase(const Vfs& vfs, |
| 572 | kj::Path path, |
| 573 | kj::Maybe<kj::WriteMode> maybeMode, |
| 574 | size_t sqliteMaxMemoryBytes, |
| 575 | SqliteObserver& sqliteObserver, |
| 576 | kj::Maybe<const ActorAccountLimits&> actorAccountLimits) |
| 577 | : vfs(vfs), |
| 578 | path(kj::mv(path)), |
| 579 | readOnly(maybeMode == kj::none), |
| 580 | sqliteObserver(sqliteObserver), |
| 581 | sqliteMaxMemoryBytes(sqliteMaxMemoryBytes), |
| 582 | actorAccountLimits(actorAccountLimits) { |
| 583 | init(maybeMode); |
| 584 | } |
| 585 | |
| 586 | void SqliteDatabase::init(kj::Maybe<kj::WriteMode> maybeMode) { |
| 587 | KJ_ASSERT(maybeDb == kj::none); |
| 588 | sqlite3* db = nullptr; |
| 589 | |
| 590 | auto memoryScope = enterMemoryScope(); |
| 591 | |
| 592 | KJ_IF_SOME(mode, maybeMode) { |
| 593 | int flags = SQLITE_OPEN_READWRITE; |
| 594 | if (kj::has(mode, kj::WriteMode::CREATE)) { |
| 595 | flags |= SQLITE_OPEN_CREATE; |
| 596 | |
| 597 | if (kj::has(mode, kj::WriteMode::CREATE_PARENT) && path.size() > 1) { |
| 598 | // SQLite isn't going to try to create the parent directory so let's try to create it now. |
| 599 | vfs.directory.openSubdir(path.parent(), |
| 600 | kj::WriteMode::CREATE | kj::WriteMode::MODIFY | kj::WriteMode::CREATE_PARENT); |
| 601 | } |
| 602 | } |
| 603 | KJ_REQUIRE( |
| 604 | kj::has(mode, kj::WriteMode::MODIFY), "SQLite doesn't support create-exclusive mode"); |
| 605 | |
| 606 | KJ_IF_SOME(rootedPath, vfs.tryAppend(path)) { |
| 607 | // If we can get the path rooted in the VFS's directory, use the system's default VFS instead |
| 608 | // TODO(bug): This doesn't honor vfs.options. (This branch is only used on Windows.) |
| 609 | SQLITE_CALL_NODB( |
| 610 | sqlite3_open_v2(rootedPath.toNativeString(true).cStr(), &db, flags, nullptr)); |
| 611 | } else { |
| 612 | SQLITE_CALL_NODB(sqlite3_open_v2(path.toString().cStr(), &db, flags, vfs.getName().cStr())); |
| 613 | } |
| 614 | } else { |
| 615 | KJ_IF_SOME(rootedPath, vfs.tryAppend(path)) { |
| 616 | // If we can get the path rooted in the VFS's directory, use the system's default VFS instead |
| 617 | // TODO(bug): This doesn't honor vfs.options. (This branch is only used on Windows.) |
| 618 | SQLITE_CALL_NODB(sqlite3_open_v2( |
| 619 | rootedPath.toNativeString(true).cStr(), &db, SQLITE_OPEN_READONLY, nullptr)); |
| 620 | } else { |
| 621 | SQLITE_CALL_NODB( |
| 622 | sqlite3_open_v2(path.toString().cStr(), &db, SQLITE_OPEN_READONLY, vfs.getName().cStr())); |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | KJ_ON_SCOPE_FAILURE(sqlite3_close_v2(db)); |
| 627 | |
| 628 | setupSecurity(db); |
| 629 | |
| 630 | maybeDb = *db; |
| 631 | } |
| 632 | |
| 633 | SqliteDatabase::~SqliteDatabase() noexcept(false) { |
| 634 | sqlite3* db = &KJ_UNWRAP_OR(maybeDb, return); |
| 635 | |
| 636 | auto memoryScope = enterMemoryScope(); |
| 637 | |
| 638 | auto err = sqlite3_close(db); |
| 639 | if (err == SQLITE_BUSY) { |
| 640 | KJ_LOG(ERROR, "sqlite database destroyed while dependent objects still exist"); |
| 641 | // SQLite actually provides a lazy-close API which we might as well use here instead of leaking |
| 642 | // memory. |
| 643 | err = sqlite3_close_v2(db); |
| 644 | } |
| 645 | |
| 646 | KJ_REQUIRE(err == SQLITE_OK, sqlite3_errstr(err)) { |
| 647 | break; |
| 648 | } |
| 649 | } |
| 650 | |
| 651 | SqliteDatabase::operator sqlite3*() { |
| 652 | return &KJ_ASSERT_NONNULL(maybeDb, "previous reset() failed"); |
| 653 | } |
| 654 | |
| 655 | SqliteMemoryScope SqliteDatabase::enterMemoryScope() { |
| 656 | return SqliteMemoryScope(sqliteMemoryBytes, sqliteMaxMemoryBytes); |
| 657 | } |
| 658 | |
| 659 | bool SqliteDatabase::observedCriticalError() { |
| 660 | return criticalErrorOccurred; |
| 661 | } |
| 662 | |
| 663 | void SqliteDatabase::notifyWrite(bool allowUnconfirmed) { |
| 664 | KJ_IF_SOME(cb, onWriteCallback) { |
| 665 | cb(allowUnconfirmed); |
| 666 | } |
| 667 | } |
| 668 | |
| 669 | void SqliteDatabase::handleCriticalError(kj::Maybe<int> errorCode, |
| 670 | kj::StringPtr errorMessage, |
| 671 | kj::Maybe<const kj::Exception&> maybeException) { |
| 672 | KJ_IF_SOME(code, errorCode) { |
| 673 | // Only errors listed in https://www.sqlite.org/lang_transaction.html#response_to_errors_within_a_transaction |
| 674 | // should be considered here as SQLITE auto rollbacks the transaction when we hit these errors |
| 675 | if (code == SQLITE_FULL || code == SQLITE_IOERR || code == SQLITE_NOMEM || |
| 676 | code == SQLITE_INTERRUPT) { |
| 677 | |
| 678 | sqlite3* db = &KJ_ASSERT_NONNULL(maybeDb, "previous reset() failed"); |
| 679 | // We are in a transaction |
| 680 | if (inTransaction || !savepoints.empty()) { |
| 681 | // The transaction was auto-rolledback, re-enabling the auto commit mode, so we should fail |
| 682 | if (sqlite3_get_autocommit(db) != 0) { |
| 683 | criticalErrorOccurred = true; |
| 684 | KJ_IF_SOME(cb, onCriticalErrorCallback) { |
| 685 | cb(errorMessage, maybeException.map([](const kj::Exception& e) { return e.clone(); })); |
| 686 | } |
| 687 | } |
| 688 | } |
| 689 | } |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | kj::StringPtr SqliteDatabase::getCurrentQueryForDebug() { |
| 694 | KJ_IF_SOME(s, currentStatement) { |
| 695 | return sqlite3_normalized_sql(&s); |
| 696 | } else { |
| 697 | return "(no statement is running)"; |
| 698 | } |
| 699 | } |
| 700 | |
| 701 | void SqliteDatabase::applyChange(const StateChange& change) { |
| 702 | KJ_SWITCH_ONEOF(change) { |
| 703 | KJ_CASE_ONEOF(none, NoChange) { |
| 704 | // Nothing. |
| 705 | } |
| 706 | |
| 707 | KJ_CASE_ONEOF(begin, BeginTxn) { |
| 708 | KJ_IF_SOME(name, begin.savepointName) { |
| 709 | savepoints.add( |
| 710 | Savepoint{.name = kj::str(name), .rollbackCallbackIndex = rollbackCallbacks.size()}); |
| 711 | } else { |
| 712 | KJ_ASSERT(savepoints.empty(), |
| 713 | "BEGIN TRANSACTION should have failed when savepoints are present?"); |
| 714 | KJ_ASSERT( |
| 715 | !inTransaction, "BEGIN TRANSACTION should have failed when already in a transaction?"); |
| 716 | KJ_ASSERT(rollbackCallbacks.empty(), |
| 717 | "we shouldn't have been keeping rollback callbacks with no transaction open!"); |
| 718 | inTransaction = true; |
| 719 | } |
| 720 | } |
| 721 | |
| 722 | KJ_CASE_ONEOF(commit, CommitTxn) { |
| 723 | KJ_IF_SOME(name, commit.savepointName) { |
| 724 | // According to https://www.sqlite.org/lang_savepoint.html, releasing a savepoint also |
| 725 | // releases all later savepoints. In theory it seems like savepoints shouldn't need to |
| 726 | // be LIFO like this, but the docs say they are! |
| 727 | for (;;) { |
| 728 | KJ_ASSERT(!savepoints.empty(), "released a savepoint that didn't exist?"); |
| 729 | auto sp = kj::mv(savepoints.back()); |
| 730 | savepoints.removeLast(); |
| 731 | if (sp.name == name) break; |
| 732 | } |
| 733 | } else { |
| 734 | KJ_ASSERT(inTransaction, "COMMIT TRANSACTION without BEGIN TRANSACTION?"); |
| 735 | |
| 736 | // Since BEGIN TRANSACTION cannot be nested within a savepoint, this must have released |
| 737 | // all savepoints implicitly. |
| 738 | savepoints.clear(); |
| 739 | inTransaction = false; |
| 740 | } |
| 741 | |
| 742 | if (savepoints.empty() && !inTransaction) { |
| 743 | // Transaction stack is empty, so the transaction is committed. We can release the rollback |
| 744 | // callbacks. |
| 745 | rollbackCallbacks.clear(); |
| 746 | } |
| 747 | } |
| 748 | |
| 749 | KJ_CASE_ONEOF(rollback, RollbackTxn) { |
| 750 | KJ_IF_SOME(name, rollback.savepointName) { |
| 751 | for (;;) { |
| 752 | KJ_ASSERT(!savepoints.empty(), "released a savepoint that didn't exist?"); |
| 753 | if (savepoints.back().name == name) { |
| 754 | // Found the savepoint. |
| 755 | // Call all rollback callbacks later than the savepoint. |
| 756 | size_t index = savepoints.back().rollbackCallbackIndex; |
| 757 | KJ_ASSERT(rollbackCallbacks.size() >= index); |
| 758 | while (rollbackCallbacks.size() > index) { |
| 759 | rollbackCallbacks.back()(); |
| 760 | rollbackCallbacks.removeLast(); |
| 761 | } |
| 762 | |
| 763 | // NOTE: Rolling back to a savepoint does not actually release the savepoint. Hence |
| 764 | // we save this savepoint as the last item in `savepoints`. It must be released |
| 765 | // separately. |
| 766 | break; |
| 767 | } |
| 768 | |
| 769 | savepoints.removeLast(); |
| 770 | } |
| 771 | } else { |
| 772 | KJ_ASSERT(inTransaction, "ROLLBACK TRANSACTION without BEGIN TRANSACTION?"); |
| 773 | |
| 774 | savepoints.clear(); |
| 775 | inTransaction = false; |
| 776 | |
| 777 | while (!rollbackCallbacks.empty()) { |
| 778 | rollbackCallbacks.back()(); |
| 779 | rollbackCallbacks.removeLast(); |
| 780 | } |
| 781 | } |
| 782 | } |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | // Set up the regulator that will be used for authorizer callbacks while preparing this |
| 787 | // statement. |
| 788 | SqliteDatabase::StatementAndEffect SqliteDatabase::prepareSql(const Regulator& regulator, |
| 789 | kj::StringPtr sqlCode, |
| 790 | uint prepFlags, |
| 791 | Multi multi, |
| 792 | kj::Maybe<kj::Vector<Statement>&> prelude) { |
| 793 | sqlite3* db = &KJ_ASSERT_NONNULL(maybeDb, "previous reset() failed"); |
| 794 | |
| 795 | ParseContext parseContext; |
| 796 | KJ_ASSERT(currentParseContext == kj::none, "recursive prepareSql()?"); |
| 797 | KJ_DEFER(currentParseContext = kj::none); |
| 798 | currentParseContext = parseContext; |
| 799 | |
| 800 | KJ_ASSERT(currentRegulator == kj::none, |
| 801 | "can't prepare statements inside executeWithRegulator() callback"); |
| 802 | KJ_DEFER(currentRegulator = kj::none); |
| 803 | currentRegulator = regulator; |
| 804 | |
| 805 | // If we fail, we need to discard any statements we added to the prelude, because the next time |
| 806 | // the statement runs they'll be parsed again and added again. |
| 807 | uint preludeInitialSize = 0; |
| 808 | KJ_IF_SOME(p, prelude) { |
| 809 | preludeInitialSize = p.size(); |
| 810 | } |
| 811 | KJ_ON_SCOPE_FAILURE({ |
| 812 | KJ_IF_SOME(p, prelude) { |
| 813 | while (p.size() > preludeInitialSize) { |
| 814 | p.removeLast(); |
| 815 | } |
| 816 | } else { |
| 817 | // (else block needed to squelch spurious clang warning) |
| 818 | } |
| 819 | }); |
| 820 | |
| 821 | auto memoryScope = enterMemoryScope(); |
| 822 | |
| 823 | for (;;) { |
| 824 | sqlite3_stmt* result; |
| 825 | const char* tail; |
| 826 | |
| 827 | SQLITE_CALL_SCOPE { |
| 828 | auto prepareResult = |
| 829 | sqlite3_prepare_v3(db, sqlCode.begin(), sqlCode.size(), prepFlags, &result, &tail); |
| 830 | |
| 831 | // If we had an auth error specifically, check if we recorded a better error message during |
| 832 | // the authorizer callback. |
| 833 | if (prepareResult == SQLITE_AUTH) { |
| 834 | KJ_IF_SOME(error, parseContext.authError) { |
| 835 | // Throw the tailored auth error. |
| 836 | kj::throwFatalException(kj::mv(error)); |
| 837 | } |
| 838 | // we don't have a better error, so fall back to SQLITE_CALL_FAILED below |
| 839 | } |
| 840 | |
| 841 | if (prepareResult != SQLITE_OK) { |
| 842 | SQLITE_CALL_FAILED("sqlite3_prepare_v3", prepareResult); |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | SQLITE_REQUIRE(result != nullptr, kj::none, "SQL code did not contain a statement.", sqlCode); |
| 847 | auto ownResult = ownSqlite(result); |
| 848 | |
| 849 | while (*tail == ' ' || *tail == '\t' || *tail == '\n' || *tail == '\r' || *tail == '\v' || |
| 850 | *tail == '\f') |
| 851 | ++tail; |
| 852 | |
| 853 | switch (multi) { |
| 854 | case SINGLE: |
| 855 | SQLITE_REQUIRE(tail == sqlCode.end(), kj::none, |
| 856 | "A prepared SQL statement must contain only one statement.", tail); |
| 857 | break; |
| 858 | |
| 859 | case MULTI: |
| 860 | if (tail != sqlCode.end()) { |
| 861 | // There are more statements after this one, so execute this statement now. |
| 862 | |
| 863 | SQLITE_REQUIRE(sqlite3_bind_parameter_count(result) == 0, kj::none, |
| 864 | "When executing multiple SQL statements in a single call, only the last statement " |
| 865 | "can have parameters."); |
| 866 | |
| 867 | // Be sure to call the onWrite callback if necessary for this statement. |
| 868 | KJ_IF_SOME(cb, onWriteCallback) { |
| 869 | if (!sqlite3_stmt_readonly(result)) { |
| 870 | // The callback is allowed to invoke queries of its own, so we have to un-set the |
| 871 | // regulator and parse context while we call it. |
| 872 | currentRegulator = kj::none; |
| 873 | KJ_DEFER(currentRegulator = regulator); |
| 874 | currentParseContext = kj::none; |
| 875 | KJ_DEFER(currentParseContext = parseContext); |
| 876 | cb(false); // prepareSql doesn't have access to allowUnconfirmed, use safe default |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | // This isn't the last statement in the code. Execute it immediately. |
| 881 | SQLITE_CALL_SCOPE { |
| 882 | auto start = sqliteObserver.now(); |
| 883 | auto dbWalSizeBefore = sqliteObserver.getDbWalSize(); |
| 884 | |
| 885 | int err = sqlite3_step(result); |
| 886 | int extendedCode = sqlite3_extended_errcode(db); |
| 887 | |
| 888 | kj::Duration queryLatency = sqliteObserver.now() - start; |
| 889 | auto dbWalBytesWritten = sqliteObserver.getDbWalSize() - dbWalSizeBefore; |
| 890 | auto rowsRead = sqlite3_stmt_status(result, LIBSQL_STMTSTATUS_ROWS_READ, 0); |
| 891 | auto rowsWritten = sqlite3_stmt_status(result, LIBSQL_STMTSTATUS_ROWS_WRITTEN, 0); |
| 892 | |
| 893 | kj::Maybe<kj::String> queryStatement; |
| 894 | kj::Maybe<kj::String> queryErrorDescription; |
| 895 | try { |
| 896 | kj::StringPtr statement = sqlite3_sql(result); |
| 897 | queryStatement = kj::heapString( |
| 898 | statement.slice(0, kj::min(statement.size(), RA_MAX_METRICS_QUERY_SIZE))); |
| 899 | } catch (kj::Exception& e) { |
| 900 | kj::StringPtr errorDescription = e.getDescription(); |
| 901 | queryErrorDescription = kj::heapString(errorDescription.slice( |
| 902 | 0, kj::min(RA_MAX_METRICS_QUERY_SIZE, errorDescription.size()))); |
| 903 | } |
| 904 | |
| 905 | // Report queryEvent for this statement |
| 906 | sqliteObserver.reportQueryEvent(kj::mv(queryStatement), rowsRead, rowsWritten, |
| 907 | queryLatency, dbWalBytesWritten, err, extendedCode, regulator.shouldAddQueryStats(), |
| 908 | kj::mv(queryErrorDescription)); |
| 909 | |
| 910 | if (err == SQLITE_DONE) { |
| 911 | // good |
| 912 | } else if (err == SQLITE_ROW) { |
| 913 | // Intermediate statement returned results. We will discard. |
| 914 | } else { |
| 915 | SQLITE_CALL_FAILED("sqlite3_step()", err); |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | // Apply any state changes from executing the statement. |
| 920 | applyChange(parseContext.stateChange); |
| 921 | |
| 922 | KJ_IF_SOME(p, prelude) { |
| 923 | p.add(Statement(*this, regulator, |
| 924 | StatementAndEffect{.statement = kj::mv(ownResult), |
| 925 | .stateChange = kj::mv(parseContext.stateChange)})); |
| 926 | } |
| 927 | |
| 928 | // Reset parse context for next statement. |
| 929 | parseContext = {}; |
| 930 | |
| 931 | // Reduce `sqlCode` to include only what we haven't already executed. |
| 932 | sqlCode = kj::StringPtr(tail, sqlCode.end()); |
| 933 | |
| 934 | continue; |
| 935 | } |
| 936 | break; |
| 937 | } |
| 938 | |
| 939 | return {.statement = kj::mv(ownResult), .stateChange = kj::mv(parseContext.stateChange)}; |
| 940 | } |
| 941 | } |
| 942 | |
| 943 | SqliteDatabase::IngestResult SqliteDatabase::ingestSql( |
| 944 | const Regulator& regulator, kj::StringPtr sqlCode) { |
| 945 | uint64_t rowsRead = 0; |
| 946 | uint64_t rowsWritten = 0; |
| 947 | uint64_t statementCount = 0; |
| 948 | |
| 949 | // While there's still some input SQL to process |
| 950 | while (sqlCode.begin() != sqlCode.end()) { |
| 951 | // And there are still valid statements: |
| 952 | auto statementLength = sqlite3_complete_length(sqlCode.begin(), 1); |
| 953 | if (!statementLength) break; |
| 954 | |
| 955 | // Slice off the next valid statement SQL |
| 956 | auto nextStatement = kj::str(sqlCode.first(statementLength)); |
| 957 | // Create a Query object, which will prepare & execute it |
| 958 | auto q = Query(*this, QueryOptions{.regulator = regulator}, nextStatement); |
| 959 | |
| 960 | rowsRead += q.getRowsRead(); |
| 961 | rowsWritten += q.getRowsWritten(); |
| 962 | statementCount++; |
| 963 | sqlCode = sqlCode.slice(statementLength); |
| 964 | } |
| 965 | |
| 966 | // Return the leftover buffer |
| 967 | return {.remainder = sqlCode, |
| 968 | .rowsRead = rowsRead, |
| 969 | .rowsWritten = rowsWritten, |
| 970 | .statementCount = statementCount}; |
| 971 | } |
| 972 | |
| 973 | void SqliteDatabase::executeWithRegulator( |
| 974 | const Regulator& regulator, kj::FunctionParam<void()> func) { |
| 975 | // currentRegulator would only be set if we're running this method while running something else |
| 976 | // with a regulator. I'm not sure what the ramifications are, so for now, we'll just assume that |
| 977 | // we can only call executeWithRegulator when no regulator is currently set. |
| 978 | KJ_REQUIRE(currentRegulator == kj::none); |
| 979 | |
| 980 | currentRegulator = regulator; |
| 981 | KJ_DEFER(currentRegulator = kj::none); |
| 982 | |
| 983 | auto memoryScope = enterMemoryScope(); |
| 984 | func(); |
| 985 | } |
| 986 | |
| 987 | void SqliteDatabase::reset() { |
| 988 | KJ_REQUIRE(!readOnly, "can't reset() read-only database"); |
| 989 | |
| 990 | // If transactions are open during reset(), whatever had the transaction open is going to get |
| 991 | // confused at best, or lose data at worst. Let's just not allow this. |
| 992 | KJ_REQUIRE(!inTransaction && savepoints.empty(), "can't reset() a database during a transaction"); |
| 993 | |
| 994 | auto memoryScope = enterMemoryScope(); |
| 995 | |
| 996 | // Temporarily disable the on-write callback while resetting. |
| 997 | auto writeCb = kj::mv(onWriteCallback); |
| 998 | KJ_DEFER(onWriteCallback = kj::mv(writeCb)); |
| 999 | |
| 1000 | KJ_IF_SOME(db, maybeDb) { |
| 1001 | for (auto& listener: resetListeners) { |
| 1002 | listener.beforeSqliteReset(); |
| 1003 | } |
| 1004 | |
| 1005 | auto err = sqlite3_close(&db); |
| 1006 | KJ_REQUIRE(err == SQLITE_OK, "can't reset() database because dependent objects still exist", |
| 1007 | sqlite3_errstr(err)); |
| 1008 | |
| 1009 | maybeDb = kj::none; |
| 1010 | vfs.directory.remove(path); |
| 1011 | } |
| 1012 | |
| 1013 | KJ_ON_SCOPE_FAILURE(maybeDb = kj::none); |
| 1014 | init(kj::WriteMode::CREATE | kj::WriteMode::MODIFY); |
| 1015 | |
| 1016 | KJ_IF_SOME(resetCb, afterResetCallback) { |
| 1017 | resetCb(*this); |
| 1018 | } |
| 1019 | } |
| 1020 | |
| 1021 | bool SqliteDatabase::isAuthorized(int actionCode, |
| 1022 | kj::Maybe<kj::StringPtr> param1, |
| 1023 | kj::Maybe<kj::StringPtr> param2, |
| 1024 | kj::Maybe<kj::StringPtr> dbName, |
| 1025 | kj::Maybe<kj::StringPtr> triggerName) { |
| 1026 | const Regulator& regulator = KJ_UNWRAP_OR(currentRegulator, { |
| 1027 | // We're not currently preparing a statement, so we didn't expect the authorizer callback to |
| 1028 | // run. We blanket-deny in this case as a precaution. |
| 1029 | KJ_LOG(ERROR, "SQLite authorizer callback invoked at unexpected time", kj::getStackTrace()); |
| 1030 | return false; |
| 1031 | }); |
| 1032 | |
| 1033 | KJ_IF_SOME(t, triggerName) { |
| 1034 | if (!regulator.isAllowedTrigger(t)) { |
| 1035 | // Log an error because it seems really suspicious if a trigger runs when it's not allowed. |
| 1036 | // I want to understand if this can even happen. |
| 1037 | KJ_LOG(ERROR, "disallowed trigger somehow ran in trusted scope?", t, kj::getStackTrace()); |
| 1038 | |
| 1039 | // TODO(security): Is it better to return SQLITE_IGNORE to ignore the trigger? I don't fully |
| 1040 | // understand the implications of SQLITE_IGNORE. The documentation mentions that in the |
| 1041 | // case of SQLITE_DELETE, it doesn't actually ignore the delete, which is weird. Hopefully |
| 1042 | // it's impossible for people to register a trigger on protected tables in the first place, |
| 1043 | // so triggers will never run. |
| 1044 | return false; |
| 1045 | } |
| 1046 | } |
| 1047 | |
| 1048 | // For some reason, for these two operations, SQLite sends the DB Name through as param1, with |
| 1049 | // the table name (for ALTER_TABLE) in param2 instead of param1 like all other table operations. |
| 1050 | // For simplicity, and because the following comment precedes sqlite3_set_authorizer in sqlite.h: |
| 1051 | // |
| 1052 | // > The 5th parameter to the authorizer callback is the name of the database |
| 1053 | // > ("main", "temp", etc.) if applicable. |
| 1054 | // |
| 1055 | // we are treating this as an SQLite bug and swapping the values around. |
| 1056 | if (actionCode == SQLITE_ALTER_TABLE || actionCode == SQLITE_DETACH) { |
| 1057 | auto swap = param1; // contains dbName |
| 1058 | param1 = param2; // contains table name (for SQLITE_ALTER_TABLE, null otherwise) |
| 1059 | param2 = dbName; // should always be null |
| 1060 | dbName = swap; |
| 1061 | } |
| 1062 | |
| 1063 | KJ_IF_SOME(d, dbName) { |
| 1064 | if (d == "temp"_kj) { |
| 1065 | return isAuthorizedTemp(actionCode, param1, param2, regulator); |
| 1066 | } else if (d != "main"_kj) { |
| 1067 | // We don't allow opening multiple databases (except for 'main' and the 'temp' |
| 1068 | // temporary database), as our storage engine is not designed to track multiple |
| 1069 | // files on-disk. |
| 1070 | return false; |
| 1071 | } |
| 1072 | } |
| 1073 | |
| 1074 | if (®ulator == &TRUSTED && actionCode != SQLITE_TRANSACTION && |
| 1075 | actionCode != SQLITE_SAVEPOINT) { |
| 1076 | // Everything is allowed for trusted queries. (But transactions and savepoints need special |
| 1077 | // handling below.) |
| 1078 | return true; |
| 1079 | } |
| 1080 | |
| 1081 | switch (actionCode) { |
| 1082 | // --------------------------------------------------------------- |
| 1083 | // Stuff that is (sometimes) allowed |
| 1084 | |
| 1085 | case SQLITE_SELECT: /* NULL NULL */ |
| 1086 | // Yes, SELECT statements are allowed. (Note that if the SELECT names any tables, a separate |
| 1087 | // SQLITE_READ will be authorized for each one.) |
| 1088 | KJ_ASSERT(param1 == kj::none); |
| 1089 | KJ_ASSERT(param2 == kj::none); |
| 1090 | return true; |
| 1091 | |
| 1092 | case SQLITE_CREATE_TABLE: /* Table Name NULL */ |
| 1093 | case SQLITE_DELETE: /* Table Name NULL */ |
| 1094 | case SQLITE_DROP_TABLE: /* Table Name NULL */ |
| 1095 | case SQLITE_INSERT: /* Table Name NULL */ |
| 1096 | case SQLITE_CREATE_VIEW: /* View Name NULL */ |
| 1097 | case SQLITE_DROP_VIEW: /* View Name NULL */ |
| 1098 | case SQLITE_REINDEX: /* Index Name NULL */ |
| 1099 | KJ_ASSERT(param2 == kj::none); |
| 1100 | return regulator.isAllowedName(KJ_ASSERT_NONNULL(param1)); |
| 1101 | |
| 1102 | case SQLITE_ANALYZE: /* Table Name NULL */ |
| 1103 | KJ_ASSERT(param2 == kj::none); |
| 1104 | // We allow all names (including names where isAllowedName() would return false) because |
| 1105 | // `PRAGMA optimize` issues an ANALYZE statement with no arguments and a SQLite ANALYZE |
| 1106 | // statement with no parameters will analyze all tables, including otherwise restricted |
| 1107 | // tables. |
| 1108 | // |
| 1109 | // The ANALYZE statement records information about the distribution of rows in each index in |
| 1110 | // the database in a special sqlite_stat1 table. While the sqlite_stat1 table leaks metadata |
| 1111 | // about restricted tables (like the names of indices and the sizes of those tables), the |
| 1112 | // sqlite_stat1 does not contain data from the restricted tables. As such, it's OK to allow |
| 1113 | // users to ANALYZE restricted tables. |
| 1114 | // |
| 1115 | // Note that users can *modify* the sqlite_stat1 table, which means that they can make the |
| 1116 | // query planner work in suboptimal ways by writing bogus data to the table. |
| 1117 | // |
| 1118 | // See https://www.sqlite.org/fileformat2.html#stat1tab for more details. |
| 1119 | return true; |
| 1120 | |
| 1121 | case SQLITE_ALTER_TABLE: /* Table Name NULL (modified) */ |
| 1122 | return regulator.isAllowedName(KJ_ASSERT_NONNULL(param1)); |
| 1123 | |
| 1124 | case SQLITE_READ: /* Table Name Column Name */ |
| 1125 | case SQLITE_UPDATE: /* Table Name Column Name */ |
| 1126 | return regulator.isAllowedName(KJ_ASSERT_NONNULL(param1)); |
| 1127 | |
| 1128 | case SQLITE_CREATE_INDEX: /* Index Name Table Name */ |
| 1129 | case SQLITE_DROP_INDEX: /* Index Name Table Name */ |
| 1130 | case SQLITE_CREATE_TRIGGER: /* Trigger Name Table Name */ |
| 1131 | case SQLITE_DROP_TRIGGER: /* Trigger Name Table Name */ |
| 1132 | return regulator.isAllowedName(KJ_ASSERT_NONNULL(param1)) && |
| 1133 | regulator.isAllowedName(KJ_ASSERT_NONNULL(param2)); |
| 1134 | |
| 1135 | case SQLITE_TRANSACTION: /* Operation NULL */ |
| 1136 | { |
| 1137 | if (!regulator.allowTransactions()) { |
| 1138 | return false; |
| 1139 | } |
| 1140 | |
| 1141 | kj::StringPtr op = KJ_ASSERT_NONNULL(param1); |
| 1142 | StateChange change; |
| 1143 | if (op == "BEGIN") { |
| 1144 | change = BeginTxn{kj::none}; |
| 1145 | } else if (op == "COMMIT") { |
| 1146 | change = CommitTxn{kj::none}; |
| 1147 | } else if (op == "ROLLBACK") { |
| 1148 | change = RollbackTxn{kj::none}; |
| 1149 | } else { |
| 1150 | KJ_FAIL_ASSERT("unknown SQLITE_TRANSACTION op", op); |
| 1151 | } |
| 1152 | KJ_IF_SOME(ctx, currentParseContext) { |
| 1153 | ctx.stateChange = kj::mv(change); |
| 1154 | } |
| 1155 | |
| 1156 | KJ_ASSERT(param2 == kj::none); |
| 1157 | return true; |
| 1158 | } |
| 1159 | |
| 1160 | case SQLITE_SAVEPOINT: /* Operation Savepoint Name */ |
| 1161 | { |
| 1162 | kj::String name = kj::str(KJ_ASSERT_NONNULL(param2)); |
| 1163 | if (!regulator.allowTransactions() || !regulator.isAllowedName(name)) { |
| 1164 | return false; |
| 1165 | } |
| 1166 | |
| 1167 | kj::StringPtr op = KJ_ASSERT_NONNULL(param1); |
| 1168 | StateChange change; |
| 1169 | if (op == "BEGIN") { |
| 1170 | change = BeginTxn{kj::mv(name)}; |
| 1171 | } else if (op == "RELEASE") { |
| 1172 | change = CommitTxn{kj::mv(name)}; |
| 1173 | } else if (op == "ROLLBACK") { |
| 1174 | change = RollbackTxn{kj::mv(name)}; |
| 1175 | } else { |
| 1176 | KJ_FAIL_ASSERT("unknown SQLITE_TRANSACTION op", op); |
| 1177 | } |
| 1178 | KJ_IF_SOME(ctx, currentParseContext) { |
| 1179 | ctx.stateChange = kj::mv(change); |
| 1180 | } |
| 1181 | |
| 1182 | return true; |
| 1183 | } |
| 1184 | |
| 1185 | case SQLITE_PRAGMA: /* Pragma Name 1st arg or NULL */ |
| 1186 | // We currently only permit a few pragmas. |
| 1187 | { |
| 1188 | kj::StringPtr pragma = KJ_ASSERT_NONNULL(param1); |
| 1189 | |
| 1190 | if (pragma == "table_list") { |
| 1191 | // Annoyingly, this will list internal tables. However, the existence of these tables |
| 1192 | // isn't really a secret, we just don't want people to access them. |
| 1193 | return true; |
| 1194 | // TODO function_list & pragma_list should be authorized but return |
| 1195 | // ALLOWED_SQLITE_FUNCTIONS & ALLOWED_[READ|WRITE]_PRAGMAS |
| 1196 | // respectively |
| 1197 | } else if (pragma == "table_info" || pragma == "table_xinfo") { |
| 1198 | // Allow if the specific named table is not protected. |
| 1199 | KJ_IF_SOME(name, param2) { |
| 1200 | return regulator.isAllowedName(name); |
| 1201 | } else { |
| 1202 | return false; // shouldn't happen? |
| 1203 | } |
| 1204 | } |
| 1205 | |
| 1206 | static const kj::HashMap<kj::StringPtr, PragmaSignature> allowedPragmas = []() { |
| 1207 | kj::HashMap<kj::StringPtr, PragmaSignature> result; |
| 1208 | for (auto& [name, signature]: ALLOWED_PRAGMAS) { |
| 1209 | result.insert(name, signature); |
| 1210 | } |
| 1211 | return result; |
| 1212 | }(); |
| 1213 | |
| 1214 | PragmaSignature sig = KJ_UNWRAP_OR(allowedPragmas.find(pragma), return false); |
| 1215 | switch (sig) { |
| 1216 | case PragmaSignature::NO_ARG: |
| 1217 | return param2 == kj::none; |
| 1218 | case PragmaSignature::BOOLEAN: { |
| 1219 | // We allow omitting the argument in order to read back the current value. |
| 1220 | auto val = KJ_UNWRAP_OR(param2, return true).asArray(); |
| 1221 | |
| 1222 | // SQLite offers many different ways to express booleans... |
| 1223 | |
| 1224 | // They can be quoted. Remove quotes if present. |
| 1225 | if (val.size() >= 2 && (val.front() == '\'' || val.front() == '\"') && |
| 1226 | val.back() == val.front()) { |
| 1227 | val = val.slice(1, val.size() - 1); |
| 1228 | } |
| 1229 | |
| 1230 | // Compare against every possible representation. Case-insensitive! |
| 1231 | return strncasecmp(val.begin(), "true", 4) == 0 || |
| 1232 | strncasecmp(val.begin(), "false", 5) == 0 || |
| 1233 | strncasecmp(val.begin(), "yes", 3) == 0 || strncasecmp(val.begin(), "no", 2) == 0 || |
| 1234 | strncasecmp(val.begin(), "on", 2) == 0 || strncasecmp(val.begin(), "off", 3) == 0 || |
| 1235 | strncasecmp(val.begin(), "1", 1) == 0 || strncasecmp(val.begin(), "0", 1) == 0; |
| 1236 | } |
| 1237 | case PragmaSignature::OBJECT_NAME: { |
| 1238 | // Argument is required. |
| 1239 | auto val = KJ_UNWRAP_OR(param2, return false); |
| 1240 | return regulator.isAllowedName(val); |
| 1241 | } |
| 1242 | case PragmaSignature::OPTIONAL_OBJECT_NAME: { |
| 1243 | auto val = KJ_UNWRAP_OR(param2, return true); |
| 1244 | return regulator.isAllowedName(val); |
| 1245 | } |
| 1246 | case PragmaSignature::NULL_OR_NUMBER: { |
| 1247 | // Argument is not required |
| 1248 | auto val = KJ_UNWRAP_OR(param2, return true); |
| 1249 | // val is allowed if it parses to an integer |
| 1250 | return val.tryParseAs<int32_t>() != kj::none; |
| 1251 | } |
| 1252 | case PragmaSignature::NULL_NUMBER_OR_OBJECT_NAME: { |
| 1253 | // Argument is not required |
| 1254 | auto val = KJ_UNWRAP_OR(param2, return true); |
| 1255 | // val is allowed if it parses to an integer |
| 1256 | if (val.tryParseAs<uint>() != kj::none) return true; |
| 1257 | // Otherwise, val must be the name of an object the user has access to |
| 1258 | return regulator.isAllowedName(val); |
| 1259 | } |
| 1260 | } |
| 1261 | KJ_UNREACHABLE; |
| 1262 | } |
| 1263 | |
| 1264 | return false; |
| 1265 | |
| 1266 | case SQLITE_FUNCTION: /* NULL Function Name */ |
| 1267 | { |
| 1268 | static const kj::HashSet<kj::StringPtr> allowSet = []() { |
| 1269 | kj::HashSet<kj::StringPtr> result; |
| 1270 | for (const kj::StringPtr& func: ALLOWED_SQLITE_FUNCTIONS) { |
| 1271 | result.insert(func); |
| 1272 | } |
| 1273 | return result; |
| 1274 | }(); |
| 1275 | return allowSet.contains(KJ_ASSERT_NONNULL(param2)); |
| 1276 | } |
| 1277 | |
| 1278 | // --------------------------------------------------------------- |
| 1279 | // Stuff that is never allowed |
| 1280 | |
| 1281 | case SQLITE_CREATE_VTABLE: /* Table Name Module Name */ |
| 1282 | case SQLITE_DROP_VTABLE: /* Table Name Module Name */ |
| 1283 | // Virtual tables are tables backed by some native-code callbacks. |
| 1284 | // We don't support these except for FTS5 (Full Text Search) https://www.sqlite.org/fts5.html |
| 1285 | // (Which also includes fts5vocab: "[fts5vocab] is available whenever FTS5 is") |
| 1286 | { |
| 1287 | KJ_IF_SOME(moduleName, param2) { |
| 1288 | if (strcasecmp(moduleName.begin(), "fts5") == 0 || |
| 1289 | strcasecmp(moduleName.begin(), "fts5vocab") == 0) { |
| 1290 | if (util::Autogate::isEnabled(util::AutogateKey::SQL_RESTRICT_RESERVED_NAMES)) { |
| 1291 | return regulator.isAllowedName(KJ_ASSERT_NONNULL(param1)); |
| 1292 | } |
| 1293 | auto& tableName = KJ_ASSERT_NONNULL(param1); |
| 1294 | if (tableName.size() >= 4 && strncasecmp(tableName.begin(), "_cf_", 4) == 0) { |
| 1295 | LOG_WARNING_PERIODICALLY("FTS5 virtual table uses reserved _cf_ prefix"); |
| 1296 | } |
| 1297 | return true; |
| 1298 | } |
| 1299 | } |
| 1300 | return false; |
| 1301 | } |
| 1302 | |
| 1303 | case SQLITE_ATTACH: /* Filename NULL */ |
| 1304 | case SQLITE_DETACH: /* Table Name NULL (modified) */ |
| 1305 | // We do not support attached databases. It seems unlikely that we ever will. |
| 1306 | return false; |
| 1307 | |
| 1308 | case SQLITE_CREATE_TEMP_TABLE: /* Table Name NULL */ |
| 1309 | case SQLITE_DROP_TEMP_TABLE: /* Table Name NULL */ |
| 1310 | case SQLITE_CREATE_TEMP_INDEX: /* Index Name Table Name */ |
| 1311 | case SQLITE_DROP_TEMP_INDEX: /* Index Name Table Name */ |
| 1312 | case SQLITE_CREATE_TEMP_TRIGGER: /* Trigger Name Table Name */ |
| 1313 | case SQLITE_DROP_TEMP_TRIGGER: /* Trigger Name Table Name */ |
| 1314 | case SQLITE_CREATE_TEMP_VIEW: /* View Name NULL */ |
| 1315 | case SQLITE_DROP_TEMP_VIEW: /* View Name NULL */ |
| 1316 | // TODO(someday): Allow temporary tables. Creating a temporary table actually causes |
| 1317 | // SQLite to open a separate temporary file to place the data in. Currently, our storage |
| 1318 | // engine has no support for this. |
| 1319 | return false; |
| 1320 | |
| 1321 | case SQLITE_RECURSIVE: /* NULL NULL */ |
| 1322 | // Recursive select, this is fine. |
| 1323 | return true; |
| 1324 | |
| 1325 | case SQLITE_COPY: /* No longer used */ |
| 1326 | // These are operations we simply don't support today. |
| 1327 | return false; |
| 1328 | |
| 1329 | default: |
| 1330 | KJ_LOG(WARNING, "unknown SQLite action", actionCode); |
| 1331 | return false; |
| 1332 | } |
| 1333 | } |
| 1334 | |
| 1335 | // Temp databases have very restricted operations |
| 1336 | bool SqliteDatabase::isAuthorizedTemp(int actionCode, |
| 1337 | const kj::Maybe<kj::StringPtr>& param1, |
| 1338 | const kj::Maybe<kj::StringPtr>& param2, |
| 1339 | const Regulator& regulator) { |
| 1340 | |
| 1341 | switch (actionCode) { |
| 1342 | case SQLITE_READ: /* Table Name Column Name */ |
| 1343 | case SQLITE_UPDATE: /* Table Name Column Name */ |
| 1344 | return regulator.isAllowedName(KJ_ASSERT_NONNULL(param1)); |
| 1345 | default: |
| 1346 | return false; |
| 1347 | } |
| 1348 | } |
| 1349 | |
| 1350 | // Set up security restrictions. |
| 1351 | // See: https://www.sqlite.org/security.html |
| 1352 | void SqliteDatabase::setupSecurity(sqlite3* db) { |
| 1353 | // 1. Set defensive mode. |
| 1354 | SQLITE_CALL_NODB(sqlite3_db_config(db, SQLITE_DBCONFIG_DEFENSIVE, 1, nullptr)); |
| 1355 | |
| 1356 | // 2. Reduce limits |
| 1357 | // We use the suggested limits from the web site. Note that sqlite3_limit() does NOT return an |
| 1358 | // error code; it returns the old limit. |
| 1359 | |
| 1360 | // This limit is set higher than what is suggested on sqlite.org/security.html |
| 1361 | // because we want to allow storing values of 1MiB, and we added some extra |
| 1362 | // padding on top of that |
| 1363 | sqlite3_limit(db, SQLITE_LIMIT_LENGTH, 2200000); |
| 1364 | sqlite3_limit(db, SQLITE_LIMIT_SQL_LENGTH, 100000); |
| 1365 | sqlite3_limit(db, SQLITE_LIMIT_COLUMN, 100); |
| 1366 | sqlite3_limit(db, SQLITE_LIMIT_EXPR_DEPTH, 100); |
| 1367 | // Enforces limits on UNION/UNION ALL/INTERSECT/etc |
| 1368 | // https://www.sqlite.org/limits.html#max_compound_select |
| 1369 | sqlite3_limit(db, SQLITE_LIMIT_COMPOUND_SELECT, 5); |
| 1370 | sqlite3_limit(db, SQLITE_LIMIT_VDBE_OP, 25000); |
| 1371 | // For SQLITE_LIMIT_FUNCTION_ARG we use the default instead of the "security" recommendation |
| 1372 | // because there are too many valid use cases for large argument lists, especially json_object. |
| 1373 | sqlite3_limit(db, SQLITE_LIMIT_FUNCTION_ARG, 127); |
| 1374 | sqlite3_limit(db, SQLITE_LIMIT_ATTACHED, 0); |
| 1375 | sqlite3_limit(db, SQLITE_LIMIT_LIKE_PATTERN_LENGTH, 50); |
| 1376 | sqlite3_limit(db, SQLITE_LIMIT_VARIABLE_NUMBER, 100); |
| 1377 | sqlite3_limit(db, SQLITE_LIMIT_TRIGGER_DEPTH, 10); |
| 1378 | sqlite3_limit(db, SQLITE_LIMIT_WORKER_THREADS, 0); |
| 1379 | |
| 1380 | // 3. Setup authorizer. |
| 1381 | SQLITE_CALL_NODB(sqlite3_set_authorizer(db, |
| 1382 | [](void* userdata, int actionCode, const char* param1, const char* param2, const char* dbName, |
| 1383 | const char* triggerName) { |
| 1384 | try { |
| 1385 | return reinterpret_cast<SqliteDatabase*>(userdata)->isAuthorized(actionCode, |
| 1386 | toMaybeString(param1), toMaybeString(param2), toMaybeString(dbName), |
| 1387 | toMaybeString(triggerName)) |
| 1388 | ? SQLITE_OK |
| 1389 | : SQLITE_DENY; |
| 1390 | } catch (kj::Exception& e) { |
| 1391 | // We'll crash if we throw to SQLite. Instead, shove the error into the parse context and |
| 1392 | // report authorization denied. We'll pull it back out later. |
| 1393 | KJ_IF_SOME(context, reinterpret_cast<SqliteDatabase*>(userdata)->currentParseContext) { |
| 1394 | context.authError = kj::mv(e); |
| 1395 | } else { |
| 1396 | KJ_LOG(ERROR, e); |
| 1397 | } |
| 1398 | return SQLITE_DENY; |
| 1399 | } |
| 1400 | }, |
| 1401 | this)); |
| 1402 | |
| 1403 | // 4. Set a progress handler or use interrupt() to limit CPU time. |
| 1404 | // This happens inside LimitEnforcer. |
| 1405 | |
| 1406 | // 5. Limit process-wide heap size. |
| 1407 | // Set a 128MB "soft" limit so that SQLite will purge the page cache when per-process memory |
| 1408 | // consumption exceeds this value, and an 8 GiB "hard" limit as a defense in depth mechanism to |
| 1409 | // prevent SQLite from consuming too much per-process memory. The primary mechanism for limiting |
| 1410 | // SQLite memory consumption is the metering done in the sqlite-metering module. |
| 1411 | static bool doOnce KJ_UNUSED = []() { |
| 1412 | sqlite3_soft_heap_limit64(128u << 20); |
| 1413 | sqlite3_hard_heap_limit64( |
| 1414 | util::Autogate::isEnabled(util::AutogateKey::INCREASE_SQLITE_HARD_HEAP_LIMIT) |
| 1415 | ? (8ull << 30) // 8 GiB |
| 1416 | : (512u << 20)); // 512 MiB |
| 1417 | return false; |
| 1418 | }(); |
| 1419 | |
| 1420 | // 6. Set SQLITE_MAX_ALLOCATION_SIZE compile flag. |
| 1421 | // (handled in BUILD.sqlite3) |
| 1422 | |
| 1423 | // 7. Consider giving SQLite a fixed heap space. |
| 1424 | // This is suggested mainly for embedded systems. It involves giving SQLite a fixed preallocated |
| 1425 | // heap space which the library restricts itself to instead of using malloc. We probably don't |
| 1426 | // want this. |
| 1427 | |
| 1428 | // 8. Set the SQLITE_PRINTF_PRECISION_LIMIT compile flag. |
| 1429 | // (handled in BUILD.sqlite3) |
| 1430 | } |
| 1431 | |
| 1432 | SqliteDatabase::Statement SqliteDatabase::prepare( |
| 1433 | const Regulator& regulator, kj::StringPtr sqlCode) { |
| 1434 | return Statement( |
| 1435 | *this, regulator, prepareSql(regulator, sqlCode, SQLITE_PREPARE_PERSISTENT, SINGLE)); |
| 1436 | } |
| 1437 | |
| 1438 | SqliteDatabase::StatementAndEffect& SqliteDatabase::Statement::prepareForExecution() { |
| 1439 | for (auto& stmt: prelude) { |
| 1440 | stmt.run(); |
| 1441 | } |
| 1442 | |
| 1443 | KJ_IF_SOME(sqlCode, stmt.tryGet<kj::String>()) { |
| 1444 | // Database was reset. Recompile the statement against the new database. (This could throw, |
| 1445 | // of course, if the statement depends on tables that haven't been recreated yet.) |
| 1446 | // |
| 1447 | // We use the MULTI flag here in case this Statement was created by prepareMulti(). If multiple |
| 1448 | // statements are parsed, they'll be added to our `prelude`, and also executed immediately. |
| 1449 | stmt = db.prepareSql(regulator, sqlCode, SQLITE_PREPARE_PERSISTENT, MULTI, prelude); |
| 1450 | } |
| 1451 | |
| 1452 | return KJ_ASSERT_NONNULL(stmt.tryGet<StatementAndEffect>()); |
| 1453 | } |
| 1454 | |
| 1455 | void SqliteDatabase::Statement::beforeSqliteReset() { |
| 1456 | KJ_IF_SOME(prepared, stmt.tryGet<StatementAndEffect>()) { |
| 1457 | // Pull the original SQL code out of the statement and store it. |
| 1458 | stmt = kj::str(sqlite3_sql(prepared.statement)); |
| 1459 | } |
| 1460 | } |
| 1461 | |
| 1462 | SqliteDatabase::Statement::~Statement() noexcept(false) { |
| 1463 | // Install memory scope for sqlite3_finalize called when stmt (containing StatementAndEffect |
| 1464 | // with kj::Own<sqlite3_stmt>) is destroyed. Also covers prelude destruction. |
| 1465 | auto memoryScope = db.enterMemoryScope(); |
| 1466 | auto stmtToDestroy = kj::mv(stmt); |
| 1467 | auto preludeToDestroy = kj::mv(prelude); |
| 1468 | } |
| 1469 | |
| 1470 | SqliteDatabase::Query::Query(SqliteDatabase& db, |
| 1471 | QueryOptions options, |
| 1472 | Statement& statement, |
| 1473 | kj::ArrayPtr<const ValuePtr> bindings) |
| 1474 | : ResetListener(db), |
| 1475 | regulator(options.regulator), |
| 1476 | maybeStatement(statement.prepareForExecution()), |
| 1477 | queryEvent(this->db.sqliteObserver) { |
| 1478 | // If we throw from the constructor, the destructor won't run. Need to call destroy() explicitly. |
| 1479 | KJ_ON_SCOPE_FAILURE(destroy()); |
| 1480 | init(bindings); |
| 1481 | } |
| 1482 | |
| 1483 | SqliteDatabase::Query::Query(SqliteDatabase& db, |
| 1484 | QueryOptions options, |
| 1485 | kj::StringPtr sqlCode, |
| 1486 | kj::ArrayPtr<const ValuePtr> bindings) |
| 1487 | : ResetListener(db), |
| 1488 | regulator(options.regulator), |
| 1489 | ownStatement(db.prepareSql(regulator, sqlCode, 0, MULTI)), |
| 1490 | maybeStatement(ownStatement), |
| 1491 | queryEvent(this->db.sqliteObserver) { |
| 1492 | // If we throw from the constructor, the destructor won't run. Need to call destroy() explicitly. |
| 1493 | KJ_ON_SCOPE_FAILURE(destroy()); |
| 1494 | init(bindings); |
| 1495 | } |
| 1496 | |
| 1497 | SqliteDatabase::Query::~Query() noexcept(false) { |
| 1498 | destroy(); |
| 1499 | } |
| 1500 | |
| 1501 | void SqliteDatabase::Query::destroy() { |
| 1502 | // Install memory scope for sqlite3_reset, sqlite3_clear_bindings, and sqlite3_finalize (via |
| 1503 | // ownStatement destruction). The scope is idempotent, so this is safe even if a scope is already |
| 1504 | // active from the caller. |
| 1505 | auto memoryScope = db.enterMemoryScope(); |
| 1506 | |
| 1507 | if (regulator.shouldAddQueryStats()) { |
| 1508 | // Update the db stats that we have collected for the query. |
| 1509 | db.sqliteObserver.addQueryStats(rowsRead, rowsWritten); |
| 1510 | } |
| 1511 | |
| 1512 | queryEvent.setQueryEventStats(rowsRead, rowsWritten, !(regulator.shouldAddQueryStats())); |
| 1513 | |
| 1514 | try { |
| 1515 | kj::StringPtr statement = sqlite3_sql(getStatementAndEffect().statement); |
| 1516 | queryEvent.setQueryStatement( |
| 1517 | kj::heapString(statement.slice(0, kj::min(statement.size(), RA_MAX_METRICS_QUERY_SIZE)))); |
| 1518 | } catch (kj::Exception& e) { |
| 1519 | kj::StringPtr errorDescription = e.getDescription(); |
| 1520 | queryEvent.setQueryErrorDescription(kj::heapString( |
| 1521 | errorDescription.slice(0, kj::min(RA_MAX_METRICS_QUERY_SIZE, errorDescription.size())))); |
| 1522 | } |
| 1523 | |
| 1524 | // Move ownStatement to a local variable so that it goes out of scope while memoryScope is still |
| 1525 | // active. ownStatement is used by getStatementAndEffect(), so we cannot move it until this |
| 1526 | // point. |
| 1527 | auto ownStatementToDestroy = kj::mv(ownStatement); |
| 1528 | |
| 1529 | // We only need to reset the statement if we don't own it. If we own it, it's about to be |
| 1530 | // destroyed anyway. |
| 1531 | if (ownStatementToDestroy.statement.get() == nullptr) { |
| 1532 | KJ_IF_SOME(statement, maybeStatement) { |
| 1533 | // The error code returned by sqlite3_reset() actually represents the last error encountered |
| 1534 | // when stepping the statement. This doesn't mean that the reset failed. |
| 1535 | sqlite3_reset(statement.statement); |
| 1536 | |
| 1537 | // sqlite3_clear_bindings() returns int, but there is no documentation on how the return code |
| 1538 | // should be interpreted, so we ignore it. |
| 1539 | sqlite3_clear_bindings(statement.statement); |
| 1540 | |
| 1541 | // Reset the rows read/written counters. |
| 1542 | sqlite3_stmt_status(statement.statement, LIBSQL_STMTSTATUS_ROWS_READ, 1); |
| 1543 | sqlite3_stmt_status(statement.statement, LIBSQL_STMTSTATUS_ROWS_WRITTEN, 1); |
| 1544 | } |
| 1545 | } |
| 1546 | } |
| 1547 | |
| 1548 | void SqliteDatabase::Query::checkRequirements(size_t size) { |
| 1549 | if (regulator.shouldAddQueryStats()) { |
| 1550 | KJ_IF_SOME(actorAccountLimits, db.actorAccountLimits) { |
| 1551 | actorAccountLimits.requireActorCanExecuteQueries(); |
| 1552 | } |
| 1553 | } |
| 1554 | |
| 1555 | sqlite3_stmt* statement = getStatement(); |
| 1556 | |
| 1557 | SQLITE_REQUIRE(!sqlite3_stmt_busy(statement), kj::none, |
| 1558 | "A SQL prepared statement can only be executed once at a time."); |
| 1559 | SQLITE_REQUIRE(size == sqlite3_bind_parameter_count(statement), kj::none, |
| 1560 | "Wrong number of parameter bindings for SQL query."); |
| 1561 | |
| 1562 | KJ_IF_SOME(cb, db.onWriteCallback) { |
| 1563 | if (!sqlite3_stmt_readonly(statement)) { |
| 1564 | cb(allowUnconfirmed); |
| 1565 | } |
| 1566 | } |
| 1567 | } |
| 1568 | |
| 1569 | void SqliteDatabase::Query::init(kj::ArrayPtr<const ValuePtr> bindings) { |
| 1570 | checkRequirements(bindings.size()); |
| 1571 | |
| 1572 | for (auto i: kj::indices(bindings)) { |
| 1573 | bind(i, bindings[i]); |
| 1574 | } |
| 1575 | |
| 1576 | nextRow(/*first=*/true); |
| 1577 | } |
| 1578 | |
| 1579 | void SqliteDatabase::Query::bind(uint i, ValuePtr value) { |
| 1580 | sqlite3_stmt* statement = getStatement(); |
| 1581 | |
| 1582 | KJ_SWITCH_ONEOF(value) { |
| 1583 | KJ_CASE_ONEOF(blob, kj::ArrayPtr<const byte>) { |
| 1584 | SQLITE_CALL(sqlite3_bind_blob(statement, i + 1, blob.begin(), blob.size(), SQLITE_STATIC)); |
| 1585 | } |
| 1586 | KJ_CASE_ONEOF(text, kj::StringPtr) { |
| 1587 | SQLITE_CALL(sqlite3_bind_text(statement, i + 1, text.begin(), text.size(), SQLITE_STATIC)); |
| 1588 | } |
| 1589 | KJ_CASE_ONEOF(n, int64_t) { |
| 1590 | SQLITE_CALL(sqlite3_bind_int64(statement, i + 1, static_cast<long long>(n))); |
| 1591 | } |
| 1592 | KJ_CASE_ONEOF(x, double) { |
| 1593 | SQLITE_CALL(sqlite3_bind_double(statement, i + 1, x)); |
| 1594 | } |
| 1595 | KJ_CASE_ONEOF(_, decltype(nullptr)) { |
| 1596 | SQLITE_CALL(sqlite3_bind_null(statement, i + 1)); |
| 1597 | } |
| 1598 | } |
| 1599 | } |
| 1600 | |
| 1601 | uint64_t SqliteDatabase::Query::getRowsRead() { |
| 1602 | sqlite3_stmt* statement = getStatement(); |
| 1603 | KJ_REQUIRE(statement != nullptr); |
| 1604 | return sqlite3_stmt_status(statement, LIBSQL_STMTSTATUS_ROWS_READ, 0); |
| 1605 | } |
| 1606 | |
| 1607 | uint64_t SqliteDatabase::Query::getRowsWritten() { |
| 1608 | sqlite3_stmt* statement = getStatement(); |
| 1609 | return sqlite3_stmt_status(statement, LIBSQL_STMTSTATUS_ROWS_WRITTEN, 0); |
| 1610 | } |
| 1611 | |
| 1612 | void SqliteDatabase::Query::bind(uint i, kj::ArrayPtr<const byte> value) { |
| 1613 | auto memoryScope = db.enterMemoryScope(); |
| 1614 | sqlite3_stmt* statement = getStatement(); |
| 1615 | SQLITE_CALL(sqlite3_bind_blob(statement, i + 1, value.begin(), value.size(), SQLITE_STATIC)); |
| 1616 | } |
| 1617 | |
| 1618 | void SqliteDatabase::Query::bind(uint i, kj::StringPtr value) { |
| 1619 | auto memoryScope = db.enterMemoryScope(); |
| 1620 | sqlite3_stmt* statement = getStatement(); |
| 1621 | SQLITE_CALL(sqlite3_bind_text(statement, i + 1, value.begin(), value.size(), SQLITE_STATIC)); |
| 1622 | } |
| 1623 | |
| 1624 | void SqliteDatabase::Query::bind(uint i, long long value) { |
| 1625 | auto memoryScope = db.enterMemoryScope(); |
| 1626 | sqlite3_stmt* statement = getStatement(); |
| 1627 | SQLITE_CALL(sqlite3_bind_int64(statement, i + 1, value)); |
| 1628 | } |
| 1629 | |
| 1630 | void SqliteDatabase::Query::bind(uint i, double value) { |
| 1631 | auto memoryScope = db.enterMemoryScope(); |
| 1632 | sqlite3_stmt* statement = getStatement(); |
| 1633 | SQLITE_CALL(sqlite3_bind_double(statement, i + 1, value)); |
| 1634 | } |
| 1635 | |
| 1636 | void SqliteDatabase::Query::bind(uint i, decltype(nullptr)) { |
| 1637 | auto memoryScope = db.enterMemoryScope(); |
| 1638 | sqlite3_stmt* statement = getStatement(); |
| 1639 | SQLITE_CALL(sqlite3_bind_null(statement, i + 1)); |
| 1640 | } |
| 1641 | |
| 1642 | void SqliteDatabase::Query::nextRow(bool first) { |
| 1643 | auto& statementAndEffect = getStatementAndEffect(); |
| 1644 | sqlite3_stmt* statement = statementAndEffect.statement; |
| 1645 | |
| 1646 | KJ_ASSERT(db.currentStatement == kj::none, "recursive nextRow()?"); |
| 1647 | KJ_DEFER(db.currentStatement = kj::none); |
| 1648 | db.currentStatement = *statement; |
| 1649 | |
| 1650 | // The statement could be "re-prepared" during sqlite3_step, so we must set up the regulator. |
| 1651 | KJ_ASSERT(db.currentRegulator == kj::none, "nextRow() during prepare()?"); |
| 1652 | KJ_DEFER(db.currentRegulator = kj::none); |
| 1653 | db.currentRegulator = regulator; |
| 1654 | |
| 1655 | auto memoryScope = db.enterMemoryScope(); |
| 1656 | SQLITE_CALL_SCOPE { |
| 1657 | int err = sqlite3_step(statement); |
| 1658 | queryEvent.setQueryResult(err); |
| 1659 | |
| 1660 | int extendedCode = sqlite3_extended_errcode(db); |
| 1661 | queryEvent.setQueryExtendedCode(extendedCode); |
| 1662 | |
| 1663 | // TODO(perf): This is slightly inefficient to call for every row read, but not bad enough to |
| 1664 | // fix it immediately. The alternate way would be to getRowsRead/Written once when we emit it |
| 1665 | // in the Dtor, and handle the case where the statement could be null when the Query gets |
| 1666 | // destructed |
| 1667 | rowsRead = getRowsRead(); |
| 1668 | rowsWritten = getRowsWritten(); |
| 1669 | if (err == SQLITE_DONE) { |
| 1670 | done = true; |
| 1671 | } else if (err != SQLITE_ROW) { |
| 1672 | SQLITE_CALL_FAILED("sqlite3_step()", err); |
| 1673 | } |
| 1674 | } |
| 1675 | |
| 1676 | if (first) { |
| 1677 | // A statement's effect is applied on the first step. |
| 1678 | db.applyChange(statementAndEffect.stateChange); |
| 1679 | } |
| 1680 | } |
| 1681 | |
| 1682 | uint SqliteDatabase::Query::changeCount() { |
| 1683 | KJ_REQUIRE(done); |
| 1684 | KJ_DREQUIRE( |
| 1685 | columnCount() == 0, "changeCount() can only be called on INSERT/UPDATE/DELETE queries"); |
| 1686 | return sqlite3_changes(db); |
| 1687 | } |
| 1688 | |
| 1689 | uint SqliteDatabase::Query::columnCount() { |
| 1690 | auto memoryScope = db.enterMemoryScope(); |
| 1691 | sqlite3_stmt* statement = getStatement(); |
| 1692 | return sqlite3_column_count(statement); |
| 1693 | } |
| 1694 | |
| 1695 | SqliteDatabase::Query::ValuePtr SqliteDatabase::Query::getValue(uint column) { |
| 1696 | auto memoryScope = db.enterMemoryScope(); |
| 1697 | sqlite3_stmt* statement = getStatement(); |
| 1698 | switch (sqlite3_column_type(statement, column)) { |
| 1699 | case SQLITE_INTEGER: |
| 1700 | return getInt64(column); |
| 1701 | case SQLITE_FLOAT: |
| 1702 | return getDouble(column); |
| 1703 | case SQLITE_TEXT: |
| 1704 | return getText(column); |
| 1705 | case SQLITE_BLOB: |
| 1706 | return getBlob(column); |
| 1707 | case SQLITE_NULL: |
| 1708 | return nullptr; |
| 1709 | } |
| 1710 | KJ_UNREACHABLE; |
| 1711 | } |
| 1712 | |
| 1713 | kj::StringPtr SqliteDatabase::Query::getColumnName(uint column) { |
| 1714 | auto memoryScope = db.enterMemoryScope(); |
| 1715 | sqlite3_stmt* statement = getStatement(); |
| 1716 | return sqlite3_column_name(statement, column); |
| 1717 | } |
| 1718 | |
| 1719 | kj::ArrayPtr<const byte> SqliteDatabase::Query::getBlob(uint column) { |
| 1720 | auto memoryScope = db.enterMemoryScope(); |
| 1721 | sqlite3_stmt* statement = getStatement(); |
| 1722 | const byte* ptr = reinterpret_cast<const byte*>(sqlite3_column_blob(statement, column)); |
| 1723 | return kj::arrayPtr(ptr, sqlite3_column_bytes(statement, column)); |
| 1724 | } |
| 1725 | |
| 1726 | kj::StringPtr SqliteDatabase::Query::getText(uint column) { |
| 1727 | auto memoryScope = db.enterMemoryScope(); |
| 1728 | sqlite3_stmt* statement = getStatement(); |
| 1729 | const char* ptr = reinterpret_cast<const char*>(sqlite3_column_text(statement, column)); |
| 1730 | return kj::StringPtr(ptr, sqlite3_column_bytes(statement, column)); |
| 1731 | } |
| 1732 | |
| 1733 | int SqliteDatabase::Query::getInt(uint column) { |
| 1734 | auto memoryScope = db.enterMemoryScope(); |
| 1735 | sqlite3_stmt* statement = getStatement(); |
| 1736 | return sqlite3_column_int(statement, column); |
| 1737 | } |
| 1738 | |
| 1739 | int64_t SqliteDatabase::Query::getInt64(uint column) { |
| 1740 | auto memoryScope = db.enterMemoryScope(); |
| 1741 | sqlite3_stmt* statement = getStatement(); |
| 1742 | return sqlite3_column_int64(statement, column); |
| 1743 | } |
| 1744 | |
| 1745 | double SqliteDatabase::Query::getDouble(uint column) { |
| 1746 | auto memoryScope = db.enterMemoryScope(); |
| 1747 | sqlite3_stmt* statement = getStatement(); |
| 1748 | return sqlite3_column_double(statement, column); |
| 1749 | } |
| 1750 | |
| 1751 | bool SqliteDatabase::Query::isNull(uint column) { |
| 1752 | auto memoryScope = db.enterMemoryScope(); |
| 1753 | sqlite3_stmt* statement = getStatement(); |
| 1754 | return sqlite3_column_type(statement, column) == SQLITE_NULL; |
| 1755 | } |
| 1756 | |
| 1757 | SqliteDatabase::StatementAndEffect& SqliteDatabase::Query::getStatementAndEffect() { |
| 1758 | return KJ_UNWRAP_OR(maybeStatement, { |
| 1759 | regulator.onError(kj::none, "SQLite query was canceled because the database was deleted."); |
| 1760 | KJ_FAIL_REQUIRE("query canceled because reset() was called on the database"); |
| 1761 | }); |
| 1762 | } |
| 1763 | |
| 1764 | void SqliteDatabase::Query::beforeSqliteReset() { |
| 1765 | // Note that if we don't own the statement, then `maybeStatement` is probably already dangling |
| 1766 | // here. Luckily, we don't need to reset it or anything because the statement will be destroyed |
| 1767 | // by Statement::beforeSqliteReset(). |
| 1768 | maybeStatement = kj::none; |
| 1769 | ownStatement = {}; |
| 1770 | } |
| 1771 | |
| 1772 | // ======================================================================================= |
| 1773 | // VFS |
| 1774 | |
| 1775 | // ----------------------------------------------------------------------------- |
| 1776 | // Code to wrap SQLite's native VFS so that it can be rooted in some `kj::Directory`, where that |
| 1777 | // directory points at a real disk directory. |
| 1778 | // |
| 1779 | // A native disk `kj::Directory` -- at least on Unix -- wraps an open file descriptor, pointing at |
| 1780 | // a directory. It does NOT keep track of the directory's path on disk. In fact, the directory can |
| 1781 | // be moved or renamed, and `kj::Directory` will continue to point at it. |
| 1782 | // |
| 1783 | // There is no portable way to query the current path of a directory. In order to open files within |
| 1784 | // a directory given only the directory descriptor, you must use syscalls like `openat()`, which |
| 1785 | // take a directory file descriptor to use as the root. |
| 1786 | // |
| 1787 | // SQLite's native VFS, however, is not openat()-aware. Luckily, it _does_ provide the ability to |
| 1788 | // redirect its syscalls to custom implementations. So we can intercept `open()` and make it use |
| 1789 | // `openat()` instead! With a little thread-local hackery, we can make sure to use the desired root |
| 1790 | // directory descriptor from the `kj::Directory`. |
| 1791 | // |
| 1792 | // Of course, SQLite also lets us virtualize the whole filesystem at a higher level. Why go to all |
| 1793 | // the bother to hack it at a low level rather than just implement an entire VFS based on the |
| 1794 | // `kj::Directory` interface? The problem is, SQLite's native VFS contains a ton of code to handle |
| 1795 | // all sorts of corner cases and do things just right. When our files are actually on real disk, |
| 1796 | // we want to leverage all that code. If we can just make it interpret paths differently, then we |
| 1797 | // can reuse the rest of the implementation. |
| 1798 | |
| 1799 | #if !_WIN32 |
| 1800 | namespace { |
| 1801 | |
| 1802 | static thread_local int currentVfsRoot = AT_FDCWD; |
| 1803 | // We will tell SQLite to use alternate implementations of path-oriented syscalls which use the |
| 1804 | // `*at()` versions of the calls with `currentVfsRoot` as the directory descriptor. When the |
| 1805 | // descriptor is `AT_FDCWD`, this will naturally reproduce the behavior of the non-`at()` versions. |
| 1806 | // We temporarily swap this for a real descriptor when our custom VFS wrapper is being invoked. |
| 1807 | |
| 1808 | static int replaced_open(const char* path, int flags, int mode) { |
| 1809 | return openat(currentVfsRoot, path, flags, mode); |
| 1810 | } |
| 1811 | static int replaced_access(const char* path, int type) { |
| 1812 | return faccessat(currentVfsRoot, path, type, 0); |
| 1813 | } |
| 1814 | static char* replaced_getcwd(char* buf, size_t size) noexcept { |
| 1815 | KJ_REQUIRE(currentVfsRoot == AT_FDCWD, |
| 1816 | "SQLite custom VFS shouldn't call getcwd() because we overrode xFullPathname"); |
| 1817 | return getcwd(buf, size); |
| 1818 | } |
| 1819 | static int replaced_stat(const char* path, struct stat* stats) { |
| 1820 | return fstatat(currentVfsRoot, path, stats, 0); |
| 1821 | } |
| 1822 | static int replaced_unlink(const char* path) { |
| 1823 | return unlinkat(currentVfsRoot, path, 0); |
| 1824 | } |
| 1825 | static int replaced_mkdir(const char* path, mode_t mode) { |
| 1826 | return mkdirat(currentVfsRoot, path, mode); |
| 1827 | } |
| 1828 | static int replaced_rmdir(const char* path) { |
| 1829 | return unlinkat(currentVfsRoot, path, AT_REMOVEDIR); |
| 1830 | } |
| 1831 | static ssize_t replaced_readlink(const char* path, char* buf, size_t len) { |
| 1832 | return readlinkat(currentVfsRoot, path, buf, len); |
| 1833 | } |
| 1834 | static int replaced_lstat(const char* path, struct stat* stats) { |
| 1835 | return fstatat(currentVfsRoot, path, stats, AT_SYMLINK_NOFOLLOW); |
| 1836 | } |
| 1837 | |
| 1838 | }; // namespace |
| 1839 | |
| 1840 | // The sqlite3_file implementation we use when wrapping the native filesystem. |
| 1841 | struct SqliteDatabase::Vfs::WrappedNativeFileImpl: public sqlite3_file { |
| 1842 | const Vfs* vfs; |
| 1843 | int rootFd; |
| 1844 | |
| 1845 | // It's expected that the wrapped sqlite_file begins in memory immediately after this object. |
| 1846 | sqlite3_file* getWrapped() { |
| 1847 | return reinterpret_cast<sqlite3_file*>(this + 1); |
| 1848 | } |
| 1849 | |
| 1850 | static const sqlite3_io_methods METHOD_TABLE; |
| 1851 | }; |
| 1852 | |
| 1853 | // This completely nutso template generates wrapper functions for each of the function pointer |
| 1854 | // members of sqlite3_vfs. The wrapper function temporarily sets `currentVfsRoot` to the FD |
| 1855 | // of the directory from the SqliteDatabase::Vfs instance in use, then invokes the same function |
| 1856 | // on the underlying native VFS. |
| 1857 | template <typename Result, |
| 1858 | typename... Params, |
| 1859 | Result (*sqlite3_vfs::*slot)(sqlite3_vfs* vfs, Params...)> |
| 1860 | struct SqliteDatabase::Vfs::MethodWrapperHack<Result (*sqlite3_vfs::*)(sqlite3_vfs* vfs, Params...), |
| 1861 | slot> { |
| 1862 | static Result wrapper(sqlite3_vfs* vfs, Params... params) noexcept { |
| 1863 | auto& self = *reinterpret_cast<SqliteDatabase::Vfs*>(vfs->pAppData); |
| 1864 | KJ_ASSERT(currentVfsRoot == AT_FDCWD); |
| 1865 | currentVfsRoot = self.rootFd; |
| 1866 | KJ_DEFER(currentVfsRoot = AT_FDCWD); |
| 1867 | return (self.native.*slot)(&self.native, params...); |
| 1868 | } |
| 1869 | }; |
| 1870 | |
| 1871 | // Specialization of MethodWrapperHack for wrapping methods of sqlite_file, aka |
| 1872 | // sqlite3_io_methods. Unfortunately, some file methods go back and perform filesystem ops. In |
| 1873 | // particular, accessing shared memory associated with a file actually opens another adjacent |
| 1874 | // file. |
| 1875 | template <typename Result, |
| 1876 | typename... Params, |
| 1877 | Result (*sqlite3_io_methods::*slot)(sqlite3_file* file, Params...)> |
| 1878 | struct SqliteDatabase::Vfs:: |
| 1879 | MethodWrapperHack<Result (*sqlite3_io_methods::*)(sqlite3_file* file, Params...), slot> { |
| 1880 | static Result wrapper(sqlite3_file* file, Params... params) noexcept { |
| 1881 | auto wrapper = static_cast<WrappedNativeFileImpl*>(file); |
| 1882 | file = wrapper->getWrapped(); |
| 1883 | KJ_ASSERT(currentVfsRoot == AT_FDCWD); |
| 1884 | currentVfsRoot = wrapper->rootFd; |
| 1885 | KJ_DEFER(currentVfsRoot = AT_FDCWD); |
| 1886 | return (file->pMethods->*slot)(file, params...); |
| 1887 | } |
| 1888 | }; |
| 1889 | |
| 1890 | // clang-format off |
| 1891 | // |
| 1892 | // The code below has a lot of lambdas inside struct initializers, which clang-format does not |
| 1893 | // handle well, making it extremely hard to read if we leave the formatter on. |
| 1894 | |
| 1895 | const sqlite3_io_methods SqliteDatabase::Vfs::WrappedNativeFileImpl::METHOD_TABLE = { |
| 1896 | .iVersion = 3, |
| 1897 | |
| 1898 | #define WRAP(name) \ |
| 1899 | .name = \ |
| 1900 | &MethodWrapperHack<decltype(&sqlite3_io_methods::name), &sqlite3_io_methods::name>::wrapper |
| 1901 | |
| 1902 | WRAP(xClose), |
| 1903 | WRAP(xRead), |
| 1904 | WRAP(xWrite), |
| 1905 | WRAP(xTruncate), |
| 1906 | WRAP(xSync), |
| 1907 | WRAP(xFileSize), |
| 1908 | WRAP(xLock), |
| 1909 | WRAP(xUnlock), |
| 1910 | WRAP(xCheckReservedLock), |
| 1911 | WRAP(xFileControl), |
| 1912 | WRAP(xSectorSize), |
| 1913 | .xDeviceCharacteristics = [](sqlite3_file* file) noexcept -> int { |
| 1914 | auto wrapper = static_cast<WrappedNativeFileImpl*>(file); |
| 1915 | file = wrapper->getWrapped(); |
| 1916 | KJ_ASSERT(currentVfsRoot == AT_FDCWD); |
| 1917 | currentVfsRoot = wrapper->rootFd; |
| 1918 | KJ_DEFER(currentVfsRoot = AT_FDCWD); |
| 1919 | return (file->pMethods->xDeviceCharacteristics)(file) | |
| 1920 | wrapper->vfs->options.deviceCharacteristics; |
| 1921 | }, |
| 1922 | |
| 1923 | WRAP(xShmMap), |
| 1924 | WRAP(xShmLock), |
| 1925 | WRAP(xShmBarrier), |
| 1926 | WRAP(xShmUnmap), |
| 1927 | |
| 1928 | WRAP(xFetch), |
| 1929 | WRAP(xUnfetch), |
| 1930 | #undef WRAP |
| 1931 | }; |
| 1932 | |
| 1933 | // The native VFS gives us the ability to override its syscalls. We need to do so, in |
| 1934 | // particular to force them to use the *at() versions of the calls that accept a directory FD |
| 1935 | // to use as the root. |
| 1936 | // |
| 1937 | // Unfortunately, these overrides are global for the process, with no ability to pass down any |
| 1938 | // context to them. So, we stash the current root FD in `currentVfsRoot` whenever we call into |
| 1939 | // the native VFS. We also don't want to interfere with anything else in the process that is |
| 1940 | // using SQLite directly, so we make sure that when we're not specifically trying to invoke |
| 1941 | // our wrapper, then `currentVfsRoot` is `AT_FDCWD`, which causes the *at() syscalls to match |
| 1942 | // their non-at() versions. |
| 1943 | sqlite3_vfs SqliteDatabase::Vfs::makeWrappedNativeVfs() { |
| 1944 | static bool registerOnce KJ_UNUSED = ([&]() { |
| 1945 | #define REPLACE_SYSCALL(name) \ |
| 1946 | native.xSetSystemCall(&native, #name, (sqlite3_syscall_ptr)replaced_##name); |
| 1947 | REPLACE_SYSCALL(open); |
| 1948 | REPLACE_SYSCALL(access); |
| 1949 | REPLACE_SYSCALL(getcwd); |
| 1950 | REPLACE_SYSCALL(stat); |
| 1951 | REPLACE_SYSCALL(unlink); |
| 1952 | REPLACE_SYSCALL(mkdir); |
| 1953 | REPLACE_SYSCALL(rmdir); |
| 1954 | REPLACE_SYSCALL(readlink); |
| 1955 | REPLACE_SYSCALL(lstat); |
| 1956 | #undef REPLACE_SYSCALL |
| 1957 | |
| 1958 | return true; |
| 1959 | })(); |
| 1960 | |
| 1961 | // We construct a sqlite3_vfs that is basically a copy of the native VFS, except each method is |
| 1962 | // wrapped so that it sets `currentVfsRoot` while running. |
| 1963 | return { |
| 1964 | .iVersion = kj::min(3, native.iVersion), |
| 1965 | .szOsFile = native.szOsFile + static_cast<int>(sizeof(WrappedNativeFileImpl)), |
| 1966 | .mxPathname = native.mxPathname, |
| 1967 | .pNext = nullptr, |
| 1968 | .zName = name.cStr(), |
| 1969 | .pAppData = this, |
| 1970 | |
| 1971 | .xOpen = [](sqlite3_vfs* vfs, sqlite3_filename zName, sqlite3_file* file, int flags, |
| 1972 | int* pOutFlags) -> int { |
| 1973 | // We have to wrap xOpen explicitly because we need to further wrap each created file. |
| 1974 | // |
| 1975 | // My trick here is to prefix the native file with a second vtable. So the layout of the |
| 1976 | // `sqlite3_file` that we construct is actually a simple `struct sqlite_file` (which just |
| 1977 | // contains a single pointer to sqlite3_io_methods, i.e. the vtable pointer) _followed by_ |
| 1978 | // the regular native file structure. |
| 1979 | |
| 1980 | auto wrapper = static_cast<WrappedNativeFileImpl*>(file); |
| 1981 | file = wrapper->getWrapped(); |
| 1982 | file->pMethods = nullptr; |
| 1983 | |
| 1984 | // Set up currentVfsRoot. |
| 1985 | auto& self = *reinterpret_cast<const SqliteDatabase::Vfs*>(vfs->pAppData); |
| 1986 | KJ_ASSERT(currentVfsRoot == AT_FDCWD); |
| 1987 | currentVfsRoot = self.rootFd; |
| 1988 | KJ_DEFER(currentVfsRoot = AT_FDCWD); |
| 1989 | |
| 1990 | int result = self.native.xOpen(&self.native, zName, file, flags, pOutFlags); |
| 1991 | |
| 1992 | // `xOpen` setting `pMethods` to non-null indicates that `xClose` is needed, i.e. the file |
| 1993 | // has been constructed. We need our wrapper to match. |
| 1994 | if (file->pMethods == nullptr) { |
| 1995 | wrapper->pMethods = nullptr; |
| 1996 | } else { |
| 1997 | wrapper->pMethods = &WrappedNativeFileImpl::METHOD_TABLE; |
| 1998 | wrapper->vfs = &self; |
| 1999 | wrapper->rootFd = self.rootFd; |
| 2000 | } |
| 2001 | |
| 2002 | return result; |
| 2003 | }, |
| 2004 | |
| 2005 | #define WRAP(name) \ |
| 2006 | .name = &MethodWrapperHack<decltype(&sqlite3_vfs::name), &sqlite3_vfs::name>::wrapper |
| 2007 | |
| 2008 | WRAP(xDelete), |
| 2009 | WRAP(xAccess), |
| 2010 | .xFullPathname = [](sqlite3_vfs*, const char* zName, int nOut, char* zOut) -> int { |
| 2011 | // Override xFullPathname so that it doesn't rewrite the path at all. |
| 2012 | size_t len = kj::min(strlen(zName), nOut - 1); |
| 2013 | memcpy(zOut, zName, len); |
| 2014 | zOut[len] = 0; |
| 2015 | return SQLITE_OK; |
| 2016 | }, |
| 2017 | |
| 2018 | .xDlOpen = nullptr, |
| 2019 | .xDlError = nullptr, |
| 2020 | .xDlSym = nullptr, |
| 2021 | .xDlClose = nullptr, |
| 2022 | // There is no dlopenat(), but we don't need to support these anyway. |
| 2023 | |
| 2024 | WRAP(xRandomness), |
| 2025 | WRAP(xSleep), |
| 2026 | WRAP(xCurrentTime), |
| 2027 | WRAP(xGetLastError), |
| 2028 | WRAP(xCurrentTimeInt64), |
| 2029 | |
| 2030 | .xSetSystemCall = nullptr, |
| 2031 | .xGetSystemCall = nullptr, |
| 2032 | .xNextSystemCall = nullptr, |
| 2033 | // We don't support further overriding syscalls. |
| 2034 | #undef WRAP |
| 2035 | }; |
| 2036 | } |
| 2037 | #endif // #if !_WIN32 |
| 2038 | |
| 2039 | // ----------------------------------------------------------------------------- |
| 2040 | // Code to implement a true SQLite VFS based on `kj::Directory`. |
| 2041 | // |
| 2042 | // This VFS implementation actually delegates to the KJ filesystem interface for everything. |
| 2043 | // This is used only when given a `kj::Directory` that does NOT represent a native file, i.e. |
| 2044 | // one where `getFd()` returns null. This is mainly used for unit tests which want to use in-memory |
| 2045 | // directories. |
| 2046 | |
| 2047 | // Implementation of sqlite3_file. |
| 2048 | // |
| 2049 | // Weirdly, for sqlite3_file, SQLite uses a C++-like inheritance approach, with a separate |
| 2050 | // virtual table that can be shared among all files of the same type. This is different from the |
| 2051 | // way sqlite3_vfs works, where the function pointers are inlined into the sqlite3_vfs struct. |
| 2052 | // In any case, as a result, `FileImpl`, unlike `VfsImpl`, is NOT just a namespace struct, but |
| 2053 | // an actual instance. |
| 2054 | struct SqliteDatabase::Vfs::FileImpl: public sqlite3_file { |
| 2055 | const Vfs& vfs; |
| 2056 | kj::Maybe<const kj::File&> writableFile; |
| 2057 | kj::Own<const kj::ReadableFile> file; |
| 2058 | |
| 2059 | kj::Maybe<kj::Own<Lock>> lock; |
| 2060 | // Rather complicatedly, SQLite doesn't consider the -shm file to be a separate file that it |
| 2061 | // opens via the VFS, but rather a facet of the database file itself. We implement it using an |
| 2062 | // entirely different interface anyway. |
| 2063 | // |
| 2064 | // We leave this null if the file is not the main database file. |
| 2065 | |
| 2066 | FileImpl(const Vfs& vfs, kj::Own<const kj::File> file, kj::Maybe<kj::Own<Lock>> lock) |
| 2067 | : sqlite3_file{.pMethods = &FILE_METHOD_TABLE}, |
| 2068 | vfs(vfs), |
| 2069 | writableFile(*file), |
| 2070 | file(kj::mv(file)), |
| 2071 | lock(kj::mv(lock)) {} |
| 2072 | FileImpl(const Vfs& vfs, kj::Own<const kj::ReadableFile> file, kj::Maybe<kj::Own<Lock>> lock) |
| 2073 | : sqlite3_file{.pMethods = &FILE_METHOD_TABLE}, |
| 2074 | vfs(vfs), |
| 2075 | file(kj::mv(file)), |
| 2076 | lock(kj::mv(lock)) {} |
| 2077 | |
| 2078 | static const sqlite3_io_methods FILE_METHOD_TABLE; |
| 2079 | }; |
| 2080 | |
| 2081 | const sqlite3_io_methods SqliteDatabase::Vfs::FileImpl::FILE_METHOD_TABLE = { |
| 2082 | .iVersion = 3, |
| 2083 | #define WRAP_METHOD(errorCode, block) \ |
| 2084 | auto& self KJ_UNUSED = *static_cast<FileImpl*>(file); \ |
| 2085 | try block catch (kj::Exception& e) { \ |
| 2086 | reportVfsErrorCaught(kj::mv(e)); \ |
| 2087 | return errorCode; \ |
| 2088 | } |
| 2089 | |
| 2090 | .xClose = [](sqlite3_file* file) noexcept -> int { |
| 2091 | WRAP_METHOD(SQLITE_OK, { |
| 2092 | auto& self = *static_cast<FileImpl*>(file); |
| 2093 | |
| 2094 | // Caller will free the object's memory, but knows nothing of destructors. |
| 2095 | kj::dtor(self); |
| 2096 | |
| 2097 | return SQLITE_OK; // return value is ignored by SQLite |
| 2098 | }); |
| 2099 | }, |
| 2100 | |
| 2101 | .xRead = [](sqlite3_file* file, void* buffer, int iAmt, sqlite3_int64 iOfst) noexcept -> int { |
| 2102 | WRAP_METHOD(SQLITE_IOERR_READ, { |
| 2103 | auto bytes = kj::arrayPtr(reinterpret_cast<byte*>(buffer), iAmt); |
| 2104 | size_t actual = self.file->read(iOfst, bytes); |
| 2105 | |
| 2106 | if (actual < iAmt) { |
| 2107 | bytes.slice(actual).fill(0); |
| 2108 | return SQLITE_IOERR_SHORT_READ; |
| 2109 | } else { |
| 2110 | return SQLITE_OK; |
| 2111 | } |
| 2112 | }); |
| 2113 | }, |
| 2114 | |
| 2115 | .xWrite = |
| 2116 | [](sqlite3_file* file, const void* buffer, int iAmt, sqlite3_int64 iOfst) noexcept -> int { |
| 2117 | WRAP_METHOD(SQLITE_IOERR_WRITE, { |
| 2118 | KJ_IF_SOME(writableFile, self.writableFile) { |
| 2119 | auto bytes = kj::arrayPtr(reinterpret_cast<const byte*>(buffer), iAmt); |
| 2120 | writableFile.write(iOfst, bytes); |
| 2121 | return SQLITE_OK; |
| 2122 | } else { |
| 2123 | return SQLITE_READONLY; |
| 2124 | } |
| 2125 | }); |
| 2126 | }, |
| 2127 | |
| 2128 | .xTruncate = [](sqlite3_file* file, sqlite3_int64 size) noexcept -> int { |
| 2129 | WRAP_METHOD(SQLITE_IOERR_TRUNCATE, { |
| 2130 | KJ_IF_SOME(writableFile, self.writableFile) { |
| 2131 | writableFile.truncate(size); |
| 2132 | return SQLITE_OK; |
| 2133 | } else { |
| 2134 | return SQLITE_READONLY; |
| 2135 | } |
| 2136 | }); |
| 2137 | }, |
| 2138 | |
| 2139 | .xSync = [](sqlite3_file* file, int flags) noexcept -> int { |
| 2140 | WRAP_METHOD(SQLITE_IOERR_FSYNC, { |
| 2141 | if (flags & SQLITE_SYNC_DATAONLY) { |
| 2142 | self.file->datasync(); |
| 2143 | } else { |
| 2144 | self.file->sync(); |
| 2145 | } |
| 2146 | return SQLITE_OK; |
| 2147 | }); |
| 2148 | }, |
| 2149 | |
| 2150 | .xFileSize = [](sqlite3_file* file, sqlite3_int64* pSize) noexcept -> int { |
| 2151 | WRAP_METHOD(SQLITE_IOERR_FSTAT, { |
| 2152 | *pSize = self.file->stat().size; |
| 2153 | return SQLITE_OK; |
| 2154 | }); |
| 2155 | }, |
| 2156 | |
| 2157 | .xLock = [](sqlite3_file* file, int level) noexcept -> int { |
| 2158 | // Verify that our enum's values match the SQLite constants. (We didn't want to include |
| 2159 | // sqlite3.h in our header, so defined a parallel enum.) |
| 2160 | static_assert(Lock::UNLOCKED == SQLITE_LOCK_NONE); |
| 2161 | static_assert(Lock::SHARED == SQLITE_LOCK_SHARED); |
| 2162 | static_assert(Lock::RESERVED == SQLITE_LOCK_RESERVED); |
| 2163 | static_assert(Lock::PENDING == SQLITE_LOCK_PENDING); |
| 2164 | static_assert(Lock::EXCLUSIVE == SQLITE_LOCK_EXCLUSIVE); |
| 2165 | |
| 2166 | WRAP_METHOD(SQLITE_IOERR_LOCK, { |
| 2167 | auto& lock = *KJ_ASSERT_NONNULL(self.lock, "xLock called on file that isn't main database?"); |
| 2168 | if (lock.tryIncreaseLevel(static_cast<Lock::Level>(level))) { |
| 2169 | return SQLITE_OK; |
| 2170 | } else { |
| 2171 | return SQLITE_BUSY; |
| 2172 | } |
| 2173 | }); |
| 2174 | }, |
| 2175 | |
| 2176 | .xUnlock = [](sqlite3_file* file, int level) noexcept -> int { |
| 2177 | WRAP_METHOD(SQLITE_IOERR_UNLOCK, { |
| 2178 | auto& lock = *KJ_ASSERT_NONNULL(self.lock, "xLock called on file that isn't main database?"); |
| 2179 | lock.decreaseLevel(static_cast<Lock::Level>(level)); |
| 2180 | return SQLITE_OK; |
| 2181 | }); |
| 2182 | }, |
| 2183 | |
| 2184 | .xCheckReservedLock = [](sqlite3_file* file, int* pResOut) noexcept -> int { |
| 2185 | WRAP_METHOD(SQLITE_IOERR_CHECKRESERVEDLOCK, { |
| 2186 | auto& lock = *KJ_ASSERT_NONNULL(self.lock, "xLock called on file that isn't main database?"); |
| 2187 | *pResOut = lock.checkReservedLock(); |
| 2188 | return SQLITE_OK; |
| 2189 | }); |
| 2190 | }, |
| 2191 | |
| 2192 | .xFileControl = [](sqlite3_file* file, int op, void* pArg) noexcept -> int { |
| 2193 | // Apparently we can return SQLITE_NOTFOUND for controls we don't implement. |
| 2194 | return SQLITE_NOTFOUND; |
| 2195 | }, |
| 2196 | .xSectorSize = [](sqlite3_file* file) noexcept -> int { |
| 2197 | // This function doesn't return a status code, it returns the size. It's largely a performance |
| 2198 | // hint, I think. For in-memory file systems, it has no real meaning. 4096 is the value of |
| 2199 | // SQLITE_DEFAULT_SECTOR_SIZE in the SQLite codebase, though the comments also say the result |
| 2200 | // is "almost always 512". |
| 2201 | return 4096; |
| 2202 | }, |
| 2203 | .xDeviceCharacteristics = [](sqlite3_file* file) noexcept -> int { |
| 2204 | WRAP_METHOD(SQLITE_IOERR, { return self.vfs.options.deviceCharacteristics; }); |
| 2205 | }, |
| 2206 | |
| 2207 | .xShmMap = |
| 2208 | [](sqlite3_file* file, int iRegion, int szRegion, int bExtend, void volatile** pp) noexcept |
| 2209 | -> int { |
| 2210 | WRAP_METHOD(SQLITE_IOERR_SHMMAP, { |
| 2211 | KJ_ASSERT(iRegion >= 0); |
| 2212 | KJ_ASSERT(szRegion >= 0); |
| 2213 | auto& lock = *KJ_ASSERT_NONNULL(self.lock, "xShmMap called on file that isn't main database?"); |
| 2214 | |
| 2215 | auto bytes = lock.getSharedMemoryRegion(iRegion, szRegion, bExtend); |
| 2216 | if (bytes == nullptr) { |
| 2217 | *pp = nullptr; |
| 2218 | } else { |
| 2219 | *pp = bytes.begin(); |
| 2220 | } |
| 2221 | return SQLITE_OK; |
| 2222 | }); |
| 2223 | }, |
| 2224 | .xShmLock = [](sqlite3_file* file, int offset, int n, int flags) noexcept -> int { |
| 2225 | WRAP_METHOD(SQLITE_IOERR_SHMLOCK, { |
| 2226 | auto& lock = *KJ_ASSERT_NONNULL(self.lock, "xShmMap called on file that isn't main database?"); |
| 2227 | if (flags & SQLITE_SHM_LOCK) { |
| 2228 | if (flags & SQLITE_SHM_EXCLUSIVE) { |
| 2229 | if (!lock.tryLockWalExclusive(offset, n)) return SQLITE_BUSY; |
| 2230 | } else { |
| 2231 | KJ_ASSERT(flags & SQLITE_SHM_SHARED); |
| 2232 | if (!lock.tryLockWalShared(offset, n)) return SQLITE_BUSY; |
| 2233 | } |
| 2234 | } else { |
| 2235 | KJ_ASSERT(flags & SQLITE_SHM_UNLOCK); |
| 2236 | if (flags & SQLITE_SHM_EXCLUSIVE) { |
| 2237 | lock.unlockWalExclusive(offset, n); |
| 2238 | } else { |
| 2239 | KJ_ASSERT(flags & SQLITE_SHM_SHARED); |
| 2240 | lock.unlockWalShared(offset, n); |
| 2241 | } |
| 2242 | } |
| 2243 | return SQLITE_OK; |
| 2244 | }); |
| 2245 | }, |
| 2246 | .xShmBarrier = [](sqlite3_file*) noexcept -> void { |
| 2247 | // I don't quite get why this is virtualized. The native implementation does |
| 2248 | // __sync_synchronize() (equivalent to below, I think) and also "for redundancy" locks and |
| 2249 | // unlocks a mutex. |
| 2250 | std::atomic_thread_fence(std::memory_order_acq_rel); |
| 2251 | }, |
| 2252 | .xShmUnmap = [](sqlite3_file* file, int deleteFlag) noexcept -> int { |
| 2253 | WRAP_METHOD(SQLITE_OK, { |
| 2254 | auto& lock = *KJ_ASSERT_NONNULL(self.lock, "xShmMap called on file that isn't main database?"); |
| 2255 | if (deleteFlag) { |
| 2256 | lock.clearSharedMemory(); |
| 2257 | } |
| 2258 | return SQLITE_OK; // return value is ignored by sqlite |
| 2259 | }); |
| 2260 | }, |
| 2261 | |
| 2262 | .xFetch = [](sqlite3_file* file, sqlite3_int64 iOfst, int iAmt, void** pp) noexcept -> int { |
| 2263 | // This is essentially requesting an mmap(). kj::File supports mmap(). Great, right? |
| 2264 | // |
| 2265 | // Well, there's a problem. We mostly use this VFS implementation to wrap an in-memory |
| 2266 | // `kj::File`. Such files support mmap by returning a pointer into the backing store. But |
| 2267 | // while such a mapping exists, the backing store cannot be resized. So write()s that extend |
| 2268 | // the file may fail. This does not work for SQLite's use case. |
| 2269 | // |
| 2270 | // So, alas, we must act like we don't support this. Luckily, SQLite has fallbacks for this. |
| 2271 | *pp = nullptr; |
| 2272 | return SQLITE_OK; |
| 2273 | }, |
| 2274 | .xUnfetch = [](sqlite3_file* file, sqlite3_int64 iOfst, void* p) noexcept -> int { |
| 2275 | // Shouldn't ever be called since xFetch() always produces null? But the native implementation |
| 2276 | // return SQLITE_OK even when mmap is disabled so we will too. |
| 2277 | return SQLITE_OK; |
| 2278 | }, |
| 2279 | #undef WRAP_METHOD |
| 2280 | }; |
| 2281 | |
| 2282 | // SQLite VFS implementation based on abstract `kj::Directory`. This is used only when the |
| 2283 | // directory is NOT a true disk directory. |
| 2284 | // |
| 2285 | // This is a namespace-struct defining static methods to fill in the function pointers of |
| 2286 | // sqlite3_vfs. |
| 2287 | sqlite3_vfs SqliteDatabase::Vfs::makeKjVfs() { |
| 2288 | return { |
| 2289 | .iVersion = kj::min(3, native.iVersion), .szOsFile = sizeof(FileImpl), |
| 2290 | |
| 2291 | // We have no real limit on paths but SQLite likes to allocate buffers of this size whenever |
| 2292 | // doing path stuff so making it huge would be bad. The default unix implementation uses |
| 2293 | // 512 as a limit so that "should be enough for anyone". |
| 2294 | .mxPathname = 512, |
| 2295 | |
| 2296 | .pNext = nullptr, |
| 2297 | .zName = name.cStr(), |
| 2298 | .pAppData = this, |
| 2299 | |
| 2300 | #define WRAP_METHOD(errorCode, block) \ |
| 2301 | auto& self KJ_UNUSED = *static_cast<const SqliteDatabase::Vfs*>(vfs->pAppData); \ |
| 2302 | try block catch (kj::Exception& e) { \ |
| 2303 | KJ_LOG(ERROR, "SQLite VFS I/O error", e); \ |
| 2304 | return errorCode; \ |
| 2305 | } |
| 2306 | |
| 2307 | .xOpen = [](sqlite3_vfs* vfs, sqlite3_filename zName, sqlite3_file* file, int flags, |
| 2308 | int* pOutFlags) -> int { |
| 2309 | WRAP_METHOD(SQLITE_CANTOPEN, { |
| 2310 | auto& target = *static_cast<FileImpl*>(file); |
| 2311 | |
| 2312 | if (flags & SQLITE_OPEN_READONLY) { |
| 2313 | KJ_REQUIRE(zName != nullptr, "readonly unnamed temporary file? what?"); |
| 2314 | KJ_REQUIRE(!(flags & SQLITE_OPEN_CREATE), "create readonly file? what?"); |
| 2315 | |
| 2316 | auto path = kj::Path::parse(zName); |
| 2317 | auto kjFile = KJ_UNWRAP_OR(self.directory.tryOpenFile(path), { return SQLITE_CANTOPEN; }); |
| 2318 | kj::Maybe<kj::Own<Lock>> lock; |
| 2319 | if (flags & SQLITE_OPEN_MAIN_DB) { |
| 2320 | lock = self.lockManager.lock(path, *kjFile); |
| 2321 | } |
| 2322 | |
| 2323 | kj::ctor(target, self, kj::mv(kjFile), kj::mv(lock)); |
| 2324 | } else { |
| 2325 | kj::Own<const kj::File> kjFile; |
| 2326 | kj::Maybe<kj::Own<Lock>> lock; |
| 2327 | |
| 2328 | if (zName == nullptr) { |
| 2329 | // Open a temp file. |
| 2330 | KJ_ASSERT(flags & SQLITE_OPEN_DELETEONCLOSE); |
| 2331 | KJ_ASSERT(!(flags & SQLITE_OPEN_MAIN_DB), "main DB can't be a temporary file"); |
| 2332 | kjFile = self.directory.createTemporary(); |
| 2333 | } else { |
| 2334 | kj::WriteMode mode; |
| 2335 | if (flags & SQLITE_OPEN_CREATE) { |
| 2336 | if (flags & SQLITE_OPEN_EXCLUSIVE) { |
| 2337 | mode = kj::WriteMode::CREATE; |
| 2338 | } else { |
| 2339 | mode = kj::WriteMode::CREATE | kj::WriteMode::MODIFY; |
| 2340 | } |
| 2341 | } else { |
| 2342 | mode = kj::WriteMode::MODIFY; |
| 2343 | } |
| 2344 | |
| 2345 | auto path = kj::Path::parse(zName); |
| 2346 | kjFile = |
| 2347 | KJ_UNWRAP_OR(self.directory.tryOpenFile(path, mode), { return SQLITE_CANTOPEN; }); |
| 2348 | if (flags & SQLITE_OPEN_MAIN_DB) { |
| 2349 | lock = self.lockManager.lock(path, *kjFile); |
| 2350 | } |
| 2351 | |
| 2352 | if (flags & SQLITE_OPEN_DELETEONCLOSE) { |
| 2353 | self.directory.remove(path); |
| 2354 | } |
| 2355 | } |
| 2356 | |
| 2357 | kj::ctor(target, self, kj::mv(kjFile), kj::mv(lock)); |
| 2358 | } |
| 2359 | |
| 2360 | // In theory if read-write was requested, but failed, we should retry read-only, and then |
| 2361 | // alter the pOutFlags to reflect this... I'm not going to bother. |
| 2362 | if (pOutFlags != nullptr) { |
| 2363 | *pOutFlags = flags; |
| 2364 | } |
| 2365 | |
| 2366 | return SQLITE_OK; |
| 2367 | }); |
| 2368 | }, |
| 2369 | .xDelete = [](sqlite3_vfs* vfs, const char* zName, int syncDir) -> int { |
| 2370 | WRAP_METHOD(SQLITE_IOERR_DELETE, { |
| 2371 | if (self.directory.tryRemove(kj::Path::parse(zName))) { |
| 2372 | return SQLITE_OK; |
| 2373 | } else { |
| 2374 | return SQLITE_IOERR_DELETE_NOENT; |
| 2375 | } |
| 2376 | }); |
| 2377 | }, |
| 2378 | .xAccess = [](sqlite3_vfs* vfs, const char* zName, int flags, int* pResOut) -> int { |
| 2379 | WRAP_METHOD(SQLITE_IOERR_ACCESS, { |
| 2380 | // Technically, depending on the flags, this may be checking whether the file is readable |
| 2381 | // or writable, rather than just whether it exists. However, the KJ filesystem API |
| 2382 | // assumes that all descendents of a writable directory are readable and writable, hence |
| 2383 | // this is equivalent to checking for existence. |
| 2384 | // |
| 2385 | // If we were to extend the VFS so it can wrap `kj::ReadableDirectory` then in that case |
| 2386 | // we would want to return false when querying writability. |
| 2387 | *pResOut = self.directory.exists(kj::Path::parse(zName)); |
| 2388 | return SQLITE_OK; |
| 2389 | }); |
| 2390 | }, |
| 2391 | .xFullPathname = [](sqlite3_vfs*, const char* zName, int nOut, char* zOut) -> int { |
| 2392 | // Don't rewrite the path at all. All paths are canonical. Our path parsing will reject |
| 2393 | // the existence of `.` or `..` as path components as well as leading `/`. |
| 2394 | size_t len = kj::min(strlen(zName), nOut - 1); |
| 2395 | memcpy(zOut, zName, len); |
| 2396 | zOut[len] = 0; |
| 2397 | return SQLITE_OK; |
| 2398 | }, |
| 2399 | |
| 2400 | // We don't support loading shared libraries from virtual files. |
| 2401 | .xDlOpen = nullptr, |
| 2402 | .xDlError = nullptr, |
| 2403 | .xDlSym = nullptr, |
| 2404 | .xDlClose = nullptr, |
| 2405 | |
| 2406 | // Use native implementations of these OS functions. I'm not sure why these are even part |
| 2407 | // of the VFS. (Exception: xGetLastError is actually sensibly a VFS thing, but we are allowed |
| 2408 | // to just not implement it.) |
| 2409 | .xRandomness = native.xRandomness, |
| 2410 | .xSleep = native.xSleep, |
| 2411 | .xCurrentTime = native.xCurrentTime, |
| 2412 | .xGetLastError = nullptr, |
| 2413 | .xCurrentTimeInt64 = native.xCurrentTimeInt64, |
| 2414 | |
| 2415 | // We don't support overriding any syscalls. |
| 2416 | .xSetSystemCall = nullptr, |
| 2417 | .xGetSystemCall = nullptr, |
| 2418 | .xNextSystemCall = nullptr, |
| 2419 | |
| 2420 | #undef WRAP_METHOD |
| 2421 | }; |
| 2422 | }; |
| 2423 | |
| 2424 | // clang-format on |
| 2425 | |
| 2426 | // ----------------------------------------------------------------------------- |
| 2427 | |
| 2428 | class SqliteDatabase::Vfs::DefaultLockManager final: public SqliteDatabase::LockManager { |
| 2429 | public: |
| 2430 | kj::Own<Lock> lock(kj::PathPtr path, const kj::ReadableFile& mainDatabaseFile) const override { |
| 2431 | return kj::heap<LockImpl>(*this, path); |
| 2432 | } |
| 2433 | |
| 2434 | private: |
| 2435 | class LockImpl; |
| 2436 | struct LockState; |
| 2437 | using LockMap = kj::HashMap<kj::PathPtr, LockState*>; |
| 2438 | kj::MutexGuarded<LockMap> lockMap; |
| 2439 | |
| 2440 | struct LockState: public kj::Refcounted { |
| 2441 | // Note: The refcount of this object is protected by `lockMap`'s mutex. |
| 2442 | |
| 2443 | struct Guarded { |
| 2444 | kj::Vector<kj::Array<byte>> regions; |
| 2445 | |
| 2446 | uint sharedLockCount = 0; |
| 2447 | bool hasReserved = false; |
| 2448 | bool hasPendingOrExclusive = false; |
| 2449 | |
| 2450 | uint walLocks[Lock::WAL_LOCK_COUNT] = {0, 0, 0, 0, 0, 0, 0, 0}; |
| 2451 | // Each slot contains the count of shared locks, or kj::maxValue if an exclusive lock is |
| 2452 | // held. |
| 2453 | }; |
| 2454 | |
| 2455 | const kj::Path path; |
| 2456 | kj::MutexGuarded<Guarded> guarded; |
| 2457 | |
| 2458 | LockState(kj::Path path): path(kj::mv(path)) {} |
| 2459 | }; |
| 2460 | |
| 2461 | class LockImpl final: public Lock { |
| 2462 | public: |
| 2463 | LockImpl(const DefaultLockManager& lockManager, kj::PathPtr path): lockManager(lockManager) { |
| 2464 | auto mlock = lockManager.lockMap.lockExclusive(); |
| 2465 | auto& slot = mlock->findOrCreate(path, [&]() { |
| 2466 | state = kj::refcounted<LockState>(path.clone()); |
| 2467 | return LockMap::Entry{.key = state->path, .value = state.get()}; |
| 2468 | }); |
| 2469 | if (state.get() == nullptr) { |
| 2470 | state = kj::addRef(*slot); |
| 2471 | } |
| 2472 | } |
| 2473 | |
| 2474 | ~LockImpl() noexcept(false) { |
| 2475 | // It's important that we drop the state object under lock to ensure no other thread is |
| 2476 | // in the process of grabbing it out of the map at the same time. Since we have to take a |
| 2477 | // lock here anyway, `LockState` uses regular non-atomic refcounts rather than atomic. |
| 2478 | auto mlock = lockManager.lockMap.lockExclusive(); |
| 2479 | auto stateToDrop = kj::mv(state); |
| 2480 | if (!stateToDrop->isShared()) { |
| 2481 | mlock->erase(stateToDrop->path); |
| 2482 | } |
| 2483 | } |
| 2484 | |
| 2485 | bool tryIncreaseLevel(Level newLevel) override { |
| 2486 | if (newLevel <= currentLevel) return true; |
| 2487 | |
| 2488 | auto slock = state->guarded.lockExclusive(); |
| 2489 | |
| 2490 | if (currentLevel < SHARED) { |
| 2491 | if (slock->hasPendingOrExclusive) { |
| 2492 | return false; |
| 2493 | } |
| 2494 | ++slock->sharedLockCount; |
| 2495 | currentLevel = SHARED; |
| 2496 | } |
| 2497 | |
| 2498 | if (newLevel == SHARED) { |
| 2499 | return true; |
| 2500 | } |
| 2501 | |
| 2502 | if (newLevel == RESERVED) { |
| 2503 | if (slock->hasReserved || slock->hasPendingOrExclusive) { |
| 2504 | return false; |
| 2505 | } |
| 2506 | if (currentLevel == SHARED) { |
| 2507 | KJ_ASSERT(slock->sharedLockCount > 0); |
| 2508 | --slock->sharedLockCount; |
| 2509 | } |
| 2510 | slock->hasReserved = true; |
| 2511 | currentLevel = RESERVED; |
| 2512 | return true; |
| 2513 | } |
| 2514 | |
| 2515 | // Requesting PENDING or EXCLUSIVE. If EXCLUSIVE, we still have to transition through |
| 2516 | // PENDING first, if we're not there already. |
| 2517 | if (currentLevel < PENDING) { |
| 2518 | if (currentLevel != RESERVED && slock->hasReserved) { |
| 2519 | return false; |
| 2520 | } |
| 2521 | if (slock->hasPendingOrExclusive) { |
| 2522 | return false; |
| 2523 | } |
| 2524 | if (currentLevel == SHARED) { |
| 2525 | KJ_ASSERT(slock->sharedLockCount > 0); |
| 2526 | --slock->sharedLockCount; |
| 2527 | } |
| 2528 | slock->hasReserved = false; |
| 2529 | slock->hasPendingOrExclusive = true; |
| 2530 | currentLevel = PENDING; |
| 2531 | } |
| 2532 | |
| 2533 | if (newLevel == EXCLUSIVE) { |
| 2534 | if (slock->sharedLockCount > 0) { |
| 2535 | return false; |
| 2536 | } |
| 2537 | currentLevel = EXCLUSIVE; |
| 2538 | } |
| 2539 | |
| 2540 | return true; |
| 2541 | } |
| 2542 | |
| 2543 | void decreaseLevel(Level newLevel) override { |
| 2544 | if (newLevel >= currentLevel) return; |
| 2545 | KJ_REQUIRE(newLevel <= SHARED); |
| 2546 | |
| 2547 | auto slock = state->guarded.lockExclusive(); |
| 2548 | if (currentLevel >= PENDING) { |
| 2549 | slock->hasPendingOrExclusive = false; |
| 2550 | } |
| 2551 | if (currentLevel == RESERVED) { |
| 2552 | slock->hasReserved = false; |
| 2553 | } |
| 2554 | if (currentLevel == SHARED && newLevel == UNLOCKED) { |
| 2555 | KJ_ASSERT(slock->sharedLockCount > 0); |
| 2556 | --slock->sharedLockCount; |
| 2557 | } |
| 2558 | if (newLevel == SHARED) { |
| 2559 | ++slock->sharedLockCount; |
| 2560 | } |
| 2561 | currentLevel = newLevel; |
| 2562 | } |
| 2563 | |
| 2564 | bool checkReservedLock() override { |
| 2565 | return state->guarded.lockShared()->hasReserved; |
| 2566 | } |
| 2567 | |
| 2568 | kj::ArrayPtr<byte> getSharedMemoryRegion(uint index, uint size, bool extend) override { |
| 2569 | if (extend) { |
| 2570 | auto slock = state->guarded.lockExclusive(); |
| 2571 | |
| 2572 | while (index >= slock->regions.size()) { |
| 2573 | auto newRegion = kj::heapArray<byte>(size); |
| 2574 | newRegion.asPtr().fill(0); |
| 2575 | slock->regions.add(kj::mv(newRegion)); |
| 2576 | } |
| 2577 | |
| 2578 | return slock->regions[index]; |
| 2579 | } else { |
| 2580 | auto slock = state->guarded.lockShared(); |
| 2581 | |
| 2582 | if (index >= slock->regions.size()) { |
| 2583 | return nullptr; |
| 2584 | } else { |
| 2585 | kj::ArrayPtr<const byte> region = slock->regions[index]; |
| 2586 | // const_cast OK because the caller will carefully control access to shared memory. |
| 2587 | return kj::arrayPtr(const_cast<byte*>(region.begin()), region.size()); |
| 2588 | } |
| 2589 | } |
| 2590 | } |
| 2591 | |
| 2592 | void clearSharedMemory() override { |
| 2593 | auto slock = state->guarded.lockExclusive(); |
| 2594 | slock->regions.clear(); |
| 2595 | } |
| 2596 | |
| 2597 | bool tryLockWalShared(uint start, uint count) override { |
| 2598 | auto slock = state->guarded.lockExclusive(); |
| 2599 | |
| 2600 | for (uint i = start; i < start + count; i++) { |
| 2601 | if (slock->walLocks[i] == static_cast<uint>(kj::maxValue)) { |
| 2602 | // blocked by exclusive lock |
| 2603 | return false; |
| 2604 | } |
| 2605 | } |
| 2606 | for (uint i = start; i < start + count; i++) { |
| 2607 | ++slock->walLocks[i]; |
| 2608 | } |
| 2609 | return true; |
| 2610 | } |
| 2611 | bool tryLockWalExclusive(uint start, uint count) override { |
| 2612 | auto slock = state->guarded.lockExclusive(); |
| 2613 | |
| 2614 | for (uint i = start; i < start + count; i++) { |
| 2615 | if (slock->walLocks[i] != 0) { |
| 2616 | // blocked by another lock |
| 2617 | return false; |
| 2618 | } |
| 2619 | } |
| 2620 | for (uint i = start; i < start + count; i++) { |
| 2621 | slock->walLocks[i] = kj::maxValue; |
| 2622 | } |
| 2623 | return true; |
| 2624 | } |
| 2625 | |
| 2626 | void unlockWalShared(uint start, uint count) override { |
| 2627 | auto slock = state->guarded.lockExclusive(); |
| 2628 | for (uint i = start; i < start + count; i++) { |
| 2629 | KJ_ASSERT(slock->walLocks[i] != 0); |
| 2630 | --slock->walLocks[i]; |
| 2631 | } |
| 2632 | } |
| 2633 | void unlockWalExclusive(uint start, uint count) override { |
| 2634 | auto slock = state->guarded.lockExclusive(); |
| 2635 | for (uint i = start; i < start + count; i++) { |
| 2636 | KJ_REQUIRE(slock->walLocks[i] == (uint)kj::maxValue); |
| 2637 | slock->walLocks[i] = 0; |
| 2638 | } |
| 2639 | } |
| 2640 | |
| 2641 | private: |
| 2642 | const DefaultLockManager& lockManager; |
| 2643 | kj::Own<LockState> state; |
| 2644 | Level currentLevel = UNLOCKED; |
| 2645 | }; |
| 2646 | }; |
| 2647 | |
| 2648 | SqliteDatabase::Vfs::Vfs(const kj::Directory& directory, Options options) |
| 2649 | : directory(directory), |
| 2650 | ownLockManager(kj::heap<DefaultLockManager>()), |
| 2651 | lockManager(*ownLockManager), |
| 2652 | options(kj::mv(options)), |
| 2653 | native(*sqlite3_vfs_find(nullptr)) { |
| 2654 | #if _WIN32 |
| 2655 | vfs = kj::heap(makeKjVfs()); |
| 2656 | #else |
| 2657 | KJ_IF_SOME(fd, directory.getFd()) { |
| 2658 | rootFd = fd; |
| 2659 | vfs = kj::heap(makeWrappedNativeVfs()); |
| 2660 | } else { |
| 2661 | vfs = kj::heap(makeKjVfs()); |
| 2662 | } |
| 2663 | #endif |
| 2664 | sqlite3_vfs_register(vfs, false); |
| 2665 | } |
| 2666 | |
| 2667 | SqliteDatabase::Vfs::Vfs( |
| 2668 | const kj::Directory& directory, const LockManager& lockManager, Options options) |
| 2669 | : directory(directory), |
| 2670 | lockManager(lockManager), |
| 2671 | options(kj::mv(options)), |
| 2672 | native(*sqlite3_vfs_find(nullptr)), |
| 2673 | // Always use KJ VFS when using a custom LockManager. |
| 2674 | vfs(kj::heap(makeKjVfs())) { |
| 2675 | sqlite3_vfs_register(vfs, false); |
| 2676 | } |
| 2677 | |
| 2678 | SqliteDatabase::Vfs::~Vfs() noexcept(false) { |
| 2679 | sqlite3_vfs_unregister(vfs); |
| 2680 | } |
| 2681 | |
| 2682 | kj::String SqliteDatabase::Vfs::makeName() { |
| 2683 | // A pointer to this object should be suitably unique. (Ugghhhh.) |
| 2684 | return kj::str("kj-", this); |
| 2685 | } |
| 2686 | |
| 2687 | #if _WIN32 |
| 2688 | kj::Maybe<kj::Path> SqliteDatabase::Vfs::tryAppend(kj::PathPtr suffix) const { |
| 2689 | auto handle = KJ_UNWRAP_OR_RETURN(directory.getWin32Handle(), nullptr); |
| 2690 | auto root = getPathFromWin32Handle(handle); |
| 2691 | return root.append(suffix); |
| 2692 | } |
| 2693 | #else |
| 2694 | kj::Maybe<kj::Path> SqliteDatabase::Vfs::tryAppend(kj::PathPtr suffix) const { |
| 2695 | // TODO(someday): consider implementing this on other platforms |
| 2696 | return kj::none; |
| 2697 | } |
| 2698 | #endif |
| 2699 | |
| 2700 | // ======================================================================================= |
| 2701 | |
| 2702 | } // namespace workerd |