| 1 | //===----------------------------------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | /// \file |
| 10 | /// Encapsulates \p OnDiskGraphDB and \p OnDiskKeyValueDB instances within one |
| 11 | /// directory while also restricting storage growth with a scheme of chaining |
| 12 | /// the two most recent directories (primary & upstream), where the primary |
| 13 | /// "faults-in" data from the upstream one. When the primary (most recent) |
| 14 | /// directory exceeds its intended limit a new empty directory becomes the |
| 15 | /// primary one. |
| 16 | /// |
| 17 | /// Within the top-level directory (the path that \p UnifiedOnDiskCache::open |
| 18 | /// receives) there are directories named like this: |
| 19 | /// |
| 20 | /// 'v<version>.<x>' |
| 21 | /// 'v<version>.<x+1>' |
| 22 | /// 'v<version>.<x+2>' |
| 23 | /// ... |
| 24 | /// |
| 25 | /// 'version' is the version integer for this \p UnifiedOnDiskCache's scheme and |
| 26 | /// the part after the dot is an increasing integer. The primary directory is |
| 27 | /// the one with the highest integer and the upstream one is the directory |
| 28 | /// before it. For example, if the sub-directories contained are: |
| 29 | /// |
| 30 | /// 'v1.5', 'v1.6', 'v1.7', 'v1.8' |
| 31 | /// |
| 32 | /// Then the primary one is 'v1.8', the upstream one is 'v1.7', and the rest are |
| 33 | /// unused directories that can be safely deleted at any time and by any |
| 34 | /// process. |
| 35 | /// |
| 36 | /// Contained within the top-level directory is a file named "lock" which is |
| 37 | /// used for processes to take shared or exclusive locks for the contents of the |
| 38 | /// top directory. While a \p UnifiedOnDiskCache is open it keeps a shared lock |
| 39 | /// for the top-level directory; when it closes, if the primary sub-directory |
| 40 | /// exceeded its limit, it attempts to get an exclusive lock in order to create |
| 41 | /// a new empty primary directory; if it can't get the exclusive lock it gives |
| 42 | /// up and lets the next \p UnifiedOnDiskCache instance that closes to attempt |
| 43 | /// again. |
| 44 | /// |
| 45 | /// The downside of this scheme is that while \p UnifiedOnDiskCache is open on a |
| 46 | /// directory, by any process, the storage size in that directory will keep |
| 47 | /// growing unrestricted. But the major benefit is that garbage-collection can |
| 48 | /// be triggered on a directory concurrently, at any time and by any process, |
| 49 | /// without affecting any active readers/writers in the same process or other |
| 50 | /// processes. |
| 51 | /// |
| 52 | /// The \c UnifiedOnDiskCache also provides validation and recovery on top of |
| 53 | /// the underlying on-disk storage. The low-level storage is designed to remain |
| 54 | /// coherent across regular process crashes, but may be invalid after power loss |
| 55 | /// or similar system failures. \c UnifiedOnDiskCache::validateIfNeeded allows |
| 56 | /// validating the contents once per boot and can recover by marking invalid |
| 57 | /// data for garbage collection. |
| 58 | /// |
| 59 | /// The data recovery described above requires exclusive access to the CAS, and |
| 60 | /// it is an error to attempt recovery if the CAS is open in any process/thread. |
| 61 | /// In order to maximize backwards compatibility with tools that do not perform |
| 62 | /// validation before opening the CAS, we do not attempt to get exclusive access |
| 63 | /// until recovery is actually performed, meaning as long as the data is valid |
| 64 | /// it will not conflict with concurrent use. |
| 65 | // |
| 66 | //===----------------------------------------------------------------------===// |
| 67 | |
| 68 | #include "llvm/CAS/UnifiedOnDiskCache.h" |
| 69 | #include "OnDiskCommon.h" |
| 70 | #include "llvm/ADT/STLExtras.h" |
| 71 | #include "llvm/ADT/ScopeExit.h" |
| 72 | #include "llvm/ADT/SmallString.h" |
| 73 | #include "llvm/ADT/SmallVector.h" |
| 74 | #include "llvm/ADT/StringExtras.h" |
| 75 | #include "llvm/ADT/StringRef.h" |
| 76 | #include "llvm/CAS/OnDiskCASLogger.h" |
| 77 | #include "llvm/CAS/OnDiskGraphDB.h" |
| 78 | #include "llvm/CAS/OnDiskKeyValueDB.h" |
| 79 | #include "llvm/Support/Compiler.h" |
| 80 | #include "llvm/Support/Errc.h" |
| 81 | #include "llvm/Support/Error.h" |
| 82 | #include "llvm/Support/FileSystem.h" |
| 83 | #include "llvm/Support/FileUtilities.h" |
| 84 | #include "llvm/Support/IOSandbox.h" |
| 85 | #include "llvm/Support/MemoryBuffer.h" |
| 86 | #include "llvm/Support/Path.h" |
| 87 | #include "llvm/Support/Program.h" |
| 88 | #include "llvm/Support/raw_ostream.h" |
| 89 | #include <optional> |
| 90 | |
| 91 | using namespace llvm; |
| 92 | using namespace llvm::cas; |
| 93 | using namespace llvm::cas::ondisk; |
| 94 | |
| 95 | /// FIXME: When the version of \p DBDirPrefix is bumped up we need to figure out |
| 96 | /// how to handle the leftover sub-directories of the previous version, within |
| 97 | /// the \p UnifiedOnDiskCache::collectGarbage function. |
| 98 | static constexpr StringLiteral DBDirPrefix = "v1." ; |
| 99 | |
| 100 | static constexpr StringLiteral ValidationFilename = "v1.validation" ; |
| 101 | static constexpr StringLiteral CorruptPrefix = "corrupt." ; |
| 102 | |
| 103 | ObjectID UnifiedOnDiskCache::getObjectIDFromValue(ArrayRef<char> Value) { |
| 104 | // little endian encoded. |
| 105 | assert(Value.size() == sizeof(uint64_t)); |
| 106 | return ObjectID::fromOpaqueData(Opaque: support::endian::read64le(P: Value.data())); |
| 107 | } |
| 108 | |
| 109 | UnifiedOnDiskCache::ValueBytes |
| 110 | UnifiedOnDiskCache::getValueFromObjectID(ObjectID ID) { |
| 111 | // little endian encoded. |
| 112 | UnifiedOnDiskCache::ValueBytes ValBytes; |
| 113 | static_assert(ValBytes.size() == sizeof(ID.getOpaqueData())); |
| 114 | support::endian::write64le(P: ValBytes.data(), V: ID.getOpaqueData()); |
| 115 | return ValBytes; |
| 116 | } |
| 117 | |
| 118 | Expected<std::optional<ArrayRef<char>>> |
| 119 | UnifiedOnDiskCache::faultInFromUpstreamKV(ArrayRef<uint8_t> Key) { |
| 120 | assert(UpstreamGraphDB); |
| 121 | assert(UpstreamKVDB); |
| 122 | |
| 123 | std::optional<ArrayRef<char>> UpstreamValue; |
| 124 | if (Error E = UpstreamKVDB->get(Key).moveInto(Value&: UpstreamValue)) |
| 125 | return std::move(E); |
| 126 | if (!UpstreamValue) |
| 127 | return std::nullopt; |
| 128 | |
| 129 | // The value is the \p ObjectID in the context of the upstream |
| 130 | // \p OnDiskGraphDB instance. Translate it to the context of the primary |
| 131 | // \p OnDiskGraphDB instance. |
| 132 | ObjectID UpstreamID = getObjectIDFromValue(Value: *UpstreamValue); |
| 133 | auto PrimaryID = |
| 134 | PrimaryGraphDB->getReference(Hash: UpstreamGraphDB->getDigest(Ref: UpstreamID)); |
| 135 | if (LLVM_UNLIKELY(!PrimaryID)) |
| 136 | return PrimaryID.takeError(); |
| 137 | return PrimaryKVDB->put(Key, Value: getValueFromObjectID(ID: *PrimaryID)); |
| 138 | } |
| 139 | |
| 140 | /// \returns all the 'v<version>.<x>' names of sub-directories, sorted with |
| 141 | /// ascending order of the integer after the dot. Corrupt directories, if |
| 142 | /// included, will come first. |
| 143 | static Expected<SmallVector<std::string, 4>> |
| 144 | getAllDBDirs(StringRef Path, bool IncludeCorrupt = false) { |
| 145 | struct DBDir { |
| 146 | uint64_t Order; |
| 147 | std::string Name; |
| 148 | }; |
| 149 | SmallVector<DBDir> FoundDBDirs; |
| 150 | |
| 151 | std::error_code EC; |
| 152 | for (sys::fs::directory_iterator DirI(Path, EC), DirE; !EC && DirI != DirE; |
| 153 | DirI.increment(ec&: EC)) { |
| 154 | if (DirI->type() != sys::fs::file_type::directory_file) |
| 155 | continue; |
| 156 | StringRef SubDir = sys::path::filename(path: DirI->path()); |
| 157 | if (IncludeCorrupt && SubDir.starts_with(Prefix: CorruptPrefix)) { |
| 158 | FoundDBDirs.push_back(Elt: {.Order: 0, .Name: std::string(SubDir)}); |
| 159 | continue; |
| 160 | } |
| 161 | if (!SubDir.starts_with(Prefix: DBDirPrefix)) |
| 162 | continue; |
| 163 | uint64_t Order; |
| 164 | if (SubDir.substr(Start: DBDirPrefix.size()).getAsInteger(Radix: 10, Result&: Order)) |
| 165 | return createStringError(EC: inconvertibleErrorCode(), |
| 166 | S: "unexpected directory " + DirI->path()); |
| 167 | FoundDBDirs.push_back(Elt: {.Order: Order, .Name: std::string(SubDir)}); |
| 168 | } |
| 169 | if (EC) |
| 170 | return createFileError(F: Path, EC); |
| 171 | |
| 172 | llvm::sort(C&: FoundDBDirs, Comp: [](const DBDir &LHS, const DBDir &RHS) -> bool { |
| 173 | return LHS.Order < RHS.Order; |
| 174 | }); |
| 175 | |
| 176 | SmallVector<std::string, 4> DBDirs; |
| 177 | for (DBDir &Dir : FoundDBDirs) |
| 178 | DBDirs.push_back(Elt: std::move(Dir.Name)); |
| 179 | return DBDirs; |
| 180 | } |
| 181 | |
| 182 | static Expected<SmallVector<std::string, 4>> getAllGarbageDirs(StringRef Path) { |
| 183 | auto DBDirs = getAllDBDirs(Path, /*IncludeCorrupt=*/true); |
| 184 | if (!DBDirs) |
| 185 | return DBDirs.takeError(); |
| 186 | |
| 187 | // FIXME: When the version of \p DBDirPrefix is bumped up we need to figure |
| 188 | // out how to handle the leftover sub-directories of the previous version. |
| 189 | |
| 190 | for (unsigned Keep = 2; Keep > 0 && !DBDirs->empty(); --Keep) { |
| 191 | StringRef Back(DBDirs->back()); |
| 192 | if (Back.starts_with(Prefix: CorruptPrefix)) |
| 193 | break; |
| 194 | DBDirs->pop_back(); |
| 195 | } |
| 196 | return *DBDirs; |
| 197 | } |
| 198 | |
| 199 | /// \returns Given a sub-directory named 'v<version>.<x>', it outputs the |
| 200 | /// 'v<version>.<x+1>' name. |
| 201 | static void getNextDBDirName(StringRef DBDir, llvm::raw_ostream &OS) { |
| 202 | assert(DBDir.starts_with(DBDirPrefix)); |
| 203 | uint64_t Count; |
| 204 | bool Failed = DBDir.substr(Start: DBDirPrefix.size()).getAsInteger(Radix: 10, Result&: Count); |
| 205 | assert(!Failed); |
| 206 | (void)Failed; |
| 207 | OS << DBDirPrefix << Count + 1; |
| 208 | } |
| 209 | |
| 210 | static Error validateOutOfProcess(StringRef LLVMCasBinary, StringRef RootPath, |
| 211 | bool CheckHash) { |
| 212 | SmallVector<StringRef> Args{LLVMCasBinary, "-cas" , RootPath, "-validate" }; |
| 213 | if (CheckHash) |
| 214 | Args.push_back(Elt: "-check-hash" ); |
| 215 | |
| 216 | llvm::SmallString<128> StdErrPath; |
| 217 | int StdErrFD = -1; |
| 218 | if (std::error_code EC = sys::fs::createTemporaryFile( |
| 219 | Prefix: "llvm-cas-validate-stderr" , Suffix: "txt" , ResultFD&: StdErrFD, ResultPath&: StdErrPath, |
| 220 | Flags: llvm::sys::fs::OF_Text)) |
| 221 | return createStringError(EC, S: "failed to create temporary file" ); |
| 222 | FileRemover OutputRemover(StdErrPath.c_str()); |
| 223 | |
| 224 | std::optional<llvm::StringRef> Redirects[] = { |
| 225 | {"" }, // stdin = /dev/null |
| 226 | {"" }, // stdout = /dev/null |
| 227 | StdErrPath.str(), |
| 228 | }; |
| 229 | |
| 230 | std::string ErrMsg; |
| 231 | int Result = |
| 232 | sys::ExecuteAndWait(Program: LLVMCasBinary, Args, /*Env=*/std::nullopt, Redirects, |
| 233 | /*SecondsToWait=*/120, /*MemoryLimit=*/0, ErrMsg: &ErrMsg); |
| 234 | |
| 235 | if (Result == -1) |
| 236 | return createStringError(S: "failed to exec " + join(R&: Args, Separator: " " ) + ": " + |
| 237 | ErrMsg); |
| 238 | if (Result != 0) { |
| 239 | llvm::SmallString<64> Err("cas contents invalid" ); |
| 240 | if (!ErrMsg.empty()) { |
| 241 | Err += ": " ; |
| 242 | Err += ErrMsg; |
| 243 | } |
| 244 | auto StdErrBuf = MemoryBuffer::getFile(Filename: StdErrPath.c_str()); |
| 245 | if (StdErrBuf && !(*StdErrBuf)->getBuffer().empty()) { |
| 246 | Err += ": " ; |
| 247 | Err += (*StdErrBuf)->getBuffer(); |
| 248 | } |
| 249 | return createStringError(S: Err); |
| 250 | } |
| 251 | return Error::success(); |
| 252 | } |
| 253 | |
| 254 | Error UnifiedOnDiskCache::validateActionCache() const { |
| 255 | auto ValidateRef = [this](FileOffset Offset, ArrayRef<char> Value) -> Error { |
| 256 | auto ID = ondisk::UnifiedOnDiskCache::getObjectIDFromValue(Value); |
| 257 | auto formatError = [&](Twine Msg) { |
| 258 | return createStringError( |
| 259 | EC: llvm::errc::illegal_byte_sequence, |
| 260 | S: "bad record at 0x" + |
| 261 | utohexstr(X: (unsigned)Offset.get(), /*LowerCase=*/true) + ": " + |
| 262 | Msg.str()); |
| 263 | }; |
| 264 | if (Error E = this->getGraphDB().validateObjectID(ID)) |
| 265 | return formatError(llvm::toString(E: std::move(E))); |
| 266 | return Error::success(); |
| 267 | }; |
| 268 | return getKeyValueDB().validate(CheckValue: ValidateRef); |
| 269 | } |
| 270 | |
| 271 | static Error validateInProcess(StringRef RootPath, StringRef HashName, |
| 272 | unsigned HashByteSize, bool CheckHash, |
| 273 | OnDiskGraphDB::HashingFuncT HashFn) { |
| 274 | std::shared_ptr<UnifiedOnDiskCache> UniDB; |
| 275 | if (Error E = UnifiedOnDiskCache::open(Path: RootPath, SizeLimit: std::nullopt, HashName, |
| 276 | HashByteSize) |
| 277 | .moveInto(Value&: UniDB)) |
| 278 | return E; |
| 279 | if (Error E = UniDB->getGraphDB().validate(Deep: CheckHash, Hasher: HashFn)) |
| 280 | return E; |
| 281 | if (Error E = UniDB->validateActionCache()) |
| 282 | return E; |
| 283 | return Error::success(); |
| 284 | } |
| 285 | |
| 286 | Expected<ValidationResult> UnifiedOnDiskCache::validateIfNeeded( |
| 287 | StringRef RootPath, StringRef HashName, unsigned HashByteSize, |
| 288 | bool CheckHash, OnDiskGraphDB::HashingFuncT HashFn, bool AllowRecovery, |
| 289 | bool ForceValidation, std::optional<StringRef> LLVMCasBinaryPath) { |
| 290 | if (std::error_code EC = sys::fs::create_directories(path: RootPath)) |
| 291 | return createFileError(F: RootPath, EC); |
| 292 | |
| 293 | SmallString<256> PathBuf(RootPath); |
| 294 | sys::path::append(path&: PathBuf, a: ValidationFilename); |
| 295 | int FD = -1; |
| 296 | if (std::error_code EC = sys::fs::openFileForReadWrite( |
| 297 | Name: PathBuf, ResultFD&: FD, Disp: sys::fs::CD_OpenAlways, Flags: sys::fs::OF_None)) |
| 298 | return createFileError(F: PathBuf, EC); |
| 299 | assert(FD != -1); |
| 300 | |
| 301 | sys::fs::file_t File = sys::fs::convertFDToNativeFile(FD); |
| 302 | llvm::scope_exit CloseFile([&]() { sys::fs::closeFile(F&: File); }); |
| 303 | |
| 304 | if (std::error_code EC = lockFileThreadSafe(FD, Kind: sys::fs::LockKind::Exclusive)) |
| 305 | return createFileError(F: PathBuf, EC); |
| 306 | llvm::scope_exit UnlockFD([&]() { unlockFileThreadSafe(FD); }); |
| 307 | |
| 308 | std::shared_ptr<ondisk::OnDiskCASLogger> Logger; |
| 309 | #ifndef _WIN32 |
| 310 | if (Error E = |
| 311 | ondisk::OnDiskCASLogger::openIfEnabled(Path: RootPath).moveInto(Value&: Logger)) |
| 312 | return std::move(E); |
| 313 | #endif |
| 314 | |
| 315 | SmallString<8> Bytes; |
| 316 | if (Error E = sys::fs::readNativeFileToEOF(FileHandle: File, Buffer&: Bytes)) |
| 317 | return createFileError(F: PathBuf, E: std::move(E)); |
| 318 | |
| 319 | uint64_t ValidationBootTime = 0; |
| 320 | if (!Bytes.empty() && |
| 321 | StringRef(Bytes).trim().getAsInteger(Radix: 10, Result&: ValidationBootTime)) |
| 322 | return createFileError(F: PathBuf, EC: errc::illegal_byte_sequence, |
| 323 | Fmt: "expected integer" ); |
| 324 | |
| 325 | static uint64_t BootTime = 0; |
| 326 | if (BootTime == 0) |
| 327 | if (Error E = getBootTime().moveInto(Value&: BootTime)) |
| 328 | return std::move(E); |
| 329 | |
| 330 | bool Recovered = false; |
| 331 | bool Skipped = false; |
| 332 | std::string LogValidationError; |
| 333 | |
| 334 | llvm::scope_exit Log([&] { |
| 335 | if (!Logger) |
| 336 | return; |
| 337 | Logger->logUnifiedOnDiskCacheValidateIfNeeded( |
| 338 | Path: RootPath, BootTime, ValidationTime: ValidationBootTime, CheckHash, AllowRecovery, |
| 339 | Force: ForceValidation, LLVMCas: LLVMCasBinaryPath, ValidationError: LogValidationError, Skipped, |
| 340 | Recovered); |
| 341 | }); |
| 342 | |
| 343 | if (ValidationBootTime == BootTime && !ForceValidation) { |
| 344 | Skipped = true; |
| 345 | return ValidationResult::Skipped; |
| 346 | } |
| 347 | |
| 348 | // Validate! |
| 349 | bool NeedsRecovery = false; |
| 350 | Error E = LLVMCasBinaryPath |
| 351 | ? validateOutOfProcess(LLVMCasBinary: *LLVMCasBinaryPath, RootPath, CheckHash) |
| 352 | : validateInProcess(RootPath, HashName, HashByteSize, CheckHash, |
| 353 | HashFn); |
| 354 | if (E) { |
| 355 | if (Logger) |
| 356 | LogValidationError = toStringWithoutConsuming(E); |
| 357 | if (AllowRecovery) { |
| 358 | consumeError(Err: std::move(E)); |
| 359 | NeedsRecovery = true; |
| 360 | } else { |
| 361 | return std::move(E); |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | if (NeedsRecovery) { |
| 366 | sys::path::remove_filename(path&: PathBuf); |
| 367 | sys::path::append(path&: PathBuf, a: "lock" ); |
| 368 | |
| 369 | int LockFD = -1; |
| 370 | if (std::error_code EC = sys::fs::openFileForReadWrite( |
| 371 | Name: PathBuf, ResultFD&: LockFD, Disp: sys::fs::CD_OpenAlways, Flags: sys::fs::OF_None)) |
| 372 | return createFileError(F: PathBuf, EC); |
| 373 | sys::fs::file_t LockFile = sys::fs::convertFDToNativeFile(FD: LockFD); |
| 374 | llvm::scope_exit CloseLock([&]() { sys::fs::closeFile(F&: LockFile); }); |
| 375 | if (std::error_code EC = tryLockFileThreadSafe(FD: LockFD)) { |
| 376 | if (EC == std::errc::no_lock_available) |
| 377 | return createFileError( |
| 378 | F: PathBuf, EC, |
| 379 | Fmt: "CAS validation requires exclusive access but CAS was in use" ); |
| 380 | return createFileError(F: PathBuf, EC); |
| 381 | } |
| 382 | llvm::scope_exit UnlockFD([&]() { unlockFileThreadSafe(FD: LockFD); }); |
| 383 | |
| 384 | auto DBDirs = getAllDBDirs(Path: RootPath); |
| 385 | if (!DBDirs) |
| 386 | return DBDirs.takeError(); |
| 387 | |
| 388 | for (StringRef DBDir : *DBDirs) { |
| 389 | sys::path::remove_filename(path&: PathBuf); |
| 390 | sys::path::append(path&: PathBuf, a: DBDir); |
| 391 | std::error_code EC; |
| 392 | int Attempt = 0, MaxAttempts = 100; |
| 393 | SmallString<128> GCPath; |
| 394 | for (; Attempt < MaxAttempts; ++Attempt) { |
| 395 | GCPath.assign(RHS: RootPath); |
| 396 | sys::path::append(path&: GCPath, a: CorruptPrefix + std::to_string(val: Attempt) + |
| 397 | "." + DBDir); |
| 398 | EC = sys::fs::rename(from: PathBuf, to: GCPath); |
| 399 | // Darwin uses ENOTEMPTY. Linux may return either ENOTEMPTY or EEXIST. |
| 400 | if (EC != errc::directory_not_empty && EC != errc::file_exists) |
| 401 | break; |
| 402 | } |
| 403 | if (Attempt == MaxAttempts) |
| 404 | return createStringError( |
| 405 | EC, S: "rename " + PathBuf + |
| 406 | " failed: too many CAS directories awaiting pruning" ); |
| 407 | if (EC) |
| 408 | return createStringError(EC, S: "rename " + PathBuf + " to " + GCPath + |
| 409 | " failed: " + EC.message()); |
| 410 | } |
| 411 | Recovered = true; |
| 412 | } |
| 413 | |
| 414 | if (ValidationBootTime != BootTime) { |
| 415 | // Fix filename in case we have error to report. |
| 416 | sys::path::remove_filename(path&: PathBuf); |
| 417 | sys::path::append(path&: PathBuf, a: ValidationFilename); |
| 418 | if (std::error_code EC = sys::fs::resize_file(FD, Size: 0)) |
| 419 | return createFileError(F: PathBuf, EC); |
| 420 | raw_fd_ostream OS(FD, /*shouldClose=*/false); |
| 421 | OS.seek(off: 0); // resize does not reset position |
| 422 | OS << BootTime << '\n'; |
| 423 | if (OS.has_error()) |
| 424 | return createFileError(F: PathBuf, EC: OS.error()); |
| 425 | } |
| 426 | |
| 427 | return NeedsRecovery ? ValidationResult::Recovered : ValidationResult::Valid; |
| 428 | } |
| 429 | |
| 430 | Expected<std::unique_ptr<UnifiedOnDiskCache>> |
| 431 | UnifiedOnDiskCache::open(StringRef RootPath, std::optional<uint64_t> SizeLimit, |
| 432 | StringRef HashName, unsigned HashByteSize, |
| 433 | OnDiskGraphDB::FaultInPolicy FaultInPolicy) { |
| 434 | auto BypassSandbox = sys::sandbox::scopedDisable(); |
| 435 | |
| 436 | if (std::error_code EC = sys::fs::create_directories(path: RootPath)) |
| 437 | return createFileError(F: RootPath, EC); |
| 438 | |
| 439 | SmallString<256> PathBuf(RootPath); |
| 440 | sys::path::append(path&: PathBuf, a: "lock" ); |
| 441 | int LockFD = -1; |
| 442 | if (std::error_code EC = sys::fs::openFileForReadWrite( |
| 443 | Name: PathBuf, ResultFD&: LockFD, Disp: sys::fs::CD_OpenAlways, Flags: sys::fs::OF_None)) |
| 444 | return createFileError(F: PathBuf, EC); |
| 445 | assert(LockFD != -1); |
| 446 | // Locking the directory using shared lock, which will prevent other processes |
| 447 | // from creating a new chain (essentially while a \p UnifiedOnDiskCache |
| 448 | // instance holds a shared lock the storage for the primary directory will |
| 449 | // grow unrestricted). |
| 450 | if (std::error_code EC = |
| 451 | lockFileThreadSafe(FD: LockFD, Kind: sys::fs::LockKind::Shared)) |
| 452 | return createFileError(F: PathBuf, EC); |
| 453 | |
| 454 | auto DBDirs = getAllDBDirs(Path: RootPath); |
| 455 | if (!DBDirs) |
| 456 | return DBDirs.takeError(); |
| 457 | if (DBDirs->empty()) |
| 458 | DBDirs->push_back(Elt: (Twine(DBDirPrefix) + "1" ).str()); |
| 459 | |
| 460 | std::shared_ptr<ondisk::OnDiskCASLogger> Logger; |
| 461 | #ifndef _WIN32 |
| 462 | if (Error E = |
| 463 | ondisk::OnDiskCASLogger::openIfEnabled(Path: RootPath).moveInto(Value&: Logger)) |
| 464 | return std::move(E); |
| 465 | #endif |
| 466 | |
| 467 | /// If there is only one directory open databases on it. If there are 2 or |
| 468 | /// more directories, get the most recent directories and chain them, with the |
| 469 | /// most recent being the primary one. The remaining directories are unused |
| 470 | /// data than can be garbage-collected. |
| 471 | auto UniDB = std::unique_ptr<UnifiedOnDiskCache>(new UnifiedOnDiskCache()); |
| 472 | std::unique_ptr<OnDiskGraphDB> UpstreamGraphDB; |
| 473 | std::unique_ptr<OnDiskKeyValueDB> UpstreamKVDB; |
| 474 | if (DBDirs->size() > 1) { |
| 475 | StringRef UpstreamDir = *(DBDirs->end() - 2); |
| 476 | PathBuf = RootPath; |
| 477 | sys::path::append(path&: PathBuf, a: UpstreamDir); |
| 478 | if (Error E = |
| 479 | OnDiskGraphDB::open(Path: PathBuf, HashName, HashByteSize, |
| 480 | /*UpstreamDB=*/nullptr, Logger, Policy: FaultInPolicy) |
| 481 | .moveInto(Value&: UpstreamGraphDB)) |
| 482 | return std::move(E); |
| 483 | if (Error E = OnDiskKeyValueDB::open(Path: PathBuf, HashName, KeySize: HashByteSize, |
| 484 | /*ValueName=*/"objectid" , |
| 485 | /*ValueSize=*/sizeof(uint64_t), |
| 486 | /*UnifiedCache=*/nullptr, Logger) |
| 487 | .moveInto(Value&: UpstreamKVDB)) |
| 488 | return std::move(E); |
| 489 | } |
| 490 | |
| 491 | StringRef PrimaryDir = *(DBDirs->end() - 1); |
| 492 | PathBuf = RootPath; |
| 493 | sys::path::append(path&: PathBuf, a: PrimaryDir); |
| 494 | std::unique_ptr<OnDiskGraphDB> PrimaryGraphDB; |
| 495 | if (Error E = |
| 496 | OnDiskGraphDB::open(Path: PathBuf, HashName, HashByteSize, |
| 497 | UpstreamDB: UpstreamGraphDB.get(), Logger, Policy: FaultInPolicy) |
| 498 | .moveInto(Value&: PrimaryGraphDB)) |
| 499 | return std::move(E); |
| 500 | std::unique_ptr<OnDiskKeyValueDB> PrimaryKVDB; |
| 501 | // \p UnifiedOnDiskCache does manual chaining for key-value requests, |
| 502 | // including an extra translation step of the value during fault-in. |
| 503 | if (Error E = OnDiskKeyValueDB::open(Path: PathBuf, HashName, KeySize: HashByteSize, |
| 504 | /*ValueName=*/"objectid" , |
| 505 | /*ValueSize=*/sizeof(uint64_t), |
| 506 | UnifiedCache: UniDB.get(), Logger) |
| 507 | .moveInto(Value&: PrimaryKVDB)) |
| 508 | return std::move(E); |
| 509 | |
| 510 | UniDB->RootPath = RootPath; |
| 511 | UniDB->SizeLimit = SizeLimit.value_or(u: 0); |
| 512 | UniDB->LockFD = LockFD; |
| 513 | UniDB->NeedsGarbageCollection = DBDirs->size() > 2; |
| 514 | UniDB->PrimaryDBDir = PrimaryDir; |
| 515 | UniDB->UpstreamGraphDB = std::move(UpstreamGraphDB); |
| 516 | UniDB->PrimaryGraphDB = std::move(PrimaryGraphDB); |
| 517 | UniDB->UpstreamKVDB = std::move(UpstreamKVDB); |
| 518 | UniDB->PrimaryKVDB = std::move(PrimaryKVDB); |
| 519 | UniDB->Logger = std::move(Logger); |
| 520 | |
| 521 | return std::move(UniDB); |
| 522 | } |
| 523 | |
| 524 | void UnifiedOnDiskCache::setSizeLimit(std::optional<uint64_t> SizeLimit) { |
| 525 | this->SizeLimit = SizeLimit.value_or(u: 0); |
| 526 | } |
| 527 | |
| 528 | uint64_t UnifiedOnDiskCache::getStorageSize() const { |
| 529 | uint64_t TotalSize = getPrimaryStorageSize(); |
| 530 | if (UpstreamGraphDB) |
| 531 | TotalSize += UpstreamGraphDB->getStorageSize(); |
| 532 | if (UpstreamKVDB) |
| 533 | TotalSize += UpstreamKVDB->getStorageSize(); |
| 534 | return TotalSize; |
| 535 | } |
| 536 | |
| 537 | uint64_t UnifiedOnDiskCache::getPrimaryStorageSize() const { |
| 538 | return PrimaryGraphDB->getStorageSize() + PrimaryKVDB->getStorageSize(); |
| 539 | } |
| 540 | |
| 541 | bool UnifiedOnDiskCache::hasExceededSizeLimit() const { |
| 542 | uint64_t CurSizeLimit = SizeLimit; |
| 543 | if (!CurSizeLimit) |
| 544 | return false; |
| 545 | |
| 546 | // If the hard limit is beyond 85%, declare above limit and request clean up. |
| 547 | unsigned CurrentPercent = |
| 548 | std::max(a: PrimaryGraphDB->getHardStorageLimitUtilization(), |
| 549 | b: PrimaryKVDB->getHardStorageLimitUtilization()); |
| 550 | if (CurrentPercent > 85) |
| 551 | return true; |
| 552 | |
| 553 | // We allow each of the directories in the chain to reach up to half the |
| 554 | // intended size limit. Check whether the primary directory has exceeded half |
| 555 | // the limit or not, in order to decide whether we need to start a new chain. |
| 556 | // |
| 557 | // We could check the size limit against the sum of sizes of both the primary |
| 558 | // and upstream directories but then if the upstream is significantly larger |
| 559 | // than the intended limit, it would trigger a new chain to be created before |
| 560 | // the primary has reached its own limit. Essentially in such situation we |
| 561 | // prefer reclaiming the storage later in order to have more consistent cache |
| 562 | // hits behavior. |
| 563 | return (CurSizeLimit / 2) < getPrimaryStorageSize(); |
| 564 | } |
| 565 | |
| 566 | Error UnifiedOnDiskCache::close(bool CheckSizeLimit) { |
| 567 | auto BypassSandbox = sys::sandbox::scopedDisable(); |
| 568 | |
| 569 | if (LockFD == -1) |
| 570 | return Error::success(); // already closed. |
| 571 | llvm::scope_exit CloseLock([&]() { |
| 572 | assert(LockFD >= 0); |
| 573 | sys::fs::file_t LockFile = sys::fs::convertFDToNativeFile(FD: LockFD); |
| 574 | sys::fs::closeFile(F&: LockFile); |
| 575 | LockFD = -1; |
| 576 | }); |
| 577 | |
| 578 | bool ExceededSizeLimit = CheckSizeLimit ? hasExceededSizeLimit() : false; |
| 579 | UpstreamKVDB.reset(); |
| 580 | PrimaryKVDB.reset(); |
| 581 | UpstreamGraphDB.reset(); |
| 582 | PrimaryGraphDB.reset(); |
| 583 | if (std::error_code EC = unlockFileThreadSafe(FD: LockFD)) |
| 584 | return createFileError(F: RootPath, EC); |
| 585 | |
| 586 | if (!ExceededSizeLimit) |
| 587 | return Error::success(); |
| 588 | |
| 589 | // The primary directory exceeded its intended size limit. Try to get an |
| 590 | // exclusive lock in order to create a new primary directory for next time |
| 591 | // this \p UnifiedOnDiskCache path is opened. |
| 592 | |
| 593 | if (std::error_code EC = tryLockFileThreadSafe( |
| 594 | FD: LockFD, Timeout: std::chrono::milliseconds(0), Kind: sys::fs::LockKind::Exclusive)) { |
| 595 | if (EC == errc::no_lock_available) |
| 596 | return Error::success(); // couldn't get exclusive lock, give up. |
| 597 | return createFileError(F: RootPath, EC); |
| 598 | } |
| 599 | llvm::scope_exit UnlockFile([&]() { unlockFileThreadSafe(FD: LockFD); }); |
| 600 | |
| 601 | // Managed to get an exclusive lock which means there are no other open |
| 602 | // \p UnifiedOnDiskCache instances for the same path, so we can safely start a |
| 603 | // new primary directory. To start a new primary directory we just have to |
| 604 | // create a new empty directory with the next consecutive index; since this is |
| 605 | // an atomic operation we will leave the top-level directory in a consistent |
| 606 | // state even if the process dies during this code-path. |
| 607 | |
| 608 | SmallString<256> PathBuf(RootPath); |
| 609 | raw_svector_ostream OS(PathBuf); |
| 610 | OS << sys::path::get_separator(); |
| 611 | getNextDBDirName(DBDir: PrimaryDBDir, OS); |
| 612 | if (std::error_code EC = sys::fs::create_directory(path: PathBuf)) |
| 613 | return createFileError(F: PathBuf, EC); |
| 614 | |
| 615 | NeedsGarbageCollection = true; |
| 616 | return Error::success(); |
| 617 | } |
| 618 | |
| 619 | UnifiedOnDiskCache::UnifiedOnDiskCache() = default; |
| 620 | |
| 621 | UnifiedOnDiskCache::~UnifiedOnDiskCache() { consumeError(Err: close()); } |
| 622 | |
| 623 | Error UnifiedOnDiskCache::collectGarbage(StringRef Path, |
| 624 | ondisk::OnDiskCASLogger *Logger) { |
| 625 | auto DBDirs = getAllGarbageDirs(Path); |
| 626 | if (!DBDirs) |
| 627 | return DBDirs.takeError(); |
| 628 | |
| 629 | SmallString<256> PathBuf(Path); |
| 630 | for (StringRef UnusedSubDir : *DBDirs) { |
| 631 | sys::path::append(path&: PathBuf, a: UnusedSubDir); |
| 632 | if (Logger) |
| 633 | Logger->logUnifiedOnDiskCacheCollectGarbage(Path: PathBuf); |
| 634 | if (std::error_code EC = sys::fs::remove_directories(path: PathBuf)) |
| 635 | return createFileError(F: PathBuf, EC); |
| 636 | sys::path::remove_filename(path&: PathBuf); |
| 637 | } |
| 638 | return Error::success(); |
| 639 | } |
| 640 | |
| 641 | Error UnifiedOnDiskCache::collectGarbage() { |
| 642 | return collectGarbage(Path: RootPath, Logger: Logger.get()); |
| 643 | } |
| 644 | |