1 //===- Trace.cpp - XRay Trace Loading implementation. ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // XRay log reader implementation. 11 // 12 //===----------------------------------------------------------------------===// 13 #include "llvm/XRay/Trace.h" 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/Support/DataExtractor.h" 16 #include "llvm/Support/Error.h" 17 #include "llvm/Support/FileSystem.h" 18 #include "llvm/XRay/YAMLXRayRecord.h" 19 20 using namespace llvm; 21 using namespace llvm::xray; 22 using llvm::yaml::Input; 23 24 namespace { 25 using XRayRecordStorage = 26 std::aligned_storage<sizeof(XRayRecord), alignof(XRayRecord)>::type; 27 28 // Populates the FileHeader reference by reading the first 32 bytes of the file. 29 Error readBinaryFormatHeader(StringRef Data, XRayFileHeader &FileHeader) { 30 // FIXME: Maybe deduce whether the data is little or big-endian using some 31 // magic bytes in the beginning of the file? 32 33 // First 32 bytes of the file will always be the header. We assume a certain 34 // format here: 35 // 36 // (2) uint16 : version 37 // (2) uint16 : type 38 // (4) uint32 : bitfield 39 // (8) uint64 : cycle frequency 40 // (16) - : padding 41 42 DataExtractor HeaderExtractor(Data, true, 8); 43 uint32_t OffsetPtr = 0; 44 FileHeader.Version = HeaderExtractor.getU16(&OffsetPtr); 45 FileHeader.Type = HeaderExtractor.getU16(&OffsetPtr); 46 uint32_t Bitfield = HeaderExtractor.getU32(&OffsetPtr); 47 FileHeader.ConstantTSC = Bitfield & 1uL; 48 FileHeader.NonstopTSC = Bitfield & 1uL << 1; 49 FileHeader.CycleFrequency = HeaderExtractor.getU64(&OffsetPtr); 50 std::memcpy(&FileHeader.FreeFormData, Data.bytes_begin() + OffsetPtr, 16); 51 if (FileHeader.Version != 1 && FileHeader.Version != 2) 52 return make_error<StringError>( 53 Twine("Unsupported XRay file version: ") + Twine(FileHeader.Version), 54 std::make_error_code(std::errc::invalid_argument)); 55 return Error::success(); 56 } 57 58 Error loadNaiveFormatLog(StringRef Data, XRayFileHeader &FileHeader, 59 std::vector<XRayRecord> &Records) { 60 if (Data.size() < 32) 61 return make_error<StringError>( 62 "Not enough bytes for an XRay log.", 63 std::make_error_code(std::errc::invalid_argument)); 64 65 if (Data.size() - 32 == 0 || Data.size() % 32 != 0) 66 return make_error<StringError>( 67 "Invalid-sized XRay data.", 68 std::make_error_code(std::errc::invalid_argument)); 69 70 if (auto E = readBinaryFormatHeader(Data, FileHeader)) 71 return E; 72 73 // Each record after the header will be 32 bytes, in the following format: 74 // 75 // (2) uint16 : record type 76 // (1) uint8 : cpu id 77 // (1) uint8 : type 78 // (4) sint32 : function id 79 // (8) uint64 : tsc 80 // (4) uint32 : thread id 81 // (12) - : padding 82 for (auto S = Data.drop_front(32); !S.empty(); S = S.drop_front(32)) { 83 DataExtractor RecordExtractor(S, true, 8); 84 uint32_t OffsetPtr = 0; 85 switch (auto RecordType = RecordExtractor.getU16(&OffsetPtr)) { 86 case 0: { // Normal records. 87 Records.emplace_back(); 88 auto &Record = Records.back(); 89 Record.RecordType = RecordType; 90 Record.CPU = RecordExtractor.getU8(&OffsetPtr); 91 auto Type = RecordExtractor.getU8(&OffsetPtr); 92 switch (Type) { 93 case 0: 94 Record.Type = RecordTypes::ENTER; 95 break; 96 case 1: 97 Record.Type = RecordTypes::EXIT; 98 break; 99 case 2: 100 Record.Type = RecordTypes::TAIL_EXIT; 101 break; 102 case 3: 103 Record.Type = RecordTypes::ENTER_ARG; 104 break; 105 default: 106 return make_error<StringError>( 107 Twine("Unknown record type '") + Twine(int{Type}) + "'", 108 std::make_error_code(std::errc::executable_format_error)); 109 } 110 Record.FuncId = RecordExtractor.getSigned(&OffsetPtr, sizeof(int32_t)); 111 Record.TSC = RecordExtractor.getU64(&OffsetPtr); 112 Record.TId = RecordExtractor.getU32(&OffsetPtr); 113 break; 114 } 115 case 1: { // Arg payload record. 116 auto &Record = Records.back(); 117 // Advance two bytes to avoid padding. 118 OffsetPtr += 2; 119 int32_t FuncId = RecordExtractor.getSigned(&OffsetPtr, sizeof(int32_t)); 120 auto TId = RecordExtractor.getU32(&OffsetPtr); 121 if (Record.FuncId != FuncId || Record.TId != TId) 122 return make_error<StringError>( 123 Twine("Corrupted log, found payload following non-matching " 124 "function + thread record. Record for ") + 125 Twine(Record.FuncId) + " != " + Twine(FuncId), 126 std::make_error_code(std::errc::executable_format_error)); 127 // Advance another four bytes to avoid padding. 128 OffsetPtr += 4; 129 auto Arg = RecordExtractor.getU64(&OffsetPtr); 130 Record.CallArgs.push_back(Arg); 131 break; 132 } 133 default: 134 return make_error<StringError>( 135 Twine("Unknown record type == ") + Twine(RecordType), 136 std::make_error_code(std::errc::executable_format_error)); 137 } 138 } 139 return Error::success(); 140 } 141 142 /// When reading from a Flight Data Recorder mode log, metadata records are 143 /// sparse compared to packed function records, so we must maintain state as we 144 /// read through the sequence of entries. This allows the reader to denormalize 145 /// the CPUId and Thread Id onto each Function Record and transform delta 146 /// encoded TSC values into absolute encodings on each record. 147 struct FDRState { 148 uint16_t CPUId; 149 uint16_t ThreadId; 150 uint64_t BaseTSC; 151 152 /// Encode some of the state transitions for the FDR log reader as explicit 153 /// checks. These are expectations for the next Record in the stream. 154 enum class Token { 155 NEW_BUFFER_RECORD_OR_EOF, 156 WALLCLOCK_RECORD, 157 NEW_CPU_ID_RECORD, 158 FUNCTION_SEQUENCE, 159 SCAN_TO_END_OF_THREAD_BUF, 160 CUSTOM_EVENT_DATA, 161 CALL_ARGUMENT, 162 BUFFER_EXTENTS, 163 }; 164 Token Expects; 165 166 // Each threads buffer may have trailing garbage to scan over, so we track our 167 // progress. 168 uint64_t CurrentBufferSize; 169 uint64_t CurrentBufferConsumed; 170 }; 171 172 const char *fdrStateToTwine(const FDRState::Token &state) { 173 switch (state) { 174 case FDRState::Token::NEW_BUFFER_RECORD_OR_EOF: 175 return "NEW_BUFFER_RECORD_OR_EOF"; 176 case FDRState::Token::WALLCLOCK_RECORD: 177 return "WALLCLOCK_RECORD"; 178 case FDRState::Token::NEW_CPU_ID_RECORD: 179 return "NEW_CPU_ID_RECORD"; 180 case FDRState::Token::FUNCTION_SEQUENCE: 181 return "FUNCTION_SEQUENCE"; 182 case FDRState::Token::SCAN_TO_END_OF_THREAD_BUF: 183 return "SCAN_TO_END_OF_THREAD_BUF"; 184 case FDRState::Token::CUSTOM_EVENT_DATA: 185 return "CUSTOM_EVENT_DATA"; 186 case FDRState::Token::CALL_ARGUMENT: 187 return "CALL_ARGUMENT"; 188 case FDRState::Token::BUFFER_EXTENTS: 189 return "BUFFER_EXTENTS"; 190 } 191 return "UNKNOWN"; 192 } 193 194 /// State transition when a NewBufferRecord is encountered. 195 Error processFDRNewBufferRecord(FDRState &State, uint8_t RecordFirstByte, 196 DataExtractor &RecordExtractor) { 197 198 if (State.Expects != FDRState::Token::NEW_BUFFER_RECORD_OR_EOF) 199 return make_error<StringError>( 200 Twine("Malformed log. Read New Buffer record kind out of sequence; " 201 "expected: ") + 202 fdrStateToTwine(State.Expects), 203 std::make_error_code(std::errc::executable_format_error)); 204 uint32_t OffsetPtr = 1; // 1 byte into record. 205 State.ThreadId = RecordExtractor.getU16(&OffsetPtr); 206 State.Expects = FDRState::Token::WALLCLOCK_RECORD; 207 return Error::success(); 208 } 209 210 /// State transition when an EndOfBufferRecord is encountered. 211 Error processFDREndOfBufferRecord(FDRState &State, uint8_t RecordFirstByte, 212 DataExtractor &RecordExtractor) { 213 if (State.Expects == FDRState::Token::NEW_BUFFER_RECORD_OR_EOF) 214 return make_error<StringError>( 215 Twine("Malformed log. Received EOB message without current buffer; " 216 "expected: ") + 217 fdrStateToTwine(State.Expects), 218 std::make_error_code(std::errc::executable_format_error)); 219 State.Expects = FDRState::Token::SCAN_TO_END_OF_THREAD_BUF; 220 return Error::success(); 221 } 222 223 /// State transition when a NewCPUIdRecord is encountered. 224 Error processFDRNewCPUIdRecord(FDRState &State, uint8_t RecordFirstByte, 225 DataExtractor &RecordExtractor) { 226 if (State.Expects != FDRState::Token::FUNCTION_SEQUENCE && 227 State.Expects != FDRState::Token::NEW_CPU_ID_RECORD) 228 return make_error<StringError>( 229 Twine("Malformed log. Read NewCPUId record kind out of sequence; " 230 "expected: ") + 231 fdrStateToTwine(State.Expects), 232 std::make_error_code(std::errc::executable_format_error)); 233 uint32_t OffsetPtr = 1; // Read starting after the first byte. 234 State.CPUId = RecordExtractor.getU16(&OffsetPtr); 235 State.BaseTSC = RecordExtractor.getU64(&OffsetPtr); 236 State.Expects = FDRState::Token::FUNCTION_SEQUENCE; 237 return Error::success(); 238 } 239 240 /// State transition when a TSCWrapRecord (overflow detection) is encountered. 241 Error processFDRTSCWrapRecord(FDRState &State, uint8_t RecordFirstByte, 242 DataExtractor &RecordExtractor) { 243 if (State.Expects != FDRState::Token::FUNCTION_SEQUENCE) 244 return make_error<StringError>( 245 Twine("Malformed log. Read TSCWrap record kind out of sequence; " 246 "expecting: ") + 247 fdrStateToTwine(State.Expects), 248 std::make_error_code(std::errc::executable_format_error)); 249 uint32_t OffsetPtr = 1; // Read starting after the first byte. 250 State.BaseTSC = RecordExtractor.getU64(&OffsetPtr); 251 return Error::success(); 252 } 253 254 /// State transition when a WallTimeMarkerRecord is encountered. 255 Error processFDRWallTimeRecord(FDRState &State, uint8_t RecordFirstByte, 256 DataExtractor &RecordExtractor) { 257 if (State.Expects != FDRState::Token::WALLCLOCK_RECORD) 258 return make_error<StringError>( 259 Twine("Malformed log. Read Wallclock record kind out of sequence; " 260 "expecting: ") + 261 fdrStateToTwine(State.Expects), 262 std::make_error_code(std::errc::executable_format_error)); 263 264 // TODO: Someday, reconcile the TSC ticks to wall clock time for presentation 265 // purposes. For now, we're ignoring these records. 266 State.Expects = FDRState::Token::NEW_CPU_ID_RECORD; 267 return Error::success(); 268 } 269 270 /// State transition when a CustomEventMarker is encountered. 271 Error processCustomEventMarker(FDRState &State, uint8_t RecordFirstByte, 272 DataExtractor &RecordExtractor, 273 size_t &RecordSize) { 274 // We can encounter a CustomEventMarker anywhere in the log, so we can handle 275 // it regardless of the expectation. However, we do set the expectation to 276 // read a set number of fixed bytes, as described in the metadata. 277 uint32_t OffsetPtr = 1; // Read after the first byte. 278 uint32_t DataSize = RecordExtractor.getU32(&OffsetPtr); 279 uint64_t TSC = RecordExtractor.getU64(&OffsetPtr); 280 281 // FIXME: Actually represent the record through the API. For now we only 282 // skip through the data. 283 (void)TSC; 284 RecordSize = 16 + DataSize; 285 return Error::success(); 286 } 287 288 /// State transition when an BufferExtents record is encountered. 289 Error processBufferExtents(FDRState &State, uint8_t RecordFirstByte, 290 DataExtractor &RecordExtractor) { 291 if (State.Expects != FDRState::Token::BUFFER_EXTENTS) 292 return make_error<StringError>( 293 Twine("Malformed log. Buffer Extents unexpected; expected: ") + 294 fdrStateToTwine(State.Expects), 295 std::make_error_code(std::errc::executable_format_error)); 296 uint32_t OffsetPtr = 1; // Read after the first byte. 297 State.CurrentBufferSize = RecordExtractor.getU64(&OffsetPtr); 298 State.Expects = FDRState::Token::NEW_BUFFER_RECORD_OR_EOF; 299 return Error::success(); 300 } 301 302 /// State transition when a CallArgumentRecord is encountered. 303 Error processFDRCallArgumentRecord(FDRState &State, uint8_t RecordFirstByte, 304 DataExtractor &RecordExtractor, 305 std::vector<XRayRecord> &Records) { 306 uint32_t OffsetPtr = 1; // Read starting after the first byte. 307 auto &Enter = Records.back(); 308 309 if (Enter.Type != RecordTypes::ENTER) 310 return make_error<StringError>( 311 "CallArgument needs to be right after a function entry", 312 std::make_error_code(std::errc::executable_format_error)); 313 Enter.Type = RecordTypes::ENTER_ARG; 314 Enter.CallArgs.emplace_back(RecordExtractor.getU64(&OffsetPtr)); 315 return Error::success(); 316 } 317 318 /// Advances the state machine for reading the FDR record type by reading one 319 /// Metadata Record and updating the State appropriately based on the kind of 320 /// record encountered. The RecordKind is encoded in the first byte of the 321 /// Record, which the caller should pass in because they have already read it 322 /// to determine that this is a metadata record as opposed to a function record. 323 /// 324 /// Beginning with Version 2 of the FDR log, we do not depend on the size of the 325 /// buffer, but rather use the extents to determine how far to read in the log 326 /// for this particular buffer. 327 Error processFDRMetadataRecord(FDRState &State, uint8_t RecordFirstByte, 328 DataExtractor &RecordExtractor, 329 size_t &RecordSize, 330 std::vector<XRayRecord> &Records, 331 uint16_t Version) { 332 // The remaining 7 bits are the RecordKind enum. 333 uint8_t RecordKind = RecordFirstByte >> 1; 334 switch (RecordKind) { 335 case 0: // NewBuffer 336 if (auto E = 337 processFDRNewBufferRecord(State, RecordFirstByte, RecordExtractor)) 338 return E; 339 break; 340 case 1: // EndOfBuffer 341 if (Version >= 2) 342 return make_error<StringError>( 343 "Since Version 2 of FDR logging, we no longer support EOB records.", 344 std::make_error_code(std::errc::executable_format_error)); 345 if (auto E = processFDREndOfBufferRecord(State, RecordFirstByte, 346 RecordExtractor)) 347 return E; 348 break; 349 case 2: // NewCPUId 350 if (auto E = 351 processFDRNewCPUIdRecord(State, RecordFirstByte, RecordExtractor)) 352 return E; 353 break; 354 case 3: // TSCWrap 355 if (auto E = 356 processFDRTSCWrapRecord(State, RecordFirstByte, RecordExtractor)) 357 return E; 358 break; 359 case 4: // WallTimeMarker 360 if (auto E = 361 processFDRWallTimeRecord(State, RecordFirstByte, RecordExtractor)) 362 return E; 363 break; 364 case 5: // CustomEventMarker 365 if (auto E = processCustomEventMarker(State, RecordFirstByte, 366 RecordExtractor, RecordSize)) 367 return E; 368 break; 369 case 6: // CallArgument 370 if (auto E = processFDRCallArgumentRecord(State, RecordFirstByte, 371 RecordExtractor, Records)) 372 return E; 373 break; 374 case 7: // BufferExtents 375 if (auto E = processBufferExtents(State, RecordFirstByte, RecordExtractor)) 376 return E; 377 break; 378 default: 379 // Widen the record type to uint16_t to prevent conversion to char. 380 return make_error<StringError>( 381 Twine("Illegal metadata record type: ") 382 .concat(Twine(static_cast<unsigned>(RecordKind))), 383 std::make_error_code(std::errc::executable_format_error)); 384 } 385 return Error::success(); 386 } 387 388 /// Reads a function record from an FDR format log, appending a new XRayRecord 389 /// to the vector being populated and updating the State with a new value 390 /// reference value to interpret TSC deltas. 391 /// 392 /// The XRayRecord constructed includes information from the function record 393 /// processed here as well as Thread ID and CPU ID formerly extracted into 394 /// State. 395 Error processFDRFunctionRecord(FDRState &State, uint8_t RecordFirstByte, 396 DataExtractor &RecordExtractor, 397 std::vector<XRayRecord> &Records) { 398 switch (State.Expects) { 399 case FDRState::Token::NEW_BUFFER_RECORD_OR_EOF: 400 return make_error<StringError>( 401 "Malformed log. Received Function Record before new buffer setup.", 402 std::make_error_code(std::errc::executable_format_error)); 403 case FDRState::Token::WALLCLOCK_RECORD: 404 return make_error<StringError>( 405 "Malformed log. Received Function Record when expecting wallclock.", 406 std::make_error_code(std::errc::executable_format_error)); 407 case FDRState::Token::NEW_CPU_ID_RECORD: 408 return make_error<StringError>( 409 "Malformed log. Received Function Record before first CPU record.", 410 std::make_error_code(std::errc::executable_format_error)); 411 default: 412 Records.emplace_back(); 413 auto &Record = Records.back(); 414 Record.RecordType = 0; // Record is type NORMAL. 415 // Strip off record type bit and use the next three bits. 416 uint8_t RecordType = (RecordFirstByte >> 1) & 0x07; 417 switch (RecordType) { 418 case static_cast<uint8_t>(RecordTypes::ENTER): 419 Record.Type = RecordTypes::ENTER; 420 break; 421 case static_cast<uint8_t>(RecordTypes::EXIT): 422 Record.Type = RecordTypes::EXIT; 423 break; 424 case static_cast<uint8_t>(RecordTypes::TAIL_EXIT): 425 Record.Type = RecordTypes::TAIL_EXIT; 426 break; 427 default: 428 // Cast to an unsigned integer to not interpret the record type as a char. 429 return make_error<StringError>( 430 Twine("Illegal function record type: ") 431 .concat(Twine(static_cast<unsigned>(RecordType))), 432 std::make_error_code(std::errc::executable_format_error)); 433 } 434 Record.CPU = State.CPUId; 435 Record.TId = State.ThreadId; 436 // Back up to read first 32 bits, including the 4 we pulled RecordType 437 // and RecordKind out of. The remaining 28 are FunctionId. 438 uint32_t OffsetPtr = 0; 439 // Despite function Id being a signed int on XRayRecord, 440 // when it is written to an FDR format, the top bits are truncated, 441 // so it is effectively an unsigned value. When we shift off the 442 // top four bits, we want the shift to be logical, so we read as 443 // uint32_t. 444 uint32_t FuncIdBitField = RecordExtractor.getU32(&OffsetPtr); 445 Record.FuncId = FuncIdBitField >> 4; 446 // FunctionRecords have a 32 bit delta from the previous absolute TSC 447 // or TSC delta. If this would overflow, we should read a TSCWrap record 448 // with an absolute TSC reading. 449 uint64_t NewTSC = State.BaseTSC + RecordExtractor.getU32(&OffsetPtr); 450 State.BaseTSC = NewTSC; 451 Record.TSC = NewTSC; 452 } 453 return Error::success(); 454 } 455 456 /// Reads a log in FDR mode for version 1 of this binary format. FDR mode is 457 /// defined as part of the compiler-rt project in xray_fdr_logging.h, and such 458 /// a log consists of the familiar 32 bit XRayHeader, followed by sequences of 459 /// of interspersed 16 byte Metadata Records and 8 byte Function Records. 460 /// 461 /// The following is an attempt to document the grammar of the format, which is 462 /// parsed by this function for little-endian machines. Since the format makes 463 /// use of BitFields, when we support big-endian architectures, we will need to 464 /// adjust not only the endianness parameter to llvm's RecordExtractor, but also 465 /// the bit twiddling logic, which is consistent with the little-endian 466 /// convention that BitFields within a struct will first be packed into the 467 /// least significant bits the address they belong to. 468 /// 469 /// We expect a format complying with the grammar in the following pseudo-EBNF 470 /// in Version 1 of the FDR log. 471 /// 472 /// FDRLog: XRayFileHeader ThreadBuffer* 473 /// XRayFileHeader: 32 bytes to identify the log as FDR with machine metadata. 474 /// Includes BufferSize 475 /// ThreadBuffer: NewBuffer WallClockTime NewCPUId FunctionSequence EOB 476 /// BufSize: 8 byte unsigned integer indicating how large the buffer is. 477 /// NewBuffer: 16 byte metadata record with Thread Id. 478 /// WallClockTime: 16 byte metadata record with human readable time. 479 /// NewCPUId: 16 byte metadata record with CPUId and a 64 bit TSC reading. 480 /// EOB: 16 byte record in a thread buffer plus mem garbage to fill BufSize. 481 /// FunctionSequence: NewCPUId | TSCWrap | FunctionRecord 482 /// TSCWrap: 16 byte metadata record with a full 64 bit TSC reading. 483 /// FunctionRecord: 8 byte record with FunctionId, entry/exit, and TSC delta. 484 /// 485 /// In Version 2, we make the following changes: 486 /// 487 /// ThreadBuffer: BufferExtents NewBuffer WallClockTime NewCPUId 488 /// FunctionSequence 489 /// BufferExtents: 16 byte metdata record describing how many usable bytes are 490 /// in the buffer. This is measured from the start of the buffer 491 /// and must always be at least 48 (bytes). 492 /// EOB: *deprecated* 493 Error loadFDRLog(StringRef Data, XRayFileHeader &FileHeader, 494 std::vector<XRayRecord> &Records) { 495 if (Data.size() < 32) 496 return make_error<StringError>( 497 "Not enough bytes for an XRay log.", 498 std::make_error_code(std::errc::invalid_argument)); 499 500 // For an FDR log, there are records sized 16 and 8 bytes. 501 // There actually may be no records if no non-trivial functions are 502 // instrumented. 503 if (Data.size() % 8 != 0) 504 return make_error<StringError>( 505 "Invalid-sized XRay data.", 506 std::make_error_code(std::errc::invalid_argument)); 507 508 if (auto E = readBinaryFormatHeader(Data, FileHeader)) 509 return E; 510 511 uint64_t BufferSize = 0; 512 { 513 StringRef ExtraDataRef(FileHeader.FreeFormData, 16); 514 DataExtractor ExtraDataExtractor(ExtraDataRef, true, 8); 515 uint32_t ExtraDataOffset = 0; 516 BufferSize = ExtraDataExtractor.getU64(&ExtraDataOffset); 517 } 518 519 FDRState::Token InitialExpectation; 520 switch (FileHeader.Version) { 521 case 1: 522 InitialExpectation = FDRState::Token::NEW_BUFFER_RECORD_OR_EOF; 523 break; 524 case 2: 525 InitialExpectation = FDRState::Token::BUFFER_EXTENTS; 526 break; 527 default: 528 return make_error<StringError>( 529 Twine("Unsupported version '") + Twine(FileHeader.Version) + "'", 530 std::make_error_code(std::errc::executable_format_error)); 531 } 532 FDRState State{0, 0, 0, InitialExpectation, BufferSize, 0}; 533 534 // RecordSize will tell the loop how far to seek ahead based on the record 535 // type that we have just read. 536 size_t RecordSize = 0; 537 for (auto S = Data.drop_front(32); !S.empty(); S = S.drop_front(RecordSize)) { 538 DataExtractor RecordExtractor(S, true, 8); 539 uint32_t OffsetPtr = 0; 540 if (State.Expects == FDRState::Token::SCAN_TO_END_OF_THREAD_BUF) { 541 RecordSize = State.CurrentBufferSize - State.CurrentBufferConsumed; 542 if (S.size() < RecordSize) { 543 return make_error<StringError>( 544 Twine("Incomplete thread buffer. Expected at least ") + 545 Twine(RecordSize) + " bytes but found " + Twine(S.size()), 546 make_error_code(std::errc::invalid_argument)); 547 } 548 State.CurrentBufferConsumed = 0; 549 State.Expects = FDRState::Token::NEW_BUFFER_RECORD_OR_EOF; 550 continue; 551 } 552 uint8_t BitField = RecordExtractor.getU8(&OffsetPtr); 553 bool isMetadataRecord = BitField & 0x01uL; 554 bool isBufferExtents = 555 (BitField >> 1) == 7; // BufferExtents record kind == 7 556 if (isMetadataRecord) { 557 RecordSize = 16; 558 if (auto E = 559 processFDRMetadataRecord(State, BitField, RecordExtractor, 560 RecordSize, Records, FileHeader.Version)) 561 return E; 562 } else { // Process Function Record 563 RecordSize = 8; 564 if (auto E = processFDRFunctionRecord(State, BitField, RecordExtractor, 565 Records)) 566 return E; 567 } 568 569 // The BufferExtents record is technically not part of the buffer, so we 570 // don't count the size of that record against the buffer's actual size. 571 if (!isBufferExtents) 572 State.CurrentBufferConsumed += RecordSize; 573 assert(State.CurrentBufferConsumed <= State.CurrentBufferSize); 574 if (FileHeader.Version == 2 && 575 State.CurrentBufferSize == State.CurrentBufferConsumed) { 576 // In Version 2 of the log, we don't need to scan to the end of the thread 577 // buffer if we've already consumed all the bytes we need to. 578 State.Expects = FDRState::Token::BUFFER_EXTENTS; 579 State.CurrentBufferSize = BufferSize; 580 State.CurrentBufferConsumed = 0; 581 } 582 } 583 584 // Having iterated over everything we've been given, we've either consumed 585 // everything and ended up in the end state, or were told to skip the rest. 586 bool Finished = State.Expects == FDRState::Token::SCAN_TO_END_OF_THREAD_BUF && 587 State.CurrentBufferSize == State.CurrentBufferConsumed; 588 if ((State.Expects != FDRState::Token::NEW_BUFFER_RECORD_OR_EOF && 589 State.Expects != FDRState::Token::BUFFER_EXTENTS) && 590 !Finished) 591 return make_error<StringError>( 592 Twine("Encountered EOF with unexpected state expectation ") + 593 fdrStateToTwine(State.Expects) + 594 ". Remaining expected bytes in thread buffer total " + 595 Twine(State.CurrentBufferSize - State.CurrentBufferConsumed), 596 std::make_error_code(std::errc::executable_format_error)); 597 598 return Error::success(); 599 } 600 601 Error loadYAMLLog(StringRef Data, XRayFileHeader &FileHeader, 602 std::vector<XRayRecord> &Records) { 603 YAMLXRayTrace Trace; 604 Input In(Data); 605 In >> Trace; 606 if (In.error()) 607 return make_error<StringError>("Failed loading YAML Data.", In.error()); 608 609 FileHeader.Version = Trace.Header.Version; 610 FileHeader.Type = Trace.Header.Type; 611 FileHeader.ConstantTSC = Trace.Header.ConstantTSC; 612 FileHeader.NonstopTSC = Trace.Header.NonstopTSC; 613 FileHeader.CycleFrequency = Trace.Header.CycleFrequency; 614 615 if (FileHeader.Version != 1) 616 return make_error<StringError>( 617 Twine("Unsupported XRay file version: ") + Twine(FileHeader.Version), 618 std::make_error_code(std::errc::invalid_argument)); 619 620 Records.clear(); 621 std::transform(Trace.Records.begin(), Trace.Records.end(), 622 std::back_inserter(Records), [&](const YAMLXRayRecord &R) { 623 return XRayRecord{R.RecordType, R.CPU, R.Type, R.FuncId, 624 R.TSC, R.TId, R.CallArgs}; 625 }); 626 return Error::success(); 627 } 628 } // namespace 629 630 Expected<Trace> llvm::xray::loadTraceFile(StringRef Filename, bool Sort) { 631 int Fd; 632 if (auto EC = sys::fs::openFileForRead(Filename, Fd)) { 633 return make_error<StringError>( 634 Twine("Cannot read log from '") + Filename + "'", EC); 635 } 636 637 uint64_t FileSize; 638 if (auto EC = sys::fs::file_size(Filename, FileSize)) { 639 return make_error<StringError>( 640 Twine("Cannot read log from '") + Filename + "'", EC); 641 } 642 if (FileSize < 4) { 643 return make_error<StringError>( 644 Twine("File '") + Filename + "' too small for XRay.", 645 std::make_error_code(std::errc::executable_format_error)); 646 } 647 648 // Map the opened file into memory and use a StringRef to access it later. 649 std::error_code EC; 650 sys::fs::mapped_file_region MappedFile( 651 Fd, sys::fs::mapped_file_region::mapmode::readonly, FileSize, 0, EC); 652 if (EC) { 653 return make_error<StringError>( 654 Twine("Cannot read log from '") + Filename + "'", EC); 655 } 656 auto Data = StringRef(MappedFile.data(), MappedFile.size()); 657 658 // Attempt to detect the file type using file magic. We have a slight bias 659 // towards the binary format, and we do this by making sure that the first 4 660 // bytes of the binary file is some combination of the following byte 661 // patterns: (observe the code loading them assumes they're little endian) 662 // 663 // 0x01 0x00 0x00 0x00 - version 1, "naive" format 664 // 0x01 0x00 0x01 0x00 - version 1, "flight data recorder" format 665 // 0x02 0x00 0x01 0x00 - version 2, "flight data recorder" format 666 // 667 // YAML files don't typically have those first four bytes as valid text so we 668 // try loading assuming YAML if we don't find these bytes. 669 // 670 // Only if we can't load either the binary or the YAML format will we yield an 671 // error. 672 StringRef Magic(MappedFile.data(), 4); 673 DataExtractor HeaderExtractor(Magic, true, 8); 674 uint32_t OffsetPtr = 0; 675 uint16_t Version = HeaderExtractor.getU16(&OffsetPtr); 676 uint16_t Type = HeaderExtractor.getU16(&OffsetPtr); 677 678 enum BinaryFormatType { NAIVE_FORMAT = 0, FLIGHT_DATA_RECORDER_FORMAT = 1 }; 679 680 Trace T; 681 switch (Type) { 682 case NAIVE_FORMAT: 683 if (Version == 1 || Version == 2) { 684 if (auto E = loadNaiveFormatLog(Data, T.FileHeader, T.Records)) 685 return std::move(E); 686 } else { 687 return make_error<StringError>( 688 Twine("Unsupported version for Basic/Naive Mode logging: ") + 689 Twine(Version), 690 std::make_error_code(std::errc::executable_format_error)); 691 } 692 break; 693 case FLIGHT_DATA_RECORDER_FORMAT: 694 if (Version == 1 || Version == 2) { 695 if (auto E = loadFDRLog(Data, T.FileHeader, T.Records)) 696 return std::move(E); 697 } else { 698 return make_error<StringError>( 699 Twine("Unsupported version for FDR Mode logging: ") + Twine(Version), 700 std::make_error_code(std::errc::executable_format_error)); 701 } 702 break; 703 default: 704 if (auto E = loadYAMLLog(Data, T.FileHeader, T.Records)) 705 return std::move(E); 706 } 707 708 if (Sort) 709 std::stable_sort(T.Records.begin(), T.Records.end(), 710 [&](const XRayRecord &L, const XRayRecord &R) { 711 return L.TSC < R.TSC; 712 }); 713 714 return std::move(T); 715 } 716