1 //===- Trace.cpp - XRay Trace Loading implementation. ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // XRay log reader implementation.
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/BlockIndexer.h"
19 #include "llvm/XRay/BlockVerifier.h"
20 #include "llvm/XRay/FDRRecordConsumer.h"
21 #include "llvm/XRay/FDRRecordProducer.h"
22 #include "llvm/XRay/FDRRecords.h"
23 #include "llvm/XRay/FDRTraceExpander.h"
24 #include "llvm/XRay/FileHeaderReader.h"
25 #include "llvm/XRay/YAMLXRayRecord.h"
30 using namespace llvm::xray;
31 using llvm::yaml::Input;
34 using XRayRecordStorage =
35 std::aligned_storage<sizeof(XRayRecord), alignof(XRayRecord)>::type;
37 Error loadNaiveFormatLog(StringRef Data, bool IsLittleEndian,
38 XRayFileHeader &FileHeader,
39 std::vector<XRayRecord> &Records) {
41 return make_error<StringError>(
42 "Not enough bytes for an XRay log.",
43 std::make_error_code(std::errc::invalid_argument));
45 if (Data.size() - 32 == 0 || Data.size() % 32 != 0)
46 return make_error<StringError>(
47 "Invalid-sized XRay data.",
48 std::make_error_code(std::errc::invalid_argument));
50 DataExtractor Reader(Data, IsLittleEndian, 8);
51 uint32_t OffsetPtr = 0;
52 auto FileHeaderOrError = readBinaryFormatHeader(Reader, OffsetPtr);
53 if (!FileHeaderOrError)
54 return FileHeaderOrError.takeError();
55 FileHeader = std::move(FileHeaderOrError.get());
57 // Each record after the header will be 32 bytes, in the following format:
59 // (2) uint16 : record type
62 // (4) sint32 : function id
64 // (4) uint32 : thread id
65 // (4) uint32 : process id
67 while (Reader.isValidOffset(OffsetPtr)) {
68 if (!Reader.isValidOffsetForDataOfSize(OffsetPtr, 32))
69 return createStringError(
70 std::make_error_code(std::errc::executable_format_error),
71 "Not enough bytes to read a full record at offset %d.", OffsetPtr);
72 auto PreReadOffset = OffsetPtr;
73 auto RecordType = Reader.getU16(&OffsetPtr);
74 if (OffsetPtr == PreReadOffset)
75 return createStringError(
76 std::make_error_code(std::errc::executable_format_error),
77 "Failed reading record type at offset %d.", OffsetPtr);
80 case 0: { // Normal records.
81 Records.emplace_back();
82 auto &Record = Records.back();
83 Record.RecordType = RecordType;
85 PreReadOffset = OffsetPtr;
86 Record.CPU = Reader.getU8(&OffsetPtr);
87 if (OffsetPtr == PreReadOffset)
88 return createStringError(
89 std::make_error_code(std::errc::executable_format_error),
90 "Failed reading CPU field at offset %d.", OffsetPtr);
92 PreReadOffset = OffsetPtr;
93 auto Type = Reader.getU8(&OffsetPtr);
94 if (OffsetPtr == PreReadOffset)
95 return createStringError(
96 std::make_error_code(std::errc::executable_format_error),
97 "Failed reading record type field at offset %d.", OffsetPtr);
101 Record.Type = RecordTypes::ENTER;
104 Record.Type = RecordTypes::EXIT;
107 Record.Type = RecordTypes::TAIL_EXIT;
110 Record.Type = RecordTypes::ENTER_ARG;
113 return createStringError(
114 std::make_error_code(std::errc::executable_format_error),
115 "Unknown record type '%d' at offset %d.", Type, OffsetPtr);
118 PreReadOffset = OffsetPtr;
119 Record.FuncId = Reader.getSigned(&OffsetPtr, sizeof(int32_t));
120 if (OffsetPtr == PreReadOffset)
121 return createStringError(
122 std::make_error_code(std::errc::executable_format_error),
123 "Failed reading function id field at offset %d.", OffsetPtr);
125 PreReadOffset = OffsetPtr;
126 Record.TSC = Reader.getU64(&OffsetPtr);
127 if (OffsetPtr == PreReadOffset)
128 return createStringError(
129 std::make_error_code(std::errc::executable_format_error),
130 "Failed reading TSC field at offset %d.", OffsetPtr);
132 PreReadOffset = OffsetPtr;
133 Record.TId = Reader.getU32(&OffsetPtr);
134 if (OffsetPtr == PreReadOffset)
135 return createStringError(
136 std::make_error_code(std::errc::executable_format_error),
137 "Failed reading thread id field at offset %d.", OffsetPtr);
139 PreReadOffset = OffsetPtr;
140 Record.PId = Reader.getU32(&OffsetPtr);
141 if (OffsetPtr == PreReadOffset)
142 return createStringError(
143 std::make_error_code(std::errc::executable_format_error),
144 "Failed reading process id at offset %d.", OffsetPtr);
148 case 1: { // Arg payload record.
149 auto &Record = Records.back();
151 // We skip the next two bytes of the record, because we don't need the
152 // type and the CPU record for arg payloads.
154 PreReadOffset = OffsetPtr;
155 int32_t FuncId = Reader.getSigned(&OffsetPtr, sizeof(int32_t));
156 if (OffsetPtr == PreReadOffset)
157 return createStringError(
158 std::make_error_code(std::errc::executable_format_error),
159 "Failed reading function id field at offset %d.", OffsetPtr);
161 PreReadOffset = OffsetPtr;
162 auto TId = Reader.getU32(&OffsetPtr);
163 if (OffsetPtr == PreReadOffset)
164 return createStringError(
165 std::make_error_code(std::errc::executable_format_error),
166 "Failed reading thread id field at offset %d.", OffsetPtr);
168 PreReadOffset = OffsetPtr;
169 auto PId = Reader.getU32(&OffsetPtr);
170 if (OffsetPtr == PreReadOffset)
171 return createStringError(
172 std::make_error_code(std::errc::executable_format_error),
173 "Failed reading process id field at offset %d.", OffsetPtr);
175 // Make a check for versions above 3 for the Pid field
176 if (Record.FuncId != FuncId || Record.TId != TId ||
177 (FileHeader.Version >= 3 ? Record.PId != PId : false))
178 return createStringError(
179 std::make_error_code(std::errc::executable_format_error),
180 "Corrupted log, found arg payload following non-matching "
181 "function+thread record. Record for function %d != %d at offset "
183 Record.FuncId, FuncId, OffsetPtr);
185 PreReadOffset = OffsetPtr;
186 auto Arg = Reader.getU64(&OffsetPtr);
187 if (OffsetPtr == PreReadOffset)
188 return createStringError(
189 std::make_error_code(std::errc::executable_format_error),
190 "Failed reading argument payload at offset %d.", OffsetPtr);
192 Record.CallArgs.push_back(Arg);
196 return createStringError(
197 std::make_error_code(std::errc::executable_format_error),
198 "Unknown record type '%d' at offset %d.", RecordType, OffsetPtr);
200 // Advance the offset pointer enough bytes to align to 32-byte records for
204 return Error::success();
207 /// Reads a log in FDR mode for version 1 of this binary format. FDR mode is
208 /// defined as part of the compiler-rt project in xray_fdr_logging.h, and such
209 /// a log consists of the familiar 32 bit XRayHeader, followed by sequences of
210 /// of interspersed 16 byte Metadata Records and 8 byte Function Records.
212 /// The following is an attempt to document the grammar of the format, which is
213 /// parsed by this function for little-endian machines. Since the format makes
214 /// use of BitFields, when we support big-endian architectures, we will need to
215 /// adjust not only the endianness parameter to llvm's RecordExtractor, but also
216 /// the bit twiddling logic, which is consistent with the little-endian
217 /// convention that BitFields within a struct will first be packed into the
218 /// least significant bits the address they belong to.
220 /// We expect a format complying with the grammar in the following pseudo-EBNF
221 /// in Version 1 of the FDR log.
223 /// FDRLog: XRayFileHeader ThreadBuffer*
224 /// XRayFileHeader: 32 bytes to identify the log as FDR with machine metadata.
225 /// Includes BufferSize
226 /// ThreadBuffer: NewBuffer WallClockTime NewCPUId FunctionSequence EOB
227 /// BufSize: 8 byte unsigned integer indicating how large the buffer is.
228 /// NewBuffer: 16 byte metadata record with Thread Id.
229 /// WallClockTime: 16 byte metadata record with human readable time.
230 /// Pid: 16 byte metadata record with Pid
231 /// NewCPUId: 16 byte metadata record with CPUId and a 64 bit TSC reading.
232 /// EOB: 16 byte record in a thread buffer plus mem garbage to fill BufSize.
233 /// FunctionSequence: NewCPUId | TSCWrap | FunctionRecord
234 /// TSCWrap: 16 byte metadata record with a full 64 bit TSC reading.
235 /// FunctionRecord: 8 byte record with FunctionId, entry/exit, and TSC delta.
237 /// In Version 2, we make the following changes:
239 /// ThreadBuffer: BufferExtents NewBuffer WallClockTime NewCPUId
241 /// BufferExtents: 16 byte metdata record describing how many usable bytes are
242 /// in the buffer. This is measured from the start of the buffer
243 /// and must always be at least 48 (bytes).
245 /// In Version 3, we make the following changes:
247 /// ThreadBuffer: BufferExtents NewBuffer WallClockTime Pid NewCPUId
249 /// EOB: *deprecated*
251 /// In Version 4, we make the following changes:
253 /// CustomEventRecord now includes the CPU data.
255 /// In Version 5, we make the following changes:
257 /// CustomEventRecord and TypedEventRecord now use TSC delta encoding similar to
258 /// what FunctionRecord instances use, and we no longer need to include the CPU
259 /// id in the CustomEventRecord.
261 Error loadFDRLog(StringRef Data, bool IsLittleEndian,
262 XRayFileHeader &FileHeader, std::vector<XRayRecord> &Records) {
264 if (Data.size() < 32)
265 return createStringError(std::make_error_code(std::errc::invalid_argument),
266 "Not enough bytes for an XRay FDR log.");
267 DataExtractor DE(Data, IsLittleEndian, 8);
269 uint32_t OffsetPtr = 0;
270 auto FileHeaderOrError = readBinaryFormatHeader(DE, OffsetPtr);
271 if (!FileHeaderOrError)
272 return FileHeaderOrError.takeError();
273 FileHeader = std::move(FileHeaderOrError.get());
275 // First we load the records into memory.
276 std::vector<std::unique_ptr<Record>> FDRRecords;
279 FileBasedRecordProducer P(FileHeader, DE, OffsetPtr);
280 LogBuilderConsumer C(FDRRecords);
281 while (DE.isValidOffsetForDataOfSize(OffsetPtr, 1)) {
282 auto R = P.produce();
284 return R.takeError();
285 if (auto E = C.consume(std::move(R.get())))
290 // Next we index the records into blocks.
291 BlockIndexer::Index Index;
293 BlockIndexer Indexer(Index);
294 for (auto &R : FDRRecords)
295 if (auto E = R->apply(Indexer))
297 if (auto E = Indexer.flush())
301 // Then we verify the consistency of the blocks.
303 for (auto &PTB : Index) {
304 auto &Blocks = PTB.second;
305 for (auto &B : Blocks) {
306 BlockVerifier Verifier;
307 for (auto *R : B.Records)
308 if (auto E = R->apply(Verifier))
310 if (auto E = Verifier.verify())
316 // This is now the meat of the algorithm. Here we sort the blocks according to
317 // the Walltime record in each of the blocks for the same thread. This allows
318 // us to more consistently recreate the execution trace in temporal order.
319 // After the sort, we then reconstitute `Trace` records using a stateful
320 // visitor associated with a single process+thread pair.
322 for (auto &PTB : Index) {
323 auto &Blocks = PTB.second;
324 llvm::sort(Blocks, [](const BlockIndexer::Block &L,
325 const BlockIndexer::Block &R) {
326 return (L.WallclockTime->seconds() < R.WallclockTime->seconds() &&
327 L.WallclockTime->nanos() < R.WallclockTime->nanos());
329 auto Adder = [&](const XRayRecord &R) { Records.push_back(R); };
330 TraceExpander Expander(Adder, FileHeader.Version);
331 for (auto &B : Blocks) {
332 for (auto *R : B.Records)
333 if (auto E = R->apply(Expander))
336 if (auto E = Expander.flush())
341 return Error::success();
344 Error loadYAMLLog(StringRef Data, XRayFileHeader &FileHeader,
345 std::vector<XRayRecord> &Records) {
350 return make_error<StringError>("Failed loading YAML Data.", In.error());
352 FileHeader.Version = Trace.Header.Version;
353 FileHeader.Type = Trace.Header.Type;
354 FileHeader.ConstantTSC = Trace.Header.ConstantTSC;
355 FileHeader.NonstopTSC = Trace.Header.NonstopTSC;
356 FileHeader.CycleFrequency = Trace.Header.CycleFrequency;
358 if (FileHeader.Version != 1)
359 return make_error<StringError>(
360 Twine("Unsupported XRay file version: ") + Twine(FileHeader.Version),
361 std::make_error_code(std::errc::invalid_argument));
364 std::transform(Trace.Records.begin(), Trace.Records.end(),
365 std::back_inserter(Records), [&](const YAMLXRayRecord &R) {
366 return XRayRecord{R.RecordType, R.CPU, R.Type,
367 R.FuncId, R.TSC, R.TId,
368 R.PId, R.CallArgs, R.Data};
370 return Error::success();
374 Expected<Trace> llvm::xray::loadTraceFile(StringRef Filename, bool Sort) {
376 if (auto EC = sys::fs::openFileForRead(Filename, Fd)) {
377 return make_error<StringError>(
378 Twine("Cannot read log from '") + Filename + "'", EC);
382 if (auto EC = sys::fs::file_size(Filename, FileSize)) {
383 return make_error<StringError>(
384 Twine("Cannot read log from '") + Filename + "'", EC);
387 return make_error<StringError>(
388 Twine("File '") + Filename + "' too small for XRay.",
389 std::make_error_code(std::errc::executable_format_error));
392 // Map the opened file into memory and use a StringRef to access it later.
394 sys::fs::mapped_file_region MappedFile(
395 Fd, sys::fs::mapped_file_region::mapmode::readonly, FileSize, 0, EC);
397 return make_error<StringError>(
398 Twine("Cannot read log from '") + Filename + "'", EC);
400 auto Data = StringRef(MappedFile.data(), MappedFile.size());
402 // TODO: Lift the endianness and implementation selection here.
403 DataExtractor LittleEndianDE(Data, true, 8);
404 auto TraceOrError = loadTrace(LittleEndianDE, Sort);
406 DataExtractor BigEndianDE(Data, false, 8);
407 TraceOrError = loadTrace(BigEndianDE, Sort);
412 Expected<Trace> llvm::xray::loadTrace(const DataExtractor &DE, bool Sort) {
413 // Attempt to detect the file type using file magic. We have a slight bias
414 // towards the binary format, and we do this by making sure that the first 4
415 // bytes of the binary file is some combination of the following byte
416 // patterns: (observe the code loading them assumes they're little endian)
418 // 0x01 0x00 0x00 0x00 - version 1, "naive" format
419 // 0x01 0x00 0x01 0x00 - version 1, "flight data recorder" format
420 // 0x02 0x00 0x01 0x00 - version 2, "flight data recorder" format
422 // YAML files don't typically have those first four bytes as valid text so we
423 // try loading assuming YAML if we don't find these bytes.
425 // Only if we can't load either the binary or the YAML format will we yield an
427 DataExtractor HeaderExtractor(DE.getData(), DE.isLittleEndian(), 8);
428 uint32_t OffsetPtr = 0;
429 uint16_t Version = HeaderExtractor.getU16(&OffsetPtr);
430 uint16_t Type = HeaderExtractor.getU16(&OffsetPtr);
432 enum BinaryFormatType { NAIVE_FORMAT = 0, FLIGHT_DATA_RECORDER_FORMAT = 1 };
437 if (Version == 1 || Version == 2 || Version == 3) {
438 if (auto E = loadNaiveFormatLog(DE.getData(), DE.isLittleEndian(),
439 T.FileHeader, T.Records))
442 return make_error<StringError>(
443 Twine("Unsupported version for Basic/Naive Mode logging: ") +
445 std::make_error_code(std::errc::executable_format_error));
448 case FLIGHT_DATA_RECORDER_FORMAT:
449 if (Version >= 1 && Version <= 5) {
450 if (auto E = loadFDRLog(DE.getData(), DE.isLittleEndian(), T.FileHeader,
454 return make_error<StringError>(
455 Twine("Unsupported version for FDR Mode logging: ") + Twine(Version),
456 std::make_error_code(std::errc::executable_format_error));
460 if (auto E = loadYAMLLog(DE.getData(), T.FileHeader, T.Records))
465 std::stable_sort(T.Records.begin(), T.Records.end(),
466 [&](const XRayRecord &L, const XRayRecord &R) {
467 return L.TSC < R.TSC;