• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- Trace.cpp - XRay Trace Loading implementation. ---------------------===//
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 // XRay log reader implementation.
10 //
11 //===----------------------------------------------------------------------===//
12 #include "llvm/XRay/Trace.h"
13 #include "llvm/ADT/STLExtras.h"
14 #include "llvm/Support/DataExtractor.h"
15 #include "llvm/Support/Error.h"
16 #include "llvm/Support/FileSystem.h"
17 #include "llvm/XRay/BlockIndexer.h"
18 #include "llvm/XRay/BlockVerifier.h"
19 #include "llvm/XRay/FDRRecordConsumer.h"
20 #include "llvm/XRay/FDRRecordProducer.h"
21 #include "llvm/XRay/FDRRecords.h"
22 #include "llvm/XRay/FDRTraceExpander.h"
23 #include "llvm/XRay/FileHeaderReader.h"
24 #include "llvm/XRay/YAMLXRayRecord.h"
25 #include <memory>
26 #include <vector>
27 
28 using namespace llvm;
29 using namespace llvm::xray;
30 using llvm::yaml::Input;
31 
32 namespace {
33 using XRayRecordStorage =
34     std::aligned_storage<sizeof(XRayRecord), alignof(XRayRecord)>::type;
35 
loadNaiveFormatLog(StringRef Data,bool IsLittleEndian,XRayFileHeader & FileHeader,std::vector<XRayRecord> & Records)36 Error loadNaiveFormatLog(StringRef Data, bool IsLittleEndian,
37                          XRayFileHeader &FileHeader,
38                          std::vector<XRayRecord> &Records) {
39   if (Data.size() < 32)
40     return make_error<StringError>(
41         "Not enough bytes for an XRay log.",
42         std::make_error_code(std::errc::invalid_argument));
43 
44   if (Data.size() - 32 == 0 || Data.size() % 32 != 0)
45     return make_error<StringError>(
46         "Invalid-sized XRay data.",
47         std::make_error_code(std::errc::invalid_argument));
48 
49   DataExtractor Reader(Data, IsLittleEndian, 8);
50   uint64_t OffsetPtr = 0;
51   auto FileHeaderOrError = readBinaryFormatHeader(Reader, OffsetPtr);
52   if (!FileHeaderOrError)
53     return FileHeaderOrError.takeError();
54   FileHeader = std::move(FileHeaderOrError.get());
55 
56   // Each record after the header will be 32 bytes, in the following format:
57   //
58   //   (2)   uint16 : record type
59   //   (1)   uint8  : cpu id
60   //   (1)   uint8  : type
61   //   (4)   sint32 : function id
62   //   (8)   uint64 : tsc
63   //   (4)   uint32 : thread id
64   //   (4)   uint32 : process id
65   //   (8)   -      : padding
66   while (Reader.isValidOffset(OffsetPtr)) {
67     if (!Reader.isValidOffsetForDataOfSize(OffsetPtr, 32))
68       return createStringError(
69           std::make_error_code(std::errc::executable_format_error),
70           "Not enough bytes to read a full record at offset %" PRId64 ".",
71           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 %" PRId64 ".", OffsetPtr);
78 
79     switch (RecordType) {
80     case 0: { // Normal records.
81       Records.emplace_back();
82       auto &Record = Records.back();
83       Record.RecordType = RecordType;
84 
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 %" PRId64 ".", OffsetPtr);
91 
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 %" PRId64 ".",
98             OffsetPtr);
99 
100       switch (Type) {
101       case 0:
102         Record.Type = RecordTypes::ENTER;
103         break;
104       case 1:
105         Record.Type = RecordTypes::EXIT;
106         break;
107       case 2:
108         Record.Type = RecordTypes::TAIL_EXIT;
109         break;
110       case 3:
111         Record.Type = RecordTypes::ENTER_ARG;
112         break;
113       default:
114         return createStringError(
115             std::make_error_code(std::errc::executable_format_error),
116             "Unknown record type '%d' at offset %" PRId64 ".", Type, OffsetPtr);
117       }
118 
119       PreReadOffset = OffsetPtr;
120       Record.FuncId = Reader.getSigned(&OffsetPtr, sizeof(int32_t));
121       if (OffsetPtr == PreReadOffset)
122         return createStringError(
123             std::make_error_code(std::errc::executable_format_error),
124             "Failed reading function id field at offset %" PRId64 ".",
125             OffsetPtr);
126 
127       PreReadOffset = OffsetPtr;
128       Record.TSC = Reader.getU64(&OffsetPtr);
129       if (OffsetPtr == PreReadOffset)
130         return createStringError(
131             std::make_error_code(std::errc::executable_format_error),
132             "Failed reading TSC field at offset %" PRId64 ".", OffsetPtr);
133 
134       PreReadOffset = OffsetPtr;
135       Record.TId = Reader.getU32(&OffsetPtr);
136       if (OffsetPtr == PreReadOffset)
137         return createStringError(
138             std::make_error_code(std::errc::executable_format_error),
139             "Failed reading thread id field at offset %" PRId64 ".", OffsetPtr);
140 
141       PreReadOffset = OffsetPtr;
142       Record.PId = Reader.getU32(&OffsetPtr);
143       if (OffsetPtr == PreReadOffset)
144         return createStringError(
145             std::make_error_code(std::errc::executable_format_error),
146             "Failed reading process id at offset %" PRId64 ".", OffsetPtr);
147 
148       break;
149     }
150     case 1: { // Arg payload record.
151       auto &Record = Records.back();
152 
153       // We skip the next two bytes of the record, because we don't need the
154       // type and the CPU record for arg payloads.
155       OffsetPtr += 2;
156       PreReadOffset = OffsetPtr;
157       int32_t FuncId = Reader.getSigned(&OffsetPtr, sizeof(int32_t));
158       if (OffsetPtr == PreReadOffset)
159         return createStringError(
160             std::make_error_code(std::errc::executable_format_error),
161             "Failed reading function id field at offset %" PRId64 ".",
162             OffsetPtr);
163 
164       PreReadOffset = OffsetPtr;
165       auto TId = Reader.getU32(&OffsetPtr);
166       if (OffsetPtr == PreReadOffset)
167         return createStringError(
168             std::make_error_code(std::errc::executable_format_error),
169             "Failed reading thread id field at offset %" PRId64 ".", OffsetPtr);
170 
171       PreReadOffset = OffsetPtr;
172       auto PId = Reader.getU32(&OffsetPtr);
173       if (OffsetPtr == PreReadOffset)
174         return createStringError(
175             std::make_error_code(std::errc::executable_format_error),
176             "Failed reading process id field at offset %" PRId64 ".",
177             OffsetPtr);
178 
179       // Make a check for versions above 3 for the Pid field
180       if (Record.FuncId != FuncId || Record.TId != TId ||
181           (FileHeader.Version >= 3 ? Record.PId != PId : false))
182         return createStringError(
183             std::make_error_code(std::errc::executable_format_error),
184             "Corrupted log, found arg payload following non-matching "
185             "function+thread record. Record for function %d != %d at offset "
186             "%" PRId64 ".",
187             Record.FuncId, FuncId, OffsetPtr);
188 
189       PreReadOffset = OffsetPtr;
190       auto Arg = Reader.getU64(&OffsetPtr);
191       if (OffsetPtr == PreReadOffset)
192         return createStringError(
193             std::make_error_code(std::errc::executable_format_error),
194             "Failed reading argument payload at offset %" PRId64 ".",
195             OffsetPtr);
196 
197       Record.CallArgs.push_back(Arg);
198       break;
199     }
200     default:
201       return createStringError(
202           std::make_error_code(std::errc::executable_format_error),
203           "Unknown record type '%d' at offset %" PRId64 ".", RecordType,
204           OffsetPtr);
205     }
206     // Advance the offset pointer enough bytes to align to 32-byte records for
207     // basic mode logs.
208     OffsetPtr += 8;
209   }
210   return Error::success();
211 }
212 
213 /// Reads a log in FDR mode for version 1 of this binary format. FDR mode is
214 /// defined as part of the compiler-rt project in xray_fdr_logging.h, and such
215 /// a log consists of the familiar 32 bit XRayHeader, followed by sequences of
216 /// of interspersed 16 byte Metadata Records and 8 byte Function Records.
217 ///
218 /// The following is an attempt to document the grammar of the format, which is
219 /// parsed by this function for little-endian machines. Since the format makes
220 /// use of BitFields, when we support big-endian architectures, we will need to
221 /// adjust not only the endianness parameter to llvm's RecordExtractor, but also
222 /// the bit twiddling logic, which is consistent with the little-endian
223 /// convention that BitFields within a struct will first be packed into the
224 /// least significant bits the address they belong to.
225 ///
226 /// We expect a format complying with the grammar in the following pseudo-EBNF
227 /// in Version 1 of the FDR log.
228 ///
229 /// FDRLog: XRayFileHeader ThreadBuffer*
230 /// XRayFileHeader: 32 bytes to identify the log as FDR with machine metadata.
231 ///     Includes BufferSize
232 /// ThreadBuffer: NewBuffer WallClockTime NewCPUId FunctionSequence EOB
233 /// BufSize: 8 byte unsigned integer indicating how large the buffer is.
234 /// NewBuffer: 16 byte metadata record with Thread Id.
235 /// WallClockTime: 16 byte metadata record with human readable time.
236 /// Pid: 16 byte metadata record with Pid
237 /// NewCPUId: 16 byte metadata record with CPUId and a 64 bit TSC reading.
238 /// EOB: 16 byte record in a thread buffer plus mem garbage to fill BufSize.
239 /// FunctionSequence: NewCPUId | TSCWrap | FunctionRecord
240 /// TSCWrap: 16 byte metadata record with a full 64 bit TSC reading.
241 /// FunctionRecord: 8 byte record with FunctionId, entry/exit, and TSC delta.
242 ///
243 /// In Version 2, we make the following changes:
244 ///
245 /// ThreadBuffer: BufferExtents NewBuffer WallClockTime NewCPUId
246 ///               FunctionSequence
247 /// BufferExtents: 16 byte metdata record describing how many usable bytes are
248 ///                in the buffer. This is measured from the start of the buffer
249 ///                and must always be at least 48 (bytes).
250 ///
251 /// In Version 3, we make the following changes:
252 ///
253 /// ThreadBuffer: BufferExtents NewBuffer WallClockTime Pid NewCPUId
254 ///               FunctionSequence
255 /// EOB: *deprecated*
256 ///
257 /// In Version 4, we make the following changes:
258 ///
259 /// CustomEventRecord now includes the CPU data.
260 ///
261 /// In Version 5, we make the following changes:
262 ///
263 /// CustomEventRecord and TypedEventRecord now use TSC delta encoding similar to
264 /// what FunctionRecord instances use, and we no longer need to include the CPU
265 /// id in the CustomEventRecord.
266 ///
loadFDRLog(StringRef Data,bool IsLittleEndian,XRayFileHeader & FileHeader,std::vector<XRayRecord> & Records)267 Error loadFDRLog(StringRef Data, bool IsLittleEndian,
268                  XRayFileHeader &FileHeader, std::vector<XRayRecord> &Records) {
269 
270   if (Data.size() < 32)
271     return createStringError(std::make_error_code(std::errc::invalid_argument),
272                              "Not enough bytes for an XRay FDR log.");
273   DataExtractor DE(Data, IsLittleEndian, 8);
274 
275   uint64_t OffsetPtr = 0;
276   auto FileHeaderOrError = readBinaryFormatHeader(DE, OffsetPtr);
277   if (!FileHeaderOrError)
278     return FileHeaderOrError.takeError();
279   FileHeader = std::move(FileHeaderOrError.get());
280 
281   // First we load the records into memory.
282   std::vector<std::unique_ptr<Record>> FDRRecords;
283 
284   {
285     FileBasedRecordProducer P(FileHeader, DE, OffsetPtr);
286     LogBuilderConsumer C(FDRRecords);
287     while (DE.isValidOffsetForDataOfSize(OffsetPtr, 1)) {
288       auto R = P.produce();
289       if (!R)
290         return R.takeError();
291       if (auto E = C.consume(std::move(R.get())))
292         return E;
293     }
294   }
295 
296   // Next we index the records into blocks.
297   BlockIndexer::Index Index;
298   {
299     BlockIndexer Indexer(Index);
300     for (auto &R : FDRRecords)
301       if (auto E = R->apply(Indexer))
302         return E;
303     if (auto E = Indexer.flush())
304       return E;
305   }
306 
307   // Then we verify the consistency of the blocks.
308   {
309     for (auto &PTB : Index) {
310       auto &Blocks = PTB.second;
311       for (auto &B : Blocks) {
312         BlockVerifier Verifier;
313         for (auto *R : B.Records)
314           if (auto E = R->apply(Verifier))
315             return E;
316         if (auto E = Verifier.verify())
317           return E;
318       }
319     }
320   }
321 
322   // This is now the meat of the algorithm. Here we sort the blocks according to
323   // the Walltime record in each of the blocks for the same thread. This allows
324   // us to more consistently recreate the execution trace in temporal order.
325   // After the sort, we then reconstitute `Trace` records using a stateful
326   // visitor associated with a single process+thread pair.
327   {
328     for (auto &PTB : Index) {
329       auto &Blocks = PTB.second;
330       llvm::sort(Blocks, [](const BlockIndexer::Block &L,
331                             const BlockIndexer::Block &R) {
332         return (L.WallclockTime->seconds() < R.WallclockTime->seconds() &&
333                 L.WallclockTime->nanos() < R.WallclockTime->nanos());
334       });
335       auto Adder = [&](const XRayRecord &R) { Records.push_back(R); };
336       TraceExpander Expander(Adder, FileHeader.Version);
337       for (auto &B : Blocks) {
338         for (auto *R : B.Records)
339           if (auto E = R->apply(Expander))
340             return E;
341       }
342       if (auto E = Expander.flush())
343         return E;
344     }
345   }
346 
347   return Error::success();
348 }
349 
loadYAMLLog(StringRef Data,XRayFileHeader & FileHeader,std::vector<XRayRecord> & Records)350 Error loadYAMLLog(StringRef Data, XRayFileHeader &FileHeader,
351                   std::vector<XRayRecord> &Records) {
352   YAMLXRayTrace Trace;
353   Input In(Data);
354   In >> Trace;
355   if (In.error())
356     return make_error<StringError>("Failed loading YAML Data.", In.error());
357 
358   FileHeader.Version = Trace.Header.Version;
359   FileHeader.Type = Trace.Header.Type;
360   FileHeader.ConstantTSC = Trace.Header.ConstantTSC;
361   FileHeader.NonstopTSC = Trace.Header.NonstopTSC;
362   FileHeader.CycleFrequency = Trace.Header.CycleFrequency;
363 
364   if (FileHeader.Version != 1)
365     return make_error<StringError>(
366         Twine("Unsupported XRay file version: ") + Twine(FileHeader.Version),
367         std::make_error_code(std::errc::invalid_argument));
368 
369   Records.clear();
370   std::transform(Trace.Records.begin(), Trace.Records.end(),
371                  std::back_inserter(Records), [&](const YAMLXRayRecord &R) {
372                    return XRayRecord{R.RecordType, R.CPU,      R.Type,
373                                      R.FuncId,     R.TSC,      R.TId,
374                                      R.PId,        R.CallArgs, R.Data};
375                  });
376   return Error::success();
377 }
378 } // namespace
379 
loadTraceFile(StringRef Filename,bool Sort)380 Expected<Trace> llvm::xray::loadTraceFile(StringRef Filename, bool Sort) {
381   Expected<sys::fs::file_t> FdOrErr = sys::fs::openNativeFileForRead(Filename);
382   if (!FdOrErr)
383     return FdOrErr.takeError();
384 
385   uint64_t FileSize;
386   if (auto EC = sys::fs::file_size(Filename, FileSize)) {
387     return make_error<StringError>(
388         Twine("Cannot read log from '") + Filename + "'", EC);
389   }
390   if (FileSize < 4) {
391     return make_error<StringError>(
392         Twine("File '") + Filename + "' too small for XRay.",
393         std::make_error_code(std::errc::executable_format_error));
394   }
395 
396   // Map the opened file into memory and use a StringRef to access it later.
397   std::error_code EC;
398   sys::fs::mapped_file_region MappedFile(
399       *FdOrErr, sys::fs::mapped_file_region::mapmode::readonly, FileSize, 0,
400       EC);
401   sys::fs::closeFile(*FdOrErr);
402   if (EC) {
403     return make_error<StringError>(
404         Twine("Cannot read log from '") + Filename + "'", EC);
405   }
406   auto Data = StringRef(MappedFile.data(), MappedFile.size());
407 
408   // TODO: Lift the endianness and implementation selection here.
409   DataExtractor LittleEndianDE(Data, true, 8);
410   auto TraceOrError = loadTrace(LittleEndianDE, Sort);
411   if (!TraceOrError) {
412     DataExtractor BigEndianDE(Data, false, 8);
413     consumeError(TraceOrError.takeError());
414     TraceOrError = loadTrace(BigEndianDE, Sort);
415   }
416   return TraceOrError;
417 }
418 
loadTrace(const DataExtractor & DE,bool Sort)419 Expected<Trace> llvm::xray::loadTrace(const DataExtractor &DE, bool Sort) {
420   // Attempt to detect the file type using file magic. We have a slight bias
421   // towards the binary format, and we do this by making sure that the first 4
422   // bytes of the binary file is some combination of the following byte
423   // patterns: (observe the code loading them assumes they're little endian)
424   //
425   //   0x01 0x00 0x00 0x00 - version 1, "naive" format
426   //   0x01 0x00 0x01 0x00 - version 1, "flight data recorder" format
427   //   0x02 0x00 0x01 0x00 - version 2, "flight data recorder" format
428   //
429   // YAML files don't typically have those first four bytes as valid text so we
430   // try loading assuming YAML if we don't find these bytes.
431   //
432   // Only if we can't load either the binary or the YAML format will we yield an
433   // error.
434   DataExtractor HeaderExtractor(DE.getData(), DE.isLittleEndian(), 8);
435   uint64_t OffsetPtr = 0;
436   uint16_t Version = HeaderExtractor.getU16(&OffsetPtr);
437   uint16_t Type = HeaderExtractor.getU16(&OffsetPtr);
438 
439   enum BinaryFormatType { NAIVE_FORMAT = 0, FLIGHT_DATA_RECORDER_FORMAT = 1 };
440 
441   Trace T;
442   switch (Type) {
443   case NAIVE_FORMAT:
444     if (Version == 1 || Version == 2 || Version == 3) {
445       if (auto E = loadNaiveFormatLog(DE.getData(), DE.isLittleEndian(),
446                                       T.FileHeader, T.Records))
447         return std::move(E);
448     } else {
449       return make_error<StringError>(
450           Twine("Unsupported version for Basic/Naive Mode logging: ") +
451               Twine(Version),
452           std::make_error_code(std::errc::executable_format_error));
453     }
454     break;
455   case FLIGHT_DATA_RECORDER_FORMAT:
456     if (Version >= 1 && Version <= 5) {
457       if (auto E = loadFDRLog(DE.getData(), DE.isLittleEndian(), T.FileHeader,
458                               T.Records))
459         return std::move(E);
460     } else {
461       return make_error<StringError>(
462           Twine("Unsupported version for FDR Mode logging: ") + Twine(Version),
463           std::make_error_code(std::errc::executable_format_error));
464     }
465     break;
466   default:
467     if (auto E = loadYAMLLog(DE.getData(), T.FileHeader, T.Records))
468       return std::move(E);
469   }
470 
471   if (Sort)
472     llvm::stable_sort(T.Records, [&](const XRayRecord &L, const XRayRecord &R) {
473       return L.TSC < R.TSC;
474     });
475 
476   return std::move(T);
477 }
478