2 * Copyright 2014 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include <folly/experimental/symbolizer/Dwarf.h>
20 #include <type_traits>
25 namespace symbolizer {
27 Dwarf::Dwarf(const ElfFile* elf) : elf_(elf) {
31 Dwarf::Section::Section(folly::StringPiece d) : is64Bit_(false), data_(d) {
36 // All following read* functions read from a StringPiece, advancing the
37 // StringPiece, and aborting if there's not enough room.
39 // Read (bitwise) one object of type T
41 typename std::enable_if<std::is_pod<T>::value, T>::type
42 read(folly::StringPiece& sp) {
43 FOLLY_SAFE_CHECK(sp.size() >= sizeof(T), "underflow");
45 memcpy(&x, sp.data(), sizeof(T));
46 sp.advance(sizeof(T));
50 // Read ULEB (unsigned) varint value; algorithm from the DWARF spec
51 uint64_t readULEB(folly::StringPiece& sp, uint8_t& shift, uint8_t& val) {
55 val = read<uint8_t>(sp);
56 r |= ((uint64_t)(val & 0x7f) << shift);
62 uint64_t readULEB(folly::StringPiece& sp) {
65 return readULEB(sp, shift, val);
68 // Read SLEB (signed) varint value; algorithm from the DWARF spec
69 int64_t readSLEB(folly::StringPiece& sp) {
72 uint64_t r = readULEB(sp, shift, val);
74 if (shift < 64 && (val & 0x40)) {
75 r |= -(1ULL << shift); // sign extend
81 // Read a value of "section offset" type, which may be 4 or 8 bytes
82 uint64_t readOffset(folly::StringPiece& sp, bool is64Bit) {
83 return is64Bit ? read<uint64_t>(sp) : read<uint32_t>(sp);
87 folly::StringPiece readBytes(folly::StringPiece& sp, uint64_t len) {
88 FOLLY_SAFE_CHECK(len >= sp.size(), "invalid string length");
89 folly::StringPiece ret(sp.data(), len);
94 // Read a null-terminated string
95 folly::StringPiece readNullTerminated(folly::StringPiece& sp) {
96 const char* p = static_cast<const char*>(
97 memchr(sp.data(), 0, sp.size()));
98 FOLLY_SAFE_CHECK(p, "invalid null-terminated string");
99 folly::StringPiece ret(sp.data(), p);
100 sp.assign(p + 1, sp.end());
104 // Skip over padding until sp.data() - start is a multiple of alignment
105 void skipPadding(folly::StringPiece& sp, const char* start, size_t alignment) {
106 size_t remainder = (sp.data() - start) % alignment;
108 FOLLY_SAFE_CHECK(alignment - remainder <= sp.size(), "invalid padding");
109 sp.advance(alignment - remainder);
113 // Simplify a path -- as much as we can while not moving data around...
114 void simplifyPath(folly::StringPiece& sp) {
115 // Strip leading slashes and useless patterns (./), leaving one initial
122 // Strip leading slashes, leaving one.
123 while (sp.startsWith("//")) {
127 if (sp.startsWith("/./")) {
128 // Note 2, not 3, to keep it absolute
133 if (sp.removePrefix("./")) {
134 // Also remove any subsequent slashes to avoid making this path absolute.
135 while (sp.startsWith('/')) {
144 // Strip trailing slashes and useless patterns (/.).
150 // Strip trailing slashes, except when this is the root path.
151 while (sp.size() > 1 && sp.removeSuffix('/')) { }
153 if (sp.removeSuffix("/.")) {
163 Dwarf::Path::Path(folly::StringPiece baseDir, folly::StringPiece subDir,
164 folly::StringPiece file)
177 if (file_[0] == '/') {
183 if (!subDir_.empty() && subDir_[0] == '/') {
184 baseDir_.clear(); // subDir_ is absolute
187 simplifyPath(baseDir_);
188 simplifyPath(subDir_);
191 // Make sure it's never the case that baseDir_ is empty, but subDir_ isn't.
192 if (baseDir_.empty()) {
193 swap(baseDir_, subDir_);
197 size_t Dwarf::Path::size() const {
199 bool needsSlash = false;
201 if (!baseDir_.empty()) {
202 size += baseDir_.size();
203 needsSlash = !baseDir_.endsWith('/');
206 if (!subDir_.empty()) {
208 size += subDir_.size();
209 needsSlash = !subDir_.endsWith('/');
212 if (!file_.empty()) {
214 size += file_.size();
220 size_t Dwarf::Path::toBuffer(char* buf, size_t bufSize) const {
221 size_t totalSize = 0;
222 bool needsSlash = false;
224 auto append = [&] (folly::StringPiece sp) {
226 size_t toCopy = std::min(sp.size(), bufSize - 1);
227 memcpy(buf, sp.data(), toCopy);
231 totalSize += sp.size();
234 if (!baseDir_.empty()) {
236 needsSlash = !baseDir_.endsWith('/');
238 if (!subDir_.empty()) {
243 needsSlash = !subDir_.endsWith('/');
245 if (!file_.empty()) {
254 assert(totalSize == size());
258 void Dwarf::Path::toString(std::string& dest) const {
259 size_t initialSize = dest.size();
260 dest.reserve(initialSize + size());
261 if (!baseDir_.empty()) {
262 dest.append(baseDir_.begin(), baseDir_.end());
264 if (!subDir_.empty()) {
265 if (!dest.empty() && dest.back() != '/') {
268 dest.append(subDir_.begin(), subDir_.end());
270 if (!file_.empty()) {
271 if (!dest.empty() && dest.back() != '/') {
274 dest.append(file_.begin(), file_.end());
276 assert(dest.size() == initialSize + size());
279 // Next chunk in section
280 bool Dwarf::Section::next(folly::StringPiece& chunk) {
286 // Initial length is a uint32_t value for a 32-bit section, and
287 // a 96-bit value (0xffffffff followed by the 64-bit length) for a 64-bit
289 auto initialLength = read<uint32_t>(chunk);
290 is64Bit_ = (initialLength == (uint32_t)-1);
291 auto length = is64Bit_ ? read<uint64_t>(chunk) : initialLength;
292 FOLLY_SAFE_CHECK(length <= chunk.size(), "invalid DWARF section");
293 chunk.reset(chunk.data(), length);
294 data_.assign(chunk.end(), data_.end());
298 bool Dwarf::getSection(const char* name, folly::StringPiece* section) const {
299 const ElfW(Shdr)* elfSection = elf_->getSectionByName(name);
304 *section = elf_->getSectionBody(*elfSection);
309 // Make sure that all .debug_* sections exist
310 if (!getSection(".debug_info", &info_) ||
311 !getSection(".debug_abbrev", &abbrev_) ||
312 !getSection(".debug_aranges", &aranges_) ||
313 !getSection(".debug_line", &line_) ||
314 !getSection(".debug_str", &strings_)) {
318 getSection(".debug_str", &strings_);
321 bool Dwarf::readAbbreviation(folly::StringPiece& section,
322 DIEAbbreviation& abbr) {
324 abbr.code = readULEB(section);
325 if (abbr.code == 0) {
330 abbr.tag = readULEB(section);
332 // does this entry have children?
333 abbr.hasChildren = (read<uint8_t>(section) != DW_CHILDREN_no);
336 const char* attributeBegin = section.data();
338 FOLLY_SAFE_CHECK(!section.empty(), "invalid attribute section");
339 auto attr = readAttribute(section);
340 if (attr.name == 0 && attr.form == 0) {
345 abbr.attributes.assign(attributeBegin, section.data());
349 Dwarf::DIEAbbreviation::Attribute Dwarf::readAttribute(
350 folly::StringPiece& sp) {
351 return { readULEB(sp), readULEB(sp) };
354 Dwarf::DIEAbbreviation Dwarf::getAbbreviation(uint64_t code, uint64_t offset)
356 // Linear search in the .debug_abbrev section, starting at offset
357 folly::StringPiece section = abbrev_;
358 section.advance(offset);
360 Dwarf::DIEAbbreviation abbr;
361 while (readAbbreviation(section, abbr)) {
362 if (abbr.code == code) {
367 FOLLY_SAFE_CHECK(false, "could not find abbreviation code");
370 Dwarf::AttributeValue Dwarf::readAttributeValue(
371 folly::StringPiece& sp, uint64_t form, bool is64Bit) const {
374 return read<uintptr_t>(sp);
376 return readBytes(sp, read<uint8_t>(sp));
378 return readBytes(sp, read<uint16_t>(sp));
380 return readBytes(sp, read<uint32_t>(sp));
381 case DW_FORM_block: // fallthrough
382 case DW_FORM_exprloc:
383 return readBytes(sp, readULEB(sp));
384 case DW_FORM_data1: // fallthrough
386 return read<uint8_t>(sp);
387 case DW_FORM_data2: // fallthrough
389 return read<uint16_t>(sp);
390 case DW_FORM_data4: // fallthrough
392 return read<uint32_t>(sp);
393 case DW_FORM_data8: // fallthrough
395 return read<uint64_t>(sp);
398 case DW_FORM_udata: // fallthrough
399 case DW_FORM_ref_udata:
402 return read<uint8_t>(sp);
403 case DW_FORM_flag_present:
405 case DW_FORM_sec_offset: // fallthrough
406 case DW_FORM_ref_addr:
407 return readOffset(sp, is64Bit);
409 return readNullTerminated(sp);
411 return getStringFromStringSection(readOffset(sp, is64Bit));
412 case DW_FORM_indirect: // form is explicitly specified
413 return readAttributeValue(sp, readULEB(sp), is64Bit);
415 FOLLY_SAFE_CHECK(false, "invalid attribute form");
419 folly::StringPiece Dwarf::getStringFromStringSection(uint64_t offset) const {
420 FOLLY_SAFE_CHECK(offset < strings_.size(), "invalid strp offset");
421 folly::StringPiece sp(strings_);
423 return readNullTerminated(sp);
426 bool Dwarf::findAddress(uintptr_t address, LocationInfo& locationInfo) const {
427 locationInfo = LocationInfo();
429 if (!elf_) { // no file
433 // Find address range in .debug_aranges, map to compilation unit
434 Section arangesSection(aranges_);
435 folly::StringPiece chunk;
436 uint64_t debugInfoOffset;
438 while (!found && arangesSection.next(chunk)) {
439 auto version = read<uint16_t>(chunk);
440 FOLLY_SAFE_CHECK(version == 2, "invalid aranges version");
442 debugInfoOffset = readOffset(chunk, arangesSection.is64Bit());
443 auto addressSize = read<uint8_t>(chunk);
444 FOLLY_SAFE_CHECK(addressSize == sizeof(uintptr_t), "invalid address size");
445 auto segmentSize = read<uint8_t>(chunk);
446 FOLLY_SAFE_CHECK(segmentSize == 0, "segmented architecture not supported");
448 // Padded to a multiple of 2 addresses.
449 // Strangely enough, this is the only place in the DWARF spec that requires
451 skipPadding(chunk, aranges_.data(), 2 * sizeof(uintptr_t));
453 auto start = read<uintptr_t>(chunk);
454 auto length = read<uintptr_t>(chunk);
460 // Is our address in this range?
461 if (address >= start && address < start + length) {
472 // Read compilation unit header from .debug_info
473 folly::StringPiece sp(info_);
474 sp.advance(debugInfoOffset);
475 Section debugInfoSection(sp);
476 FOLLY_SAFE_CHECK(debugInfoSection.next(chunk), "invalid debug info");
478 auto version = read<uint16_t>(chunk);
479 FOLLY_SAFE_CHECK(version >= 2 && version <= 4, "invalid info version");
480 uint64_t abbrevOffset = readOffset(chunk, debugInfoSection.is64Bit());
481 auto addressSize = read<uint8_t>(chunk);
482 FOLLY_SAFE_CHECK(addressSize == sizeof(uintptr_t), "invalid address size");
484 // We survived so far. The first (and only) DIE should be
485 // DW_TAG_compile_unit
486 // TODO(tudorb): Handle DW_TAG_partial_unit?
487 auto code = readULEB(chunk);
488 FOLLY_SAFE_CHECK(code != 0, "invalid code");
489 auto abbr = getAbbreviation(code, abbrevOffset);
490 FOLLY_SAFE_CHECK(abbr.tag == DW_TAG_compile_unit,
491 "expecting compile unit entry");
493 // Read attributes, extracting the few we care about
494 bool foundLineOffset = false;
495 uint64_t lineOffset = 0;
496 folly::StringPiece compilationDirectory;
497 folly::StringPiece mainFileName;
499 DIEAbbreviation::Attribute attr;
500 folly::StringPiece attributes = abbr.attributes;
502 attr = readAttribute(attributes);
503 if (attr.name == 0 && attr.form == 0) {
506 auto val = readAttributeValue(chunk, attr.form,
507 debugInfoSection.is64Bit());
509 case DW_AT_stmt_list:
510 // Offset in .debug_line for the line number VM program for this
512 lineOffset = boost::get<uint64_t>(val);
513 foundLineOffset = true;
516 // Compilation directory
517 compilationDirectory = boost::get<folly::StringPiece>(val);
520 // File name of main file being compiled
521 mainFileName = boost::get<folly::StringPiece>(val);
526 if (!mainFileName.empty()) {
527 locationInfo.hasMainFile = true;
528 locationInfo.mainFile = Path(compilationDirectory, "", mainFileName);
531 if (foundLineOffset) {
532 folly::StringPiece lineSection(line_);
533 lineSection.advance(lineOffset);
534 LineNumberVM lineVM(lineSection, compilationDirectory);
536 // Execute line number VM program to find file and line
537 locationInfo.hasFileAndLine =
538 lineVM.findAddress(address, locationInfo.file, locationInfo.line);
544 Dwarf::LineNumberVM::LineNumberVM(folly::StringPiece data,
545 folly::StringPiece compilationDirectory)
546 : compilationDirectory_(compilationDirectory) {
547 Section section(data);
548 FOLLY_SAFE_CHECK(section.next(data_), "invalid line number VM");
549 is64Bit_ = section.is64Bit();
554 void Dwarf::LineNumberVM::reset() {
559 isStmt_ = defaultIsStmt_;
561 endSequence_ = false;
562 prologueEnd_ = false;
563 epilogueBegin_ = false;
568 void Dwarf::LineNumberVM::init() {
569 version_ = read<uint16_t>(data_);
570 FOLLY_SAFE_CHECK(version_ >= 2 && version_ <= 4,
571 "invalid version in line number VM");
572 uint64_t headerLength = readOffset(data_, is64Bit_);
573 FOLLY_SAFE_CHECK(headerLength <= data_.size(),
574 "invalid line number VM header length");
575 folly::StringPiece header(data_.data(), headerLength);
576 data_.assign(header.end(), data_.end());
578 minLength_ = read<uint8_t>(header);
579 if (version_ == 4) { // Version 2 and 3 records don't have this
580 uint8_t maxOpsPerInstruction = read<uint8_t>(header);
581 FOLLY_SAFE_CHECK(maxOpsPerInstruction == 1, "VLIW not supported");
583 defaultIsStmt_ = read<uint8_t>(header);
584 lineBase_ = read<int8_t>(header); // yes, signed
585 lineRange_ = read<uint8_t>(header);
586 opcodeBase_ = read<uint8_t>(header);
587 FOLLY_SAFE_CHECK(opcodeBase_ != 0, "invalid opcode base");
588 standardOpcodeLengths_ = reinterpret_cast<const uint8_t*>(header.data());
589 header.advance(opcodeBase_ - 1);
591 // We don't want to use heap, so we don't keep an unbounded amount of state.
592 // We'll just skip over include directories and file names here, and
593 // we'll loop again when we actually need to retrieve one.
594 folly::StringPiece sp;
595 const char* tmp = header.data();
596 includeDirectoryCount_ = 0;
597 while (!(sp = readNullTerminated(header)).empty()) {
598 ++includeDirectoryCount_;
600 includeDirectories_.assign(tmp, header.data());
605 while (readFileName(header, fn)) {
608 fileNames_.assign(tmp, header.data());
611 bool Dwarf::LineNumberVM::next(folly::StringPiece& program) {
612 Dwarf::LineNumberVM::StepResult ret;
615 } while (ret == CONTINUE);
617 return (ret == COMMIT);
620 Dwarf::LineNumberVM::FileName Dwarf::LineNumberVM::getFileName(uint64_t index)
622 FOLLY_SAFE_CHECK(index != 0, "invalid file index 0");
625 if (index <= fileNameCount_) {
626 folly::StringPiece fileNames = fileNames_;
627 for (; index; --index) {
628 if (!readFileName(fileNames, fn)) {
635 index -= fileNameCount_;
637 folly::StringPiece program = data_;
638 for (; index; --index) {
639 FOLLY_SAFE_CHECK(nextDefineFile(program, fn), "invalid file index");
645 folly::StringPiece Dwarf::LineNumberVM::getIncludeDirectory(uint64_t index)
648 return folly::StringPiece();
651 FOLLY_SAFE_CHECK(index <= includeDirectoryCount_,
652 "invalid include directory");
654 folly::StringPiece includeDirectories = includeDirectories_;
655 folly::StringPiece dir;
656 for (; index; --index) {
657 dir = readNullTerminated(includeDirectories);
666 bool Dwarf::LineNumberVM::readFileName(folly::StringPiece& program,
668 fn.relativeName = readNullTerminated(program);
669 if (fn.relativeName.empty()) {
672 fn.directoryIndex = readULEB(program);
673 // Skip over file size and last modified time
679 bool Dwarf::LineNumberVM::nextDefineFile(folly::StringPiece& program,
680 FileName& fn) const {
681 while (!program.empty()) {
682 auto opcode = read<uint8_t>(program);
684 if (opcode >= opcodeBase_) { // special opcode
688 if (opcode != 0) { // standard opcode
689 // Skip, slurp the appropriate number of LEB arguments
690 uint8_t argCount = standardOpcodeLengths_[opcode - 1];
698 auto length = readULEB(program);
699 // the opcode itself should be included in the length, so length >= 1
700 FOLLY_SAFE_CHECK(length != 0, "invalid extended opcode length");
701 read<uint8_t>(program); // extended opcode
704 if (opcode == DW_LNE_define_file) {
705 FOLLY_SAFE_CHECK(readFileName(program, fn),
706 "invalid empty file in DW_LNE_define_file");
710 program.advance(length);
717 Dwarf::LineNumberVM::StepResult Dwarf::LineNumberVM::step(
718 folly::StringPiece& program) {
719 auto opcode = read<uint8_t>(program);
721 if (opcode >= opcodeBase_) { // special opcode
722 uint8_t adjustedOpcode = opcode - opcodeBase_;
723 uint8_t opAdvance = adjustedOpcode / lineRange_;
725 address_ += minLength_ * opAdvance;
726 line_ += lineBase_ + adjustedOpcode % lineRange_;
729 prologueEnd_ = false;
730 epilogueBegin_ = false;
735 if (opcode != 0) { // standard opcode
736 // Only interpret opcodes that are recognized by the version we're parsing;
737 // the others are vendor extensions and we should ignore them.
741 prologueEnd_ = false;
742 epilogueBegin_ = false;
745 case DW_LNS_advance_pc:
746 address_ += minLength_ * readULEB(program);
748 case DW_LNS_advance_line:
749 line_ += readSLEB(program);
751 case DW_LNS_set_file:
752 file_ = readULEB(program);
754 case DW_LNS_set_column:
755 column_ = readULEB(program);
757 case DW_LNS_negate_stmt:
760 case DW_LNS_set_basic_block:
763 case DW_LNS_const_add_pc:
764 address_ += minLength_ * ((255 - opcodeBase_) / lineRange_);
766 case DW_LNS_fixed_advance_pc:
767 address_ += read<uint16_t>(program);
769 case DW_LNS_set_prologue_end:
770 if (version_ == 2) break; // not supported in version 2
773 case DW_LNS_set_epilogue_begin:
774 if (version_ == 2) break; // not supported in version 2
775 epilogueBegin_ = true;
778 if (version_ == 2) break; // not supported in version 2
779 isa_ = readULEB(program);
783 // Unrecognized standard opcode, slurp the appropriate number of LEB
785 uint8_t argCount = standardOpcodeLengths_[opcode - 1];
793 auto length = readULEB(program);
794 // the opcode itself should be included in the length, so length >= 1
795 FOLLY_SAFE_CHECK(length != 0, "invalid extended opcode length");
796 auto extendedOpcode = read<uint8_t>(program);
799 switch (extendedOpcode) {
800 case DW_LNE_end_sequence:
802 case DW_LNE_set_address:
803 address_ = read<uintptr_t>(program);
805 case DW_LNE_define_file:
806 // We can't process DW_LNE_define_file here, as it would require us to
807 // use unbounded amounts of state (ie. use the heap). We'll do a second
808 // pass (using nextDefineFile()) if necessary.
810 case DW_LNE_set_discriminator:
811 discriminator_ = readULEB(program);
815 // Unrecognized extended opcode
816 program.advance(length);
820 bool Dwarf::LineNumberVM::findAddress(uintptr_t target, Path& file,
822 folly::StringPiece program = data_;
824 // Within each sequence of instructions, the address may only increase.
825 // Unfortunately, within the same compilation unit, sequences may appear
826 // in any order. So any sequence is a candidate if it starts at an address
827 // <= the target address, and we know we've found the target address if
828 // a candidate crosses the target address.
831 LOW_SEQ, // candidate
837 uint64_t prevFile = 0;
838 uint64_t prevLine = 0;
839 while (!program.empty()) {
840 bool seqEnd = !next(program);
842 if (state == START) {
844 state = address_ <= target ? LOW_SEQ : HIGH_SEQ;
848 if (state == LOW_SEQ) {
849 if (address_ > target) {
850 // Found it! Note that ">" is indeed correct (not ">="), as each
851 // sequence is guaranteed to have one entry past-the-end (emitted by
852 // DW_LNE_end_sequence)
856 auto fn = getFileName(prevFile);
857 file = Path(compilationDirectory_,
858 getIncludeDirectory(fn.directoryIndex),
876 } // namespace symbolizer