// Copyright 2014 Google Inc. All Rights Reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. // Package binutils provides access to the GNU binutils. package binutils import ( "debug/elf" "debug/macho" "debug/pe" "encoding/binary" "errors" "fmt" "io" "os" "os/exec" "path/filepath" "regexp" "runtime" "strconv" "strings" "sync" "github.com/google/pprof/internal/elfexec" "github.com/google/pprof/internal/plugin" ) // A Binutils implements plugin.ObjTool by invoking the GNU binutils. type Binutils struct { mu sync.Mutex rep *binrep } var ( objdumpLLVMVerRE = regexp.MustCompile(`LLVM version (?:(\d*)\.(\d*)\.(\d*)|.*(trunk).*)`) // Defined for testing elfOpen = elf.Open ) // binrep is an immutable representation for Binutils. It is atomically // replaced on every mutation to provide thread-safe access. type binrep struct { // Commands to invoke. llvmSymbolizer string llvmSymbolizerFound bool addr2line string addr2lineFound bool nm string nmFound bool objdump string objdumpFound bool isLLVMObjdump bool // if fast, perform symbolization using nm (symbol names only), // instead of file-line detail from the slower addr2line. fast bool } // get returns the current representation for bu, initializing it if necessary. func (bu *Binutils) get() *binrep { bu.mu.Lock() r := bu.rep if r == nil { r = &binrep{} initTools(r, "") bu.rep = r } bu.mu.Unlock() return r } // update modifies the rep for bu via the supplied function. func (bu *Binutils) update(fn func(r *binrep)) { r := &binrep{} bu.mu.Lock() defer bu.mu.Unlock() if bu.rep == nil { initTools(r, "") } else { *r = *bu.rep } fn(r) bu.rep = r } // String returns string representation of the binutils state for debug logging. func (bu *Binutils) String() string { r := bu.get() var llvmSymbolizer, addr2line, nm, objdump string if r.llvmSymbolizerFound { llvmSymbolizer = r.llvmSymbolizer } if r.addr2lineFound { addr2line = r.addr2line } if r.nmFound { nm = r.nm } if r.objdumpFound { objdump = r.objdump } return fmt.Sprintf("llvm-symbolizer=%q addr2line=%q nm=%q objdump=%q fast=%t", llvmSymbolizer, addr2line, nm, objdump, r.fast) } // SetFastSymbolization sets a toggle that makes binutils use fast // symbolization (using nm), which is much faster than addr2line but // provides only symbol name information (no file/line). func (bu *Binutils) SetFastSymbolization(fast bool) { bu.update(func(r *binrep) { r.fast = fast }) } // SetTools processes the contents of the tools option. It // expects a set of entries separated by commas; each entry is a pair // of the form t:path, where cmd will be used to look only for the // tool named t. If t is not specified, the path is searched for all // tools. func (bu *Binutils) SetTools(config string) { bu.update(func(r *binrep) { initTools(r, config) }) } func initTools(b *binrep, config string) { // paths collect paths per tool; Key "" contains the default. paths := make(map[string][]string) for _, t := range strings.Split(config, ",") { name, path := "", t if ct := strings.SplitN(t, ":", 2); len(ct) == 2 { name, path = ct[0], ct[1] } paths[name] = append(paths[name], path) } defaultPath := paths[""] b.llvmSymbolizer, b.llvmSymbolizerFound = chooseExe([]string{"llvm-symbolizer"}, []string{}, append(paths["llvm-symbolizer"], defaultPath...)) b.addr2line, b.addr2lineFound = chooseExe([]string{"addr2line"}, []string{"gaddr2line"}, append(paths["addr2line"], defaultPath...)) // The "-n" option is supported by LLVM since 2011. The output of llvm-nm // and GNU nm with "-n" option is interchangeable for our purposes, so we do // not need to differrentiate them. b.nm, b.nmFound = chooseExe([]string{"llvm-nm", "nm"}, []string{"gnm"}, append(paths["nm"], defaultPath...)) b.objdump, b.objdumpFound, b.isLLVMObjdump = findObjdump(append(paths["objdump"], defaultPath...)) } // findObjdump finds and returns path to preferred objdump binary. // Order of preference is: llvm-objdump, objdump. // On MacOS only, also looks for gobjdump with least preference. // Accepts a list of paths and returns: // a string with path to the preferred objdump binary if found, // or an empty string if not found; // a boolean if any acceptable objdump was found; // a boolean indicating if it is an LLVM objdump. func findObjdump(paths []string) (string, bool, bool) { objdumpNames := []string{"llvm-objdump", "objdump"} if runtime.GOOS == "darwin" { objdumpNames = append(objdumpNames, "gobjdump") } for _, objdumpName := range objdumpNames { if objdump, objdumpFound := findExe(objdumpName, paths); objdumpFound { cmdOut, err := exec.Command(objdump, "--version").Output() if err != nil { continue } if isLLVMObjdump(string(cmdOut)) { return objdump, true, true } if isBuObjdump(string(cmdOut)) { return objdump, true, false } } } return "", false, false } // chooseExe finds and returns path to preferred binary. names is a list of // names to search on both Linux and OSX. osxNames is a list of names specific // to OSX. names always has a higher priority than osxNames. The order of // the name within each list decides its priority (e.g. the first name has a // higher priority than the second name in the list). // // It returns a string with path to the binary and a boolean indicating if any // acceptable binary was found. func chooseExe(names, osxNames []string, paths []string) (string, bool) { if runtime.GOOS == "darwin" { names = append(names, osxNames...) } for _, name := range names { if binary, found := findExe(name, paths); found { return binary, true } } return "", false } // isLLVMObjdump accepts a string with path to an objdump binary, // and returns a boolean indicating if the given binary is an LLVM // objdump binary of an acceptable version. func isLLVMObjdump(output string) bool { fields := objdumpLLVMVerRE.FindStringSubmatch(output) if len(fields) != 5 { return false } if fields[4] == "trunk" { return true } verMajor, err := strconv.Atoi(fields[1]) if err != nil { return false } verPatch, err := strconv.Atoi(fields[3]) if err != nil { return false } if runtime.GOOS == "linux" && verMajor >= 8 { // Ensure LLVM objdump is at least version 8.0 on Linux. // Some flags, like --demangle, and double dashes for options are // not supported by previous versions. return true } if runtime.GOOS == "darwin" { // Ensure LLVM objdump is at least version 10.0.1 on MacOS. return verMajor > 10 || (verMajor == 10 && verPatch >= 1) } return false } // isBuObjdump accepts a string with path to an objdump binary, // and returns a boolean indicating if the given binary is a GNU // binutils objdump binary. No version check is performed. func isBuObjdump(output string) bool { return strings.Contains(output, "GNU objdump") } // findExe looks for an executable command on a set of paths. // If it cannot find it, returns cmd. func findExe(cmd string, paths []string) (string, bool) { for _, p := range paths { cp := filepath.Join(p, cmd) if c, err := exec.LookPath(cp); err == nil { return c, true } } return cmd, false } // Disasm returns the assembly instructions for the specified address range // of a binary. func (bu *Binutils) Disasm(file string, start, end uint64, intelSyntax bool) ([]plugin.Inst, error) { b := bu.get() if !b.objdumpFound { return nil, errors.New("cannot disasm: no objdump tool available") } args := []string{"--disassemble", "--demangle", "--no-show-raw-insn", "--line-numbers", fmt.Sprintf("--start-address=%#x", start), fmt.Sprintf("--stop-address=%#x", end)} if intelSyntax { if b.isLLVMObjdump { args = append(args, "--x86-asm-syntax=intel") } else { args = append(args, "-M", "intel") } } args = append(args, file) cmd := exec.Command(b.objdump, args...) out, err := cmd.Output() if err != nil { return nil, fmt.Errorf("%v: %v", cmd.Args, err) } return disassemble(out) } // Open satisfies the plugin.ObjTool interface. func (bu *Binutils) Open(name string, start, limit, offset uint64, relocationSymbol string) (plugin.ObjFile, error) { b := bu.get() // Make sure file is a supported executable. // This uses magic numbers, mainly to provide better error messages but // it should also help speed. if _, err := os.Stat(name); err != nil { // For testing, do not require file name to exist. if strings.Contains(b.addr2line, "testdata/") { return &fileAddr2Line{file: file{b: b, name: name}}, nil } return nil, err } // Read the first 4 bytes of the file. f, err := os.Open(name) if err != nil { return nil, fmt.Errorf("error opening %s: %v", name, err) } defer f.Close() var header [4]byte if _, err = io.ReadFull(f, header[:]); err != nil { return nil, fmt.Errorf("error reading magic number from %s: %v", name, err) } elfMagic := string(header[:]) // Match against supported file types. if elfMagic == elf.ELFMAG { f, err := b.openELF(name, start, limit, offset, relocationSymbol) if err != nil { return nil, fmt.Errorf("error reading ELF file %s: %v", name, err) } return f, nil } // Mach-O magic numbers can be big or little endian. machoMagicLittle := binary.LittleEndian.Uint32(header[:]) machoMagicBig := binary.BigEndian.Uint32(header[:]) if machoMagicLittle == macho.Magic32 || machoMagicLittle == macho.Magic64 || machoMagicBig == macho.Magic32 || machoMagicBig == macho.Magic64 { f, err := b.openMachO(name, start, limit, offset) if err != nil { return nil, fmt.Errorf("error reading Mach-O file %s: %v", name, err) } return f, nil } if machoMagicLittle == macho.MagicFat || machoMagicBig == macho.MagicFat { f, err := b.openFatMachO(name, start, limit, offset) if err != nil { return nil, fmt.Errorf("error reading fat Mach-O file %s: %v", name, err) } return f, nil } peMagic := string(header[:2]) if peMagic == "MZ" { f, err := b.openPE(name, start, limit, offset) if err != nil { return nil, fmt.Errorf("error reading PE file %s: %v", name, err) } return f, nil } return nil, fmt.Errorf("unrecognized binary format: %s", name) } func (b *binrep) openMachOCommon(name string, of *macho.File, start, limit, offset uint64) (plugin.ObjFile, error) { // Subtract the load address of the __TEXT section. Usually 0 for shared // libraries or 0x100000000 for executables. You can check this value by // running `objdump -private-headers `. textSegment := of.Segment("__TEXT") if textSegment == nil { return nil, fmt.Errorf("could not identify base for %s: no __TEXT segment", name) } if textSegment.Addr > start { return nil, fmt.Errorf("could not identify base for %s: __TEXT segment address (0x%x) > mapping start address (0x%x)", name, textSegment.Addr, start) } base := start - textSegment.Addr if b.fast || (!b.addr2lineFound && !b.llvmSymbolizerFound) { return &fileNM{file: file{b: b, name: name, base: base}}, nil } return &fileAddr2Line{file: file{b: b, name: name, base: base}}, nil } func (b *binrep) openFatMachO(name string, start, limit, offset uint64) (plugin.ObjFile, error) { of, err := macho.OpenFat(name) if err != nil { return nil, fmt.Errorf("error parsing %s: %v", name, err) } defer of.Close() if len(of.Arches) == 0 { return nil, fmt.Errorf("empty fat Mach-O file: %s", name) } var arch macho.Cpu // Use the host architecture. // TODO: This is not ideal because the host architecture may not be the one // that was profiled. E.g. an amd64 host can profile a 386 program. switch runtime.GOARCH { case "386": arch = macho.Cpu386 case "amd64", "amd64p32": arch = macho.CpuAmd64 case "arm", "armbe", "arm64", "arm64be": arch = macho.CpuArm case "ppc": arch = macho.CpuPpc case "ppc64", "ppc64le": arch = macho.CpuPpc64 default: return nil, fmt.Errorf("unsupported host architecture for %s: %s", name, runtime.GOARCH) } for i := range of.Arches { if of.Arches[i].Cpu == arch { return b.openMachOCommon(name, of.Arches[i].File, start, limit, offset) } } return nil, fmt.Errorf("architecture not found in %s: %s", name, runtime.GOARCH) } func (b *binrep) openMachO(name string, start, limit, offset uint64) (plugin.ObjFile, error) { of, err := macho.Open(name) if err != nil { return nil, fmt.Errorf("error parsing %s: %v", name, err) } defer of.Close() return b.openMachOCommon(name, of, start, limit, offset) } func (b *binrep) openELF(name string, start, limit, offset uint64, relocationSymbol string) (plugin.ObjFile, error) { ef, err := elfOpen(name) if err != nil { return nil, fmt.Errorf("error parsing %s: %v", name, err) } defer ef.Close() buildID := "" if f, err := os.Open(name); err == nil { if id, err := elfexec.GetBuildID(f); err == nil { buildID = fmt.Sprintf("%x", id) } } var ( kernelOffset *uint64 pageAligned = func(addr uint64) bool { return addr%4096 == 0 } ) if strings.Contains(name, "vmlinux") || !pageAligned(start) || !pageAligned(limit) || !pageAligned(offset) { // Reading all Symbols is expensive, and we only rarely need it so // we don't want to do it every time. But if _stext happens to be // page-aligned but isn't the same as Vaddr, we would symbolize // wrong. So if the name the addresses aren't page aligned, or if // the name is "vmlinux" we read _stext. We can be wrong if: (1) // someone passes a kernel path that doesn't contain "vmlinux" AND // (2) _stext is page-aligned AND (3) _stext is not at Vaddr symbols, err := ef.Symbols() if err != nil && err != elf.ErrNoSymbols { return nil, err } // The kernel relocation symbol (the mapping start address) can be either // _text or _stext. When profiles are generated by `perf`, which one was used is // distinguished by the mapping name for the kernel image: // '[kernel.kallsyms]_text' or '[kernel.kallsyms]_stext', respectively. If we haven't // been able to parse it from the mapping, we default to _stext. if relocationSymbol == "" { relocationSymbol = "_stext" } for _, s := range symbols { if s.Name == relocationSymbol { kernelOffset = &s.Value break } } } // Check that we can compute a base for the binary. This may not be the // correct base value, so we don't save it. We delay computing the actual base // value until we have a sample address for this mapping, so that we can // correctly identify the associated program segment that is needed to compute // the base. if _, err := elfexec.GetBase(&ef.FileHeader, elfexec.FindTextProgHeader(ef), kernelOffset, start, limit, offset); err != nil { return nil, fmt.Errorf("could not identify base for %s: %v", name, err) } if b.fast || (!b.addr2lineFound && !b.llvmSymbolizerFound) { return &fileNM{file: file{ b: b, name: name, buildID: buildID, m: &elfMapping{start: start, limit: limit, offset: offset, kernelOffset: kernelOffset}, }}, nil } return &fileAddr2Line{file: file{ b: b, name: name, buildID: buildID, m: &elfMapping{start: start, limit: limit, offset: offset, kernelOffset: kernelOffset}, }}, nil } func (b *binrep) openPE(name string, start, limit, offset uint64) (plugin.ObjFile, error) { pf, err := pe.Open(name) if err != nil { return nil, fmt.Errorf("error parsing %s: %v", name, err) } defer pf.Close() var imageBase uint64 switch h := pf.OptionalHeader.(type) { case *pe.OptionalHeader32: imageBase = uint64(h.ImageBase) case *pe.OptionalHeader64: imageBase = uint64(h.ImageBase) default: return nil, fmt.Errorf("unknown OptionalHeader %T", pf.OptionalHeader) } var base uint64 if start > 0 { base = start - imageBase } if b.fast || (!b.addr2lineFound && !b.llvmSymbolizerFound) { return &fileNM{file: file{b: b, name: name, base: base}}, nil } return &fileAddr2Line{file: file{b: b, name: name, base: base}}, nil } // elfMapping stores the parameters of a runtime mapping that are needed to // identify the ELF segment associated with a mapping. type elfMapping struct { // Runtime mapping parameters. start, limit, offset uint64 // Offset of kernel relocation symbol. Only defined for kernel images, nil otherwise. kernelOffset *uint64 } // findProgramHeader returns the program segment that matches the current // mapping and the given address, or an error if it cannot find a unique program // header. func (m *elfMapping) findProgramHeader(ef *elf.File, addr uint64) (*elf.ProgHeader, error) { // For user space executables, we try to find the actual program segment that // is associated with the given mapping. Skip this search if limit <= start. // We cannot use just a check on the start address of the mapping to tell if // it's a kernel / .ko module mapping, because with quipper address remapping // enabled, the address would be in the lower half of the address space. if m.kernelOffset != nil || m.start >= m.limit || m.limit >= (uint64(1)<<63) { // For the kernel, find the program segment that includes the .text section. return elfexec.FindTextProgHeader(ef), nil } // Fetch all the loadable segments. var phdrs []elf.ProgHeader for i := range ef.Progs { if ef.Progs[i].Type == elf.PT_LOAD { phdrs = append(phdrs, ef.Progs[i].ProgHeader) } } // Some ELF files don't contain any loadable program segments, e.g. .ko // kernel modules. It's not an error to have no header in such cases. if len(phdrs) == 0 { return nil, nil } // Get all program headers associated with the mapping. headers := elfexec.ProgramHeadersForMapping(phdrs, m.offset, m.limit-m.start) if len(headers) == 0 { return nil, errors.New("no program header matches mapping info") } if len(headers) == 1 { return headers[0], nil } // Use the file offset corresponding to the address to symbolize, to narrow // down the header. return elfexec.HeaderForFileOffset(headers, addr-m.start+m.offset) } // file implements the binutils.ObjFile interface. type file struct { b *binrep name string buildID string baseOnce sync.Once // Ensures the base, baseErr and isData are computed once. base uint64 baseErr error // Any eventual error while computing the base. isData bool // Mapping information. Relevant only for ELF files, nil otherwise. m *elfMapping } // computeBase computes the relocation base for the given binary file only if // the elfMapping field is set. It populates the base and isData fields and // returns an error. func (f *file) computeBase(addr uint64) error { if f == nil || f.m == nil { return nil } if addr < f.m.start || addr >= f.m.limit { return fmt.Errorf("specified address %x is outside the mapping range [%x, %x] for file %q", addr, f.m.start, f.m.limit, f.name) } ef, err := elfOpen(f.name) if err != nil { return fmt.Errorf("error parsing %s: %v", f.name, err) } defer ef.Close() ph, err := f.m.findProgramHeader(ef, addr) if err != nil { return fmt.Errorf("failed to find program header for file %q, ELF mapping %#v, address %x: %v", f.name, *f.m, addr, err) } base, err := elfexec.GetBase(&ef.FileHeader, ph, f.m.kernelOffset, f.m.start, f.m.limit, f.m.offset) if err != nil { return err } f.base = base f.isData = ph != nil && ph.Flags&elf.PF_X == 0 return nil } func (f *file) Name() string { return f.name } func (f *file) ObjAddr(addr uint64) (uint64, error) { f.baseOnce.Do(func() { f.baseErr = f.computeBase(addr) }) if f.baseErr != nil { return 0, f.baseErr } return addr - f.base, nil } func (f *file) BuildID() string { return f.buildID } func (f *file) SourceLine(addr uint64) ([]plugin.Frame, error) { f.baseOnce.Do(func() { f.baseErr = f.computeBase(addr) }) if f.baseErr != nil { return nil, f.baseErr } return nil, nil } func (f *file) Close() error { return nil } func (f *file) Symbols(r *regexp.Regexp, addr uint64) ([]*plugin.Sym, error) { // Get from nm a list of symbols sorted by address. cmd := exec.Command(f.b.nm, "-n", f.name) out, err := cmd.Output() if err != nil { return nil, fmt.Errorf("%v: %v", cmd.Args, err) } return findSymbols(out, f.name, r, addr) } // fileNM implements the binutils.ObjFile interface, using 'nm' to map // addresses to symbols (without file/line number information). It is // faster than fileAddr2Line. type fileNM struct { file addr2linernm *addr2LinerNM } func (f *fileNM) SourceLine(addr uint64) ([]plugin.Frame, error) { f.baseOnce.Do(func() { f.baseErr = f.computeBase(addr) }) if f.baseErr != nil { return nil, f.baseErr } if f.addr2linernm == nil { addr2liner, err := newAddr2LinerNM(f.b.nm, f.name, f.base) if err != nil { return nil, err } f.addr2linernm = addr2liner } return f.addr2linernm.addrInfo(addr) } // fileAddr2Line implements the binutils.ObjFile interface, using // llvm-symbolizer, if that's available, or addr2line to map addresses to // symbols (with file/line number information). It can be slow for large // binaries with debug information. type fileAddr2Line struct { once sync.Once file addr2liner *addr2Liner llvmSymbolizer *llvmSymbolizer isData bool } func (f *fileAddr2Line) SourceLine(addr uint64) ([]plugin.Frame, error) { f.baseOnce.Do(func() { f.baseErr = f.computeBase(addr) }) if f.baseErr != nil { return nil, f.baseErr } f.once.Do(f.init) if f.llvmSymbolizer != nil { return f.llvmSymbolizer.addrInfo(addr) } if f.addr2liner != nil { return f.addr2liner.addrInfo(addr) } return nil, fmt.Errorf("could not find local addr2liner") } func (f *fileAddr2Line) init() { if llvmSymbolizer, err := newLLVMSymbolizer(f.b.llvmSymbolizer, f.name, f.base, f.isData); err == nil { f.llvmSymbolizer = llvmSymbolizer return } if addr2liner, err := newAddr2Liner(f.b.addr2line, f.name, f.base); err == nil { f.addr2liner = addr2liner // When addr2line encounters some gcc compiled binaries, it // drops interesting parts of names in anonymous namespaces. // Fallback to NM for better function names. if nm, err := newAddr2LinerNM(f.b.nm, f.name, f.base); err == nil { f.addr2liner.nm = nm } } } func (f *fileAddr2Line) Close() error { if f.llvmSymbolizer != nil { f.llvmSymbolizer.rw.close() f.llvmSymbolizer = nil } if f.addr2liner != nil { f.addr2liner.rw.close() f.addr2liner = nil } return nil }