runtime/pprof, cmd/pprof: fix profiling for PIE

In order to support pprof for position independent executables, pprof
needs to adjust the PC addresses stored in the profile by the address at
which the program is loaded. The legacy profiling support which we use
already supports recording the GNU/Linux /proc/self/maps data
immediately after the CPU samples, so do that. Also change the pprof
symbolizer to use the information, if available, when looking up
addresses in the Go pcline data.

Fixes #15714.

Change-Id: I4bf679210ef7c51d85cf873c968ce82db8898e3e
Reviewed-on: https://go-review.googlesource.com/23525
Reviewed-by: Michael Hudson-Doyle <michael.hudson@canonical.com>

src/cmd/internal/objfile/elf.go

@@ -106,6 +106,15 @@ func (f *elfFile) goarch() string {
 	return ""
 }
 
+func (f *elfFile) loadAddress() (uint64, error) {
+	for _, p := range f.elf.Progs {
+		if p.Type == elf.PT_LOAD {
+			return p.Vaddr, nil
+		}
+	}
+	return 0, fmt.Errorf("unknown load address")
+}
+
 func (f *elfFile) dwarf() (*dwarf.Data, error) {
 	return f.elf.DWARF()
 }

src/cmd/internal/objfile/goobj.go

@@ -94,6 +94,10 @@ func (f *goobjFile) goarch() string {
 	return "GOARCH unimplemented for debug/goobj files"
 }
 
+func (f *goobjFile) loadAddress() (uint64, error) {
+	return 0, fmt.Errorf("unknown load address")
+}
+
 func (f *goobjFile) dwarf() (*dwarf.Data, error) {
 	return nil, errors.New("no DWARF data in go object file")
 }

src/cmd/internal/objfile/macho.go

@@ -125,6 +125,10 @@ func (x uint64s) Len() int           { return len(x) }
 func (x uint64s) Swap(i, j int)      { x[i], x[j] = x[j], x[i] }
 func (x uint64s) Less(i, j int) bool { return x[i] < x[j] }
 
+func (f *machoFile) loadAddress() (uint64, error) {
+	return 0, fmt.Errorf("unknown load address")
+}
+
 func (f *machoFile) dwarf() (*dwarf.Data, error) {
 	return f.macho.DWARF()
 }

src/cmd/internal/objfile/objfile.go

@@ -18,6 +18,7 @@ type rawFile interface {
 	pcln() (textStart uint64, symtab, pclntab []byte, err error)
 	text() (textStart uint64, text []byte, err error)
 	goarch() string
+	loadAddress() (uint64, error)
 	dwarf() (*dwarf.Data, error)
 }
 
@@ -95,6 +96,13 @@ func (f *File) GOARCH() string {
 	return f.raw.goarch()
 }
 
+// LoadAddress returns the expected load address of the file.
+// This differs from the actual load address for a position-independent
+// executable.
+func (f *File) LoadAddress() (uint64, error) {
+	return f.raw.loadAddress()
+}
+
 // DWARF returns DWARF debug data for the file, if any.
 // This is for cmd/pprof to locate cgo functions.
 func (f *File) DWARF() (*dwarf.Data, error) {

src/cmd/internal/objfile/pe.go

@@ -199,6 +199,10 @@ func (f *peFile) goarch() string {
 	return ""
 }
 
+func (f *peFile) loadAddress() (uint64, error) {
+	return 0, fmt.Errorf("unknown load address")
+}
+
 func (f *peFile) dwarf() (*dwarf.Data, error) {
 	return f.pe.DWARF()
 }

src/cmd/internal/objfile/plan9obj.go

@@ -147,6 +147,10 @@ func (f *plan9File) goarch() string {
 	return ""
 }
 
+func (f *plan9File) loadAddress() (uint64, error) {
+	return 0, fmt.Errorf("unknown load address")
+}
+
 func (f *plan9File) dwarf() (*dwarf.Data, error) {
 	return nil, errors.New("no DWARF data in Plan 9 file")
 }

src/cmd/pprof/pprof.go

@@ -117,6 +117,9 @@ func (*objTool) Open(name string, start uint64) (plugin.ObjFile, error) {
 		name: name,
 		file: of,
 	}
+	if load, err := of.LoadAddress(); err == nil {
+		f.offset = start - load
+	}
 	return f, nil
 }
 
@@ -169,10 +172,11 @@ func (*objTool) SetConfig(config string) {
 // (instead of invoking GNU binutils).
 // A file represents a single executable being analyzed.
 type file struct {
-	name string
-	sym  []objfile.Sym
-	file *objfile.File
-	pcln *gosym.Table
+	name   string
+	offset uint64
+	sym    []objfile.Sym
+	file   *objfile.File
+	pcln   *gosym.Table
 
 	triedDwarf bool
 	dwarf      *dwarf.Data
@@ -200,6 +204,7 @@ func (f *file) SourceLine(addr uint64) ([]plugin.Frame, error) {
 		}
 		f.pcln = pcln
 	}
+	addr -= f.offset
 	file, line, fn := f.pcln.PCToLine(addr)
 	if fn != nil {
 		frame := []plugin.Frame{

src/go/build/deps_test.go

@@ -173,7 +173,7 @@ var pkgDeps = map[string][]string{
 	"regexp":         {"L2", "regexp/syntax"},
 	"regexp/syntax":  {"L2"},
 	"runtime/debug":  {"L2", "fmt", "io/ioutil", "os", "time"},
-	"runtime/pprof":  {"L2", "fmt", "text/tabwriter"},
+	"runtime/pprof":  {"L2", "fmt", "os", "text/tabwriter"},
 	"runtime/trace":  {"L0"},
 	"text/tabwriter": {"L2"},
 

src/runtime/crash_cgo_test.go

@@ -263,3 +263,33 @@ func TestCgoPprof(t *testing.T) {
 		t.Error("missing cpuHog in pprof output")
 	}
 }
+
+func TestCgoPprofPIE(t *testing.T) {
+	if runtime.GOOS != "linux" || runtime.GOARCH != "amd64" {
+		t.Skipf("not yet supported on %s/%s", runtime.GOOS, runtime.GOARCH)
+	}
+	testenv.MustHaveGoRun(t)
+
+	exe, err := buildTestProg(t, "testprogcgo", "-ldflags=-extldflags=-pie")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	got, err := testEnv(exec.Command(exe, "CgoPprof")).CombinedOutput()
+	if err != nil {
+		t.Fatal(err)
+	}
+	fn := strings.TrimSpace(string(got))
+	defer os.Remove(fn)
+
+	top, err := exec.Command("go", "tool", "pprof", "-top", "-nodecount=1", exe, fn).CombinedOutput()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	t.Logf("%s", top)
+
+	if !bytes.Contains(top, []byte("cpuHog")) {
+		t.Error("missing cpuHog in pprof output")
+	}
+}

src/runtime/crash_test.go

@@ -69,7 +69,7 @@ func runTestProg(t *testing.T, binary, name string) string {
 	return string(got)
 }
 
-func buildTestProg(t *testing.T, binary string) (string, error) {
+func buildTestProg(t *testing.T, binary string, flags ...string) (string, error) {
 	checkStaleRuntime(t)
 
 	testprog.Lock()
@@ -86,23 +86,27 @@ func buildTestProg(t *testing.T, binary string) (string, error) {
 	if testprog.target == nil {
 		testprog.target = make(map[string]buildexe)
 	}
-	target, ok := testprog.target[binary]
+	name := binary
+	if len(flags) > 0 {
+		name += "_" + strings.Join(flags, "_")
+	}
+	target, ok := testprog.target[name]
 	if ok {
 		return target.exe, target.err
 	}
 
-	exe := filepath.Join(testprog.dir, binary+".exe")
-	cmd := exec.Command("go", "build", "-o", exe)
+	exe := filepath.Join(testprog.dir, name+".exe")
+	cmd := exec.Command("go", append([]string{"build", "-o", exe}, flags...)...)
 	cmd.Dir = "testdata/" + binary
 	out, err := testEnv(cmd).CombinedOutput()
 	if err != nil {
 		exe = ""
-		target.err = fmt.Errorf("building %s: %v\n%s", binary, err, out)
-		testprog.target[binary] = target
+		target.err = fmt.Errorf("building %s %v: %v\n%s", binary, flags, err, out)
+		testprog.target[name] = target
 		return "", target.err
 	}
 	target.exe = exe
-	testprog.target[binary] = target
+	testprog.target[name] = target
 	return exe, nil
 }
 

src/runtime/pprof/pprof.go

@@ -13,6 +13,7 @@ import (
 	"bytes"
 	"fmt"
 	"io"
+	"os"
 	"runtime"
 	"sort"
 	"strings"
@@ -620,6 +621,42 @@ func profileWriter(w io.Writer) {
 		}
 		w.Write(data)
 	}
+
+	// We are emitting the legacy profiling format, which permits
+	// a memory map following the CPU samples. The memory map is
+	// simply a copy of the GNU/Linux /proc/self/maps file. The
+	// profiler uses the memory map to map PC values in shared
+	// libraries to a shared library in the filesystem, in order
+	// to report the correct function and, if the shared library
+	// has debug info, file/line. This is particularly useful for
+	// PIE (position independent executables) as on ELF systems a
+	// PIE is simply an executable shared library.
+	//
+	// Because the profiling format expects the memory map in
+	// GNU/Linux format, we only do this on GNU/Linux for now. To
+	// add support for profiling PIE on other ELF-based systems,
+	// it may be necessary to map the system-specific mapping
+	// information to the GNU/Linux format. For a reasonably
+	// portable C++ version, see the FillProcSelfMaps function in
+	// https://github.com/gperftools/gperftools/blob/master/src/base/sysinfo.cc
+	//
+	// The code that parses this mapping for the pprof tool is
+	// ParseMemoryMap in cmd/internal/pprof/legacy_profile.go, but
+	// don't change that code, as similar code exists in other
+	// (non-Go) pprof readers. Change this code so that that code works.
+	//
+	// We ignore errors reading or copying the memory map; the
+	// profile is likely usable without it, and we have no good way
+	// to report errors.
+	if runtime.GOOS == "linux" {
+		f, err := os.Open("/proc/self/maps")
+		if err == nil {
+			io.WriteString(w, "\nMAPPED_LIBRARIES:\n")
+			io.Copy(w, f)
+			f.Close()
+		}
+	}
+
 	cpu.done <- true
 }
 

src/runtime/pprof/pprof_test.go

@@ -86,10 +86,14 @@ func TestCPUProfileMultithreaded(t *testing.T) {
 	})
 }
 
-func parseProfile(t *testing.T, bytes []byte, f func(uintptr, []uintptr)) {
+func parseProfile(t *testing.T, valBytes []byte, f func(uintptr, []uintptr)) {
 	// Convert []byte to []uintptr.
-	l := len(bytes) / int(unsafe.Sizeof(uintptr(0)))
-	val := *(*[]uintptr)(unsafe.Pointer(&bytes))
+	l := len(valBytes)
+	if i := bytes.Index(valBytes, []byte("\nMAPPED_LIBRARIES:\n")); i >= 0 {
+		l = i
+	}
+	l /= int(unsafe.Sizeof(uintptr(0)))
+	val := *(*[]uintptr)(unsafe.Pointer(&valBytes))
 	val = val[:l]
 
 	// 5 for the header, 3 for the trailer.