cmd/link: generate .xdata PE section

This CL adds a .xdata section to the PE file generated by the Go linker.
It is also the first CL of the SEH chain that adds effective support
for unwinding the Go stack, as demonstrated by the newly added tests.

The .xdata section is a standard PE section that contains
an array of unwind data info structures. This structures are used to
record the effects a function has on the stack pointer,
and where the nonvolatile registers are saved on the stack [1].

Note that this CL still does not support unwinding the cgo stack.

Updates #57302

[1] https://learn.microsoft.com/en-us/cpp/build/exception-handling-x64#struct-unwind_info

Change-Id: I6f305a51ed130b758ff9ca7b90c091e50a109a6f
Reviewed-on: https://go-review.googlesource.com/c/go/+/457455
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Davis Goodin <dagood@microsoft.com>
Run-TryBot: Quim Muntal <quimmuntal@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>

src/cmd/link/internal/ld/asmb.go

@@ -63,6 +63,9 @@ func asmb(ctxt *Link) {
 	if Segpdata.Filelen > 0 {
 		writeParallel(&wg, pdatablk, ctxt, Segpdata.Fileoff, Segpdata.Vaddr, Segpdata.Filelen)
 	}
+	if Segxdata.Filelen > 0 {
+		writeParallel(&wg, xdatablk, ctxt, Segxdata.Fileoff, Segxdata.Vaddr, Segxdata.Filelen)
+	}
 
 	wg.Wait()
 }

src/cmd/link/internal/ld/data.go

@@ -1158,6 +1158,10 @@ func pdatablk(ctxt *Link, out *OutBuf, addr int64, size int64) {
 	writeBlocks(ctxt, out, ctxt.outSem, ctxt.loader, []loader.Sym{sehp.pdata}, addr, size, zeros[:])
 }
 
+func xdatablk(ctxt *Link, out *OutBuf, addr int64, size int64) {
+	writeBlocks(ctxt, out, ctxt.outSem, ctxt.loader, []loader.Sym{sehp.xdata}, addr, size, zeros[:])
+}
+
 var covCounterDataStartOff, covCounterDataLen uint64
 
 var zeros [512]byte
@@ -1686,6 +1690,10 @@ func (ctxt *Link) dodata(symGroupType []sym.SymKind) {
 		sect.Extnum = n
 		n++
 	}
+	for _, sect := range Segxdata.Sections {
+		sect.Extnum = n
+		n++
+	}
 }
 
 // allocateDataSectionForSym creates a new sym.Section into which a
@@ -2164,7 +2172,12 @@ func (state *dodataState) allocateSEHSections(ctxt *Link) {
 	if sehp.pdata > 0 {
 		sect := state.allocateDataSectionForSym(&Segpdata, sehp.pdata, 04)
 		state.assignDsymsToSection(sect, []loader.Sym{sehp.pdata}, sym.SRODATA, aligndatsize)
-		state.checkdatsize(sym.SPDATASECT)
+		state.checkdatsize(sym.SSEHSECT)
+	}
+	if sehp.xdata > 0 {
+		sect := state.allocateNamedDataSection(&Segxdata, ".xdata", []sym.SymKind{}, 04)
+		state.assignDsymsToSection(sect, []loader.Sym{sehp.xdata}, sym.SRODATA, aligndatsize)
+		state.checkdatsize(sym.SSEHSECT)
 	}
 }
 
@@ -2719,6 +2732,21 @@ func (ctxt *Link) address() []*sym.Segment {
 		Segpdata.Length = va - Segpdata.Vaddr
 	}
 
+	if len(Segxdata.Sections) > 0 {
+		va = uint64(Rnd(int64(va), int64(*FlagRound)))
+		order = append(order, &Segxdata)
+		Segxdata.Rwx = 04
+		Segxdata.Vaddr = va
+		// Segxdata.Sections is intended to contain just one section.
+		// Loop through the slice anyway for consistency.
+		for _, s := range Segxdata.Sections {
+			va = uint64(Rnd(int64(va), int64(s.Align)))
+			s.Vaddr = va
+			va += s.Length
+		}
+		Segxdata.Length = va - Segxdata.Vaddr
+	}
+
 	va = uint64(Rnd(int64(va), int64(*FlagRound)))
 	order = append(order, &Segdwarf)
 	Segdwarf.Rwx = 06
@@ -2770,8 +2798,10 @@ func (ctxt *Link) address() []*sym.Segment {
 		}
 	}
 
-	if sect := ldr.SymSect(sehp.pdata); sect != nil {
-		ldr.AddToSymValue(sehp.pdata, int64(sect.Vaddr))
+	for _, s := range []loader.Sym{sehp.pdata, sehp.xdata} {
+		if sect := ldr.SymSect(s); sect != nil {
+			ldr.AddToSymValue(s, int64(sect.Vaddr))
+		}
 	}
 
 	if ctxt.BuildMode == BuildModeShared {

src/cmd/link/internal/ld/lib.go

@@ -329,8 +329,9 @@ var (
 	Segdata      sym.Segment
 	Segdwarf     sym.Segment
 	Segpdata     sym.Segment // windows-only
+	Segxdata     sym.Segment // windows-only
 
-	Segments = []*sym.Segment{&Segtext, &Segrodata, &Segrelrodata, &Segdata, &Segdwarf, &Segpdata}
+	Segments = []*sym.Segment{&Segtext, &Segrodata, &Segrelrodata, &Segdata, &Segdwarf, &Segpdata, &Segxdata}
 )
 
 const pkgdef = "__.PKGDEF"

src/cmd/link/internal/ld/pe.go

@@ -434,6 +434,7 @@ type peFile struct {
 	bssSect        *peSection
 	ctorsSect      *peSection
 	pdataSect      *peSection
+	xdataSect      *peSection
 	nextSectOffset uint32
 	nextFileOffset uint32
 	symtabOffset   int64 // offset to the start of symbol table
@@ -501,6 +502,8 @@ func (f *peFile) addDWARF() {
 
 // addSEH adds SEH information to the COFF file f.
 func (f *peFile) addSEH(ctxt *Link) {
+	// .pdata section can exist without the .xdata section.
+	// .xdata section depends on the .pdata section.
 	if Segpdata.Length == 0 {
 		return
 	}
@@ -512,10 +515,19 @@ func (f *peFile) addSEH(ctxt *Link) {
 	}
 	pefile.pdataSect = d
 	d.checkSegment(&Segpdata)
-	// TODO: remove extraSize once the dummy unwind info is removed from the .pdata section.
-	const extraSize = 12
-	pefile.dataDirectory[pe.IMAGE_DIRECTORY_ENTRY_EXCEPTION].VirtualAddress = d.virtualAddress + extraSize
-	pefile.dataDirectory[pe.IMAGE_DIRECTORY_ENTRY_EXCEPTION].Size = d.virtualSize - extraSize
+	pefile.dataDirectory[pe.IMAGE_DIRECTORY_ENTRY_EXCEPTION].VirtualAddress = d.virtualAddress
+	pefile.dataDirectory[pe.IMAGE_DIRECTORY_ENTRY_EXCEPTION].Size = d.virtualSize
+
+	if Segxdata.Length > 0 {
+		d = pefile.addSection(".xdata", int(Segxdata.Length), int(Segxdata.Length))
+		d.characteristics = IMAGE_SCN_CNT_INITIALIZED_DATA | IMAGE_SCN_MEM_READ
+		if ctxt.LinkMode == LinkExternal {
+			// Some gcc versions don't honor the default alignment for the .xdata section.
+			d.characteristics |= IMAGE_SCN_ALIGN_4BYTES
+		}
+		pefile.xdataSect = d
+		d.checkSegment(&Segxdata)
+	}
 }
 
 // addInitArray adds .ctors COFF section to the file f.
@@ -626,6 +638,9 @@ func (f *peFile) emitRelocations(ctxt *Link) {
 	if sehp.pdata != 0 {
 		sects = append(sects, relsect{f.pdataSect, &Segpdata, []loader.Sym{sehp.pdata}})
 	}
+	if sehp.xdata != 0 {
+		sects = append(sects, relsect{f.xdataSect, &Segxdata, []loader.Sym{sehp.xdata}})
+	}
 	for _, s := range sects {
 		s.peSect.emitRelocations(ctxt.Out, func() int {
 			var n int

src/cmd/link/internal/ld/seh.go

@@ -12,6 +12,7 @@ import (
 
 var sehp struct {
 	pdata loader.Sym
+	xdata loader.Sym
 }
 
 func writeSEH(ctxt *Link) {
@@ -29,11 +30,15 @@ func writeSEHAMD64(ctxt *Link) {
 		s.SetAlign(4)
 		return s
 	}
-	pdata := mkSecSym(".pdata", sym.SPDATASECT)
-	// TODO: the following 12 bytes represent a dummy unwind info,
-	// remove once unwind infos are encoded in the .xdata section.
-	pdata.AddUint64(ctxt.Arch, 0)
-	pdata.AddUint32(ctxt.Arch, 0)
+	pdata := mkSecSym(".pdata", sym.SSEHSECT)
+	xdata := mkSecSym(".xdata", sym.SSEHSECT)
+	// The .xdata entries have very low cardinality
+	// as it only contains frame pointer operations,
+	// which are very similar across functions.
+	// These are referenced by .pdata entries using
+	// an RVA, so it is possible, and binary-size wise,
+	// to deduplicate .xdata entries.
+	uwcache := make(map[string]int64) // aux symbol name --> .xdata offset
 	for _, s := range ctxt.Textp {
 		if fi := ldr.FuncInfo(s); !fi.Valid() || fi.TopFrame() {
 			continue
@@ -42,13 +47,20 @@ func writeSEHAMD64(ctxt *Link) {
 		if uw == 0 {
 			continue
 		}
+		name := ctxt.SymName(uw)
+		off, cached := uwcache[name]
+		if !cached {
+			off = xdata.Size()
+			uwcache[name] = off
+			xdata.AddBytes(ldr.Data(uw))
+		}
 
 		// Reference:
 		// https://learn.microsoft.com/en-us/cpp/build/exception-handling-x64#struct-runtime_function
 		pdata.AddPEImageRelativeAddrPlus(ctxt.Arch, s, 0)
 		pdata.AddPEImageRelativeAddrPlus(ctxt.Arch, s, ldr.SymSize(s))
-		// TODO: reference the .xdata symbol.
-		pdata.AddPEImageRelativeAddrPlus(ctxt.Arch, pdata.Sym(), 0)
+		pdata.AddPEImageRelativeAddrPlus(ctxt.Arch, xdata.Sym(), off)
 	}
 	sehp.pdata = pdata.Sym()
+	sehp.xdata = xdata.Sym()
 }

src/cmd/link/internal/sym/symkind.go

@@ -126,7 +126,7 @@ const (
 
 	// SEH symbol types
 	SSEHUNWINDINFO
-	SPDATASECT
+	SSEHSECT
 )
 
 // AbiSymKindToSymKind maps values read from object files (which are

src/internal/syscall/windows/syscall_windows.go

@@ -401,3 +401,4 @@ type FILE_ID_BOTH_DIR_INFO struct {
 //sys	GetVolumeInformationByHandle(file syscall.Handle, volumeNameBuffer *uint16, volumeNameSize uint32, volumeNameSerialNumber *uint32, maximumComponentLength *uint32, fileSystemFlags *uint32, fileSystemNameBuffer *uint16, fileSystemNameSize uint32) (err error) = GetVolumeInformationByHandleW
 
 //sys	RtlLookupFunctionEntry(pc uintptr, baseAddress *uintptr, table *byte) (ret uintptr) = kernel32.RtlLookupFunctionEntry
+//sys	RtlVirtualUnwind(handlerType uint32, baseAddress uintptr, pc uintptr, entry uintptr, ctxt uintptr, data *uintptr, frame *uintptr, ctxptrs *byte) (ret uintptr) = kernel32.RtlVirtualUnwind

src/runtime/runtime-seh_windows_test.go

@@ -8,7 +8,9 @@ import (
 	"internal/abi"
 	"internal/syscall/windows"
 	"runtime"
+	"slices"
 	"testing"
+	"unsafe"
 )
 
 func sehf1() int {
@@ -61,3 +63,129 @@ func TestSehLookupFunctionEntry(t *testing.T) {
 		}
 	}
 }
+
+func sehCallers() []uintptr {
+	// We don't need a real context,
+	// RtlVirtualUnwind just needs a context with
+	// valid a pc, sp and fp (aka bp).
+	ctx := runtime.NewContextStub()
+
+	pcs := make([]uintptr, 15)
+	var base, frame uintptr
+	var n int
+	for i := 0; i < len(pcs); i++ {
+		fn := windows.RtlLookupFunctionEntry(ctx.GetPC(), &base, nil)
+		if fn == 0 {
+			break
+		}
+		windows.RtlVirtualUnwind(0, base, ctx.GetPC(), fn, uintptr(unsafe.Pointer(&ctx)), nil, &frame, nil)
+		n++
+		pcs[i] = ctx.GetPC()
+	}
+	return pcs[:n]
+}
+
+// SEH unwinding does not report inlined frames.
+//
+//go:noinline
+func sehf3(pan bool) []uintptr {
+	return sehf4(pan)
+}
+
+//go:noinline
+func sehf4(pan bool) []uintptr {
+	var pcs []uintptr
+	if pan {
+		panic("sehf4")
+	}
+	pcs = sehCallers()
+	return pcs
+}
+
+func testSehCallersEqual(t *testing.T, pcs []uintptr, want []string) {
+	t.Helper()
+	got := make([]string, 0, len(want))
+	for _, pc := range pcs {
+		fn := runtime.FuncForPC(pc)
+		if fn == nil || len(got) >= len(want) {
+			break
+		}
+		name := fn.Name()
+		switch name {
+		case "runtime.deferCallSave", "runtime.runOpenDeferFrame", "runtime.panicmem":
+			// These functions are skipped as they appear inconsistently depending
+			// whether inlining is on or off.
+			continue
+		}
+		got = append(got, name)
+	}
+	if !slices.Equal(want, got) {
+		t.Fatalf("wanted %v, got %v", want, got)
+	}
+}
+
+func TestSehUnwind(t *testing.T) {
+	if runtime.GOARCH != "amd64" {
+		t.Skip("skipping amd64-only test")
+	}
+	pcs := sehf3(false)
+	testSehCallersEqual(t, pcs, []string{"runtime_test.sehCallers", "runtime_test.sehf4",
+		"runtime_test.sehf3", "runtime_test.TestSehUnwind"})
+}
+
+func TestSehUnwindPanic(t *testing.T) {
+	if runtime.GOARCH != "amd64" {
+		t.Skip("skipping amd64-only test")
+	}
+	want := []string{"runtime_test.sehCallers", "runtime_test.TestSehUnwindPanic.func1", "runtime.gopanic",
+		"runtime_test.sehf4", "runtime_test.sehf3", "runtime_test.TestSehUnwindPanic"}
+	defer func() {
+		if r := recover(); r == nil {
+			t.Fatal("did not panic")
+		}
+		pcs := sehCallers()
+		testSehCallersEqual(t, pcs, want)
+	}()
+	sehf3(true)
+}
+
+func TestSehUnwindDoublePanic(t *testing.T) {
+	if runtime.GOARCH != "amd64" {
+		t.Skip("skipping amd64-only test")
+	}
+	want := []string{"runtime_test.sehCallers", "runtime_test.TestSehUnwindDoublePanic.func1.1", "runtime.gopanic",
+		"runtime_test.TestSehUnwindDoublePanic.func1", "runtime.gopanic", "runtime_test.TestSehUnwindDoublePanic"}
+	defer func() {
+		defer func() {
+			if recover() == nil {
+				t.Fatal("did not panic")
+			}
+			pcs := sehCallers()
+			testSehCallersEqual(t, pcs, want)
+		}()
+		if recover() == nil {
+			t.Fatal("did not panic")
+		}
+		panic(2)
+	}()
+	panic(1)
+}
+
+func TestSehUnwindNilPointerPanic(t *testing.T) {
+	if runtime.GOARCH != "amd64" {
+		t.Skip("skipping amd64-only test")
+	}
+	want := []string{"runtime_test.sehCallers", "runtime_test.TestSehUnwindNilPointerPanic.func1", "runtime.gopanic",
+		"runtime.sigpanic", "runtime_test.TestSehUnwindNilPointerPanic"}
+	defer func() {
+		if r := recover(); r == nil {
+			t.Fatal("did not panic")
+		}
+		pcs := sehCallers()
+		testSehCallersEqual(t, pcs, want)
+	}()
+	var p *int
+	if *p == 3 {
+		t.Fatal("did not see nil pointer panic")
+	}
+}