[dev.link] cmd/link: support updates to contents of obj-based Syms

Add in the hooks to SymbolBuilder and to the loader to allow
the linker to make modifications to a non-external symbol (e.g.
a sym whose index is less than loader.extStart).

The basic idea is to manufacture a new external symbol with the same
name and version, then import the old symbol's content (type, data,
relocations, etc) into the payload struct for the new symbol, and
finally redirect the name lookup tables to target the new sym for the
specified name/version.

This change is needed in order to convert over the host object loaders
to avoid use of sym.Symbol.

Change-Id: I79cd42b23794e830bbdbcbcd2c500c35c351f01f
Reviewed-on: https://go-review.googlesource.com/c/go/+/211897
Reviewed-by: Cherry Zhang <cherryyz@google.com>

src/cmd/link/internal/loader/loader.go

@@ -181,6 +181,7 @@ type Loader struct {
 	attrShared           bitmap // shared symbols, indexed by ext sym index
 	attrExternal         bitmap // external symbols, indexed by ext sym index
 
+	attrReadOnly         map[Sym]bool     // readonly data for this sym
 	attrTopFrame         map[Sym]struct{} // top frame symbols
 	attrSpecial          map[Sym]struct{} // "special" frame symbols
 	attrCgoExportDynamic map[Sym]struct{} // "cgo_export_dynamic" symbols
@@ -247,6 +248,7 @@ func NewLoader(flags uint32, elfsetstring elfsetstringFunc) *Loader {
 		dynimpvers:           make(map[Sym]string),
 		localentry:           make(map[Sym]uint8),
 		extname:              make(map[Sym]string),
+		attrReadOnly:         make(map[Sym]bool),
 		attrTopFrame:         make(map[Sym]struct{}),
 		attrSpecial:          make(map[Sym]struct{}),
 		attrCgoExportDynamic: make(map[Sym]struct{}),
@@ -878,6 +880,25 @@ func (l *Loader) SetAttrCgoExportStatic(i Sym, v bool) {
 	}
 }
 
+// AttrReadOnly returns true for a symbol whose underlying data
+// is stored via a read-only mmap.
+func (l *Loader) AttrReadOnly(i Sym) bool {
+	if v, ok := l.attrReadOnly[i]; ok {
+		return v
+	}
+	if i >= l.extStart {
+		return false
+	}
+	r, _ := l.toLocal(i)
+	return r.ReadOnly()
+}
+
+// SetAttrReadOnly sets the "cgo_export_dynamic" for a symbol
+// (see AttrReadOnly).
+func (l *Loader) SetAttrReadOnly(i Sym, v bool) {
+	l.attrReadOnly[i] = v
+}
+
 // AttrSubSymbol returns true for symbols that are listed as a
 // sub-symbol of some other outer symbol. The sub/outer mechanism is
 // used when loading host objects (sections from the host object
@@ -1466,31 +1487,80 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols) {
 		nr += loadObjSyms(l, syms, o.r)
 	}
 
-	// allocate a single large slab of relocations for all live symbols
+	// Make a first pass through the external symbols, making
-	l.relocBatch = make([]sym.Reloc, nr)
+	// sure that each external symbol has a non-nil entry in
-
+	// l.Syms (note that relocations and symbol content will
-	// external symbols
+	// be copied in a later loop).
+	toConvert := make([]Sym, 0, l.max-l.extStart+1)
 	for i := l.extStart; i <= l.max; i++ {
 		if s := l.Syms[i]; s != nil {
 			s.Attr.Set(sym.AttrReachable, l.attrReachable.has(i))
-			continue // already loaded from external object
+			continue
 		}
-		sname := l.payloads[i-l.extStart].name
+		sname := l.RawSymName(i)
-		sver := l.payloads[i-l.extStart].ver
+		if !l.attrReachable.has(i) && !strings.HasPrefix(sname, "gofile..") { // XXX file symbols are used but not marked
-		if l.attrReachable.has(i) || strings.HasPrefix(sname, "gofile..") { // XXX file symbols are used but not marked
+			continue
-			s := l.allocSym(sname, sver)
+		}
-			pp := l.getPayload(i)
+		pp := l.getPayload(i)
-			if pp != nil {
+		nr += len(pp.relocs)
-				if pp.kind != sym.Sxxx || len(pp.relocs) != 0 || len(pp.data) != 0 {
+		// create and install the sym.Symbol here so that l.Syms will
-					// Unpack payload into sym. Currently there is nothing
+		// be fully populated when we do relocation processing and
-					// to do here, but eventually we'll need a real
+		// outer/sub processing below.
-					// implementation.
+		s := l.allocSym(sname, 0)
-					panic("need to handle this")
+		l.installSym(i, s)
-				}
+		toConvert = append(toConvert, i)
-			}
+	}
-			preprocess(arch, s)
+
-			s.Attr.Set(sym.AttrReachable, l.attrReachable.has(i))
+	// allocate a single large slab of relocations for all live symbols
-			l.installSym(i, s)
+	l.relocBatch = make([]sym.Reloc, nr)
+
+	// convert payload-based external symbols into sym.Symbol-based
+	for _, i := range toConvert {
+
+		// Copy kind/size/value etc.
+		pp := &l.payloads[i-l.extStart]
+		s := l.Syms[i]
+		s.Version = int16(pp.ver)
+		s.Type = pp.kind
+		s.Size = pp.size
+		s.Value = l.SymValue(i)
+
+		// Copy relocations
+		batch := l.relocBatch
+		s.R = batch[:len(pp.relocs):len(pp.relocs)]
+		l.relocBatch = batch[len(pp.relocs):]
+		l.convertRelocations(pp.relocs, s)
+
+		// Copy data
+		s.P = pp.data
+
+		// Convert outer/sub relationships
+		if outer, ok := l.outer[i]; ok {
+			s.Outer = l.Syms[outer]
+		}
+		if sub, ok := l.sub[i]; ok {
+			s.Sub = l.Syms[sub]
+		}
+
+		// Preprocess symbol and set reachability and onlist.
+		preprocess(arch, s)
+		s.Attr.Set(sym.AttrReachable, l.attrReachable.has(i))
+		s.Attr.Set(sym.AttrOnList, l.attrOnList.has(i))
+
+		// Set sub-symbol attribute. FIXME: would be better
+		// to do away with this and just use l.OuterSymbol() != 0
+		// elsewhere within the linker.
+		s.Attr.Set(sym.AttrSubSymbol, s.Outer != nil)
+
+		// Copy over dynimplib, dynimpvers, extname.
+		if l.SymExtname(i) != "" {
+			s.SetExtname(l.SymExtname(i))
+		}
+		if l.SymDynimplib(i) != "" {
+			s.SetDynimplib(l.SymDynimplib(i))
+		}
+		if l.SymDynimpvers(i) != "" {
+			s.SetDynimpvers(l.SymDynimpvers(i))
 		}
 	}
 
@@ -1499,6 +1569,10 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols) {
 		loadObjFull(l, o.r)
 	}
 
+	// Note: resolution of ABI aliases is now also handled in
+	// loader.convertRelocations, so once the host object loaders move
+	// completely to loader.Sym, we can remove the code below.
+
 	// Resolve ABI aliases for external symbols. This is only
 	// needed for internal cgo linking.
 	// (The old code does this in deadcode, but deadcode2 doesn't
@@ -1574,14 +1648,6 @@ func (l *Loader) installSym(i Sym, s *sym.Symbol) {
 	if l.Syms[i] != nil {
 		panic("sym already present in installSym")
 	}
-	if l.IsExternal(i) {
-		// temporary sanity check: make sure that the payload
-		// is empty, e.g. nobody has added symbol content already.
-		pp := l.getPayload(i)
-		if pp != nil && (len(pp.relocs) != 0 || len(pp.data) != 0) {
-			panic("expected empty payload")
-		}
-	}
 	l.Syms[i] = s
 }
 
@@ -1713,6 +1779,87 @@ func (l *Loader) LookupOrCreate(name string, version int) *sym.Symbol {
 	return s
 }
 
+// cloneToExternal takes the existing object file symbol (symIdx)
+// and creates a new external symbol that is a clone with respect
+// to name, version, type, relocations, etc. The idea here is that
+// if the linker decides it wants to update the contents of a
+// symbol originally discovered as part of an object file, it's
+// easier to do this if we make the updates to a new and similarly
+// named external copy of that symbol.
+func (l *Loader) cloneToExternal(symIdx Sym) Sym {
+	if l.IsExternal(symIdx) {
+		panic("sym is already external, no need for clone")
+	}
+
+	// Read the particulars from object.
+	osym := goobj2.Sym{}
+	r, li := l.toLocal(symIdx)
+	osym.Read(r.Reader, r.SymOff(li))
+	sname := strings.Replace(osym.Name, "\"\".", r.pkgprefix, -1)
+	sver := abiToVer(osym.ABI, r.version)
+	skind := sym.AbiSymKindToSymKind[objabi.SymKind(osym.Type)]
+
+	// Create new symbol, update version and kind.
+	ns := l.newExtSym(sname, sver)
+	pp := &l.payloads[ns-l.extStart]
+	pp.kind = skind
+	pp.ver = sver
+	pp.size = int64(osym.Siz)
+
+	// If this is a def, then copy the guts. We expect this case
+	// to be very rare (one case it may come up is with -X).
+	if li < (r.NSym() + r.NNonpkgdef()) {
+
+		// Copy relocations
+		relocs := l.Relocs(symIdx)
+		pp.relocs = relocs.ReadAll(nil)
+
+		// Copy data
+		pp.data = r.Data(li)
+
+		// Copy read-only attr
+		if r.ReadOnly() {
+			l.attrReadOnly[ns] = true
+		}
+	}
+
+	// Fix up the lookup tables if the symbol in question was
+	// present in the lookup tables. At the moment it only makes
+	// sense to do this sort of clone/update for symbols that are
+	// in the symbol table (as opposed to anonymous symbols);
+	// issue an error if we can't look up the original symbol.
+	if sver >= sym.SymVerStatic {
+		s, ok := l.extStaticSyms[nameVer{sname, sver}]
+		if !ok || s != symIdx {
+			panic("lookup failed for clone of non-external static symbol")
+		}
+		l.extStaticSyms[nameVer{sname, sver}] = ns
+	} else {
+		s, ok := l.symsByName[sver][sname]
+		if !ok || s != symIdx {
+			panic("lookup failed for clone of non-external symbol")
+		}
+		l.symsByName[sver][sname] = ns
+	}
+
+	// Add an overwrite entry (in case there are relocations against
+	// the old symbol).
+	l.overwrite[symIdx] = ns
+
+	// There may be relocations against this symbol from other symbols
+	// in the object -- we want those relocations to target the new
+	// external sym version of this symbol, not the old overwritten
+	// one. Update the rcache accordingly.
+	if li > r.NSym() {
+		r.rcacheSet(uint32(li-r.NSym()), ns)
+	}
+
+	// FIXME: copy other attributes? reachable is the main one, and we
+	// don't expect it to be set at this point.
+
+	return ns
+}
+
 // CreateExtSym creates a new external symbol with the specified name
 // without adding it to any lookup tables, returning a Sym index for it.
 func (l *Loader) CreateExtSym(name string) Sym {
@@ -1813,35 +1960,7 @@ func loadObjFull(l *Loader, r *oReader) {
 		batch := l.relocBatch
 		s.R = batch[:relocs.Count:relocs.Count]
 		l.relocBatch = batch[relocs.Count:]
-		for j := range s.R {
+		l.convertRelocations(rslice, s)
-			r := rslice[j]
-			rs := r.Sym
-			sz := r.Size
-			rt := r.Type
-			if rt == objabi.R_METHODOFF {
-				if l.attrReachable.has(rs) {
-					rt = objabi.R_ADDROFF
-				} else {
-					sz = 0
-					rs = 0
-				}
-			}
-			if rt == objabi.R_WEAKADDROFF && !l.attrReachable.has(rs) {
-				rs = 0
-				sz = 0
-			}
-			if rs != 0 && l.Syms[rs] != nil && l.Syms[rs].Type == sym.SABIALIAS {
-				rsrelocs := l.Relocs(rs)
-				rs = rsrelocs.At(0).Sym
-			}
-			s.R[j] = sym.Reloc{
-				Off:  r.Off,
-				Siz:  sz,
-				Type: rt,
-				Add:  r.Add,
-				Sym:  l.Syms[rs],
-			}
-		}
 
 		// Aux symbol info
 		isym := -1
@@ -2015,6 +2134,43 @@ func loadObjFull(l *Loader, r *oReader) {
 	}
 }
 
+// convertRelocations takes a vector of loader.Reloc relocations and
+// translates them into an equivalent set of sym.Reloc relocations on
+// the symbol "dst", performing fixups along the way for ABI aliases,
+// etc. It is assumed that the called has pre-allocated the dst symbol
+// relocations slice.
+func (l *Loader) convertRelocations(src []Reloc, dst *sym.Symbol) {
+	for j := range dst.R {
+		r := src[j]
+		rs := r.Sym
+		sz := r.Size
+		rt := r.Type
+		if rt == objabi.R_METHODOFF {
+			if l.attrReachable.has(rs) {
+				rt = objabi.R_ADDROFF
+			} else {
+				sz = 0
+				rs = 0
+			}
+		}
+		if rt == objabi.R_WEAKADDROFF && !l.attrReachable.has(rs) {
+			rs = 0
+			sz = 0
+		}
+		if rs != 0 && l.Syms[rs] != nil && l.Syms[rs].Type == sym.SABIALIAS {
+			rsrelocs := l.Relocs(rs)
+			rs = rsrelocs.At(0).Sym
+		}
+		dst.R[j] = sym.Reloc{
+			Off:  r.Off,
+			Siz:  sz,
+			Type: rt,
+			Add:  r.Add,
+			Sym:  l.Syms[rs],
+		}
+	}
+}
+
 var emptyPkg = []byte(`"".`)
 
 func patchDWARFName1(p []byte, r *oReader) ([]byte, int) {

src/cmd/link/internal/loader/loader_test.go

@@ -59,9 +59,9 @@ func TestAddMaterializedSymbol(t *testing.T) {
 	}
 
 	// Grab symbol builder pointers
-	sb1 := ldr.MakeSymbolUpdater(es1)
+	sb1, es1 := ldr.MakeSymbolUpdater(es1)
-	sb2 := ldr.MakeSymbolUpdater(es2)
+	sb2, es2 := ldr.MakeSymbolUpdater(es2)
-	sb3 := ldr.MakeSymbolUpdater(es3)
+	sb3, es3 := ldr.MakeSymbolUpdater(es3)
 
 	// Check get/set symbol type
 	es3typ := sb3.Type()
@@ -108,8 +108,8 @@ func TestAddMaterializedSymbol(t *testing.T) {
 		}
 	}
 
-	sb1 = ldr.MakeSymbolUpdater(es1)
+	sb1, es1 = ldr.MakeSymbolUpdater(es1)
-	sb2 = ldr.MakeSymbolUpdater(es2)
+	sb2, es2 = ldr.MakeSymbolUpdater(es2)
 
 	// Get/set a few other attributes
 	if ldr.AttrVisibilityHidden(es3) {
@@ -217,7 +217,7 @@ func sameRelocSlice(s1 []Reloc, s2 []Reloc) bool {
 	return true
 }
 
-type addFunc func(l *Loader, s Sym, s2 Sym)
+type addFunc func(l *Loader, s Sym, s2 Sym) Sym
 
 func TestAddDataMethods(t *testing.T) {
 	edummy := func(s *sym.Symbol, str string, off int) {}
@@ -239,47 +239,52 @@ func TestAddDataMethods(t *testing.T) {
 	}{
 		{
 			which: "AddUint8",
-			addDataFunc: func(l *Loader, s Sym, _ Sym) {
+			addDataFunc: func(l *Loader, s Sym, _ Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.AddUint8('a')
+				return ns
 			},
 			expData: []byte{'a'},
 			expKind: sym.SDATA,
 		},
 		{
 			which: "AddUintXX",
-			addDataFunc: func(l *Loader, s Sym, _ Sym) {
+			addDataFunc: func(l *Loader, s Sym, _ Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.AddUintXX(arch, 25185, 2)
+				return ns
 			},
 			expData: []byte{'a', 'b'},
 			expKind: sym.SDATA,
 		},
 		{
 			which: "SetUint8",
-			addDataFunc: func(l *Loader, s Sym, _ Sym) {
+			addDataFunc: func(l *Loader, s Sym, _ Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.AddUint8('a')
 				sb.AddUint8('b')
 				sb.SetUint8(arch, 1, 'c')
+				return ns
 			},
 			expData: []byte{'a', 'c'},
 			expKind: sym.SDATA,
 		},
 		{
 			which: "AddString",
-			addDataFunc: func(l *Loader, s Sym, _ Sym) {
+			addDataFunc: func(l *Loader, s Sym, _ Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.Addstring("hello")
+				return ns
 			},
 			expData: []byte{'h', 'e', 'l', 'l', 'o', 0},
 			expKind: sym.SNOPTRDATA,
 		},
 		{
 			which: "AddAddrPlus",
-			addDataFunc: func(l *Loader, s Sym, s2 Sym) {
+			addDataFunc: func(l *Loader, s Sym, s2 Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.AddAddrPlus(arch, s2, 3)
+				return ns
 			},
 			expData: []byte{0, 0, 0, 0, 0, 0, 0, 0},
 			expKind: sym.SDATA,
@@ -287,9 +292,10 @@ func TestAddDataMethods(t *testing.T) {
 		},
 		{
 			which: "AddAddrPlus4",
-			addDataFunc: func(l *Loader, s Sym, s2 Sym) {
+			addDataFunc: func(l *Loader, s Sym, s2 Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.AddAddrPlus4(arch, s2, 3)
+				return ns
 			},
 			expData: []byte{0, 0, 0, 0},
 			expKind: sym.SDATA,
@@ -297,9 +303,10 @@ func TestAddDataMethods(t *testing.T) {
 		},
 		{
 			which: "AddCURelativeAddrPlus",
-			addDataFunc: func(l *Loader, s Sym, s2 Sym) {
+			addDataFunc: func(l *Loader, s Sym, s2 Sym) Sym {
-				sb := l.MakeSymbolUpdater(s)
+				sb, ns := l.MakeSymbolUpdater(s)
 				sb.AddCURelativeAddrPlus(arch, s2, 7)
+				return ns
 			},
 			expData: []byte{0, 0, 0, 0, 0, 0, 0, 0},
 			expKind: sym.SDATA,
@@ -314,7 +321,7 @@ func TestAddDataMethods(t *testing.T) {
 		if mi == 0 {
 			t.Fatalf("AddExtSym failed for '" + name + "'")
 		}
-		tp.addDataFunc(ldr, mi, pmi)
+		mi = tp.addDataFunc(ldr, mi, pmi)
 		if ldr.SymType(mi) != tp.expKind {
 			t.Errorf("testing Loader.%s: expected kind %s got %s",
 				tp.which, tp.expKind, ldr.SymType(mi))

src/cmd/link/internal/loader/symbolbuilder.go

@@ -18,7 +18,7 @@ type SymbolBuilder struct {
 	l              *Loader // loader
 }
 
-// NewSymbolBuilder creates a symbol builder for use in constructing
+// MakeSymbolBuilder creates a symbol builder for use in constructing
 // an entirely new symbol.
 func (l *Loader) MakeSymbolBuilder(name string) *SymbolBuilder {
 	// for now assume that any new sym is intended to be static
@@ -31,18 +31,31 @@ func (l *Loader) MakeSymbolBuilder(name string) *SymbolBuilder {
 	return sb
 }
 
-// NewSymbolBuilder creates a symbol builder helper for an already-allocated
+// MakeSymbolUpdater creates a symbol builder helper for an existing
-// external symbol 'symIdx'.
+// symbol 'symIdx'. If 'symIdx' is not an external symbol, then create
-func (l *Loader) MakeSymbolUpdater(symIdx Sym) *SymbolBuilder {
+// a clone of it (copy name, properties, etc) fix things up so that
+// the lookup tables and caches point to the new version, not the old
+// version. Returns a SymbolBuilder and a Sym (which may be different
+// from the original if we had to clone).
+func (l *Loader) MakeSymbolUpdater(symIdx Sym) (*SymbolBuilder, Sym) {
+	if symIdx == 0 {
+		panic("can't update the null symbol")
+	}
+	if ov, ok := l.overwrite[symIdx]; ok {
+		symIdx = ov
+	}
 	if !l.IsExternal(symIdx) {
-		panic("can't build on non-external sym")
+		// Create a clone with the same name/version/kind etc.
+		symIdx = l.cloneToExternal(symIdx)
 	}
 	if l.Syms[symIdx] != nil {
 		panic("can't build if sym.Symbol already present")
 	}
+
+	// Construct updater and return.
 	sb := &SymbolBuilder{l: l, symIdx: symIdx}
 	sb.extSymPayload = &l.payloads[symIdx-l.extStart]
-	return sb
+	return sb, symIdx
 }
 
 // Getters for properties of the symbol we're working on.