cmd/compile: add -d=checkptr to validate unsafe.Pointer rules

This CL adds -d=checkptr as a compile-time option for adding
instrumentation to check that Go code is following unsafe.Pointer
safety rules dynamically. In particular, it currently checks two
things:

1. When converting unsafe.Pointer to *T, make sure the resulting
pointer is aligned appropriately for T.

2. When performing pointer arithmetic, if the result points to a Go
heap object, make sure we can find an unsafe.Pointer-typed operand
that pointed into the same object.

These checks are currently disabled for the runtime, and can also be
disabled through a new //go:nocheckptr annotation. The latter is
necessary for functions like strings.noescape, which intentionally
violate safety rules to workaround escape analysis limitations.

Fixes #22218.

Change-Id: If5a51273881d93048f74bcff10a3275c9c91da6a
Reviewed-on: https://go-review.googlesource.com/c/go/+/162237
Run-TryBot: Matthew Dempsky <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: Keith Randall <[email protected]>

Author: mdempsky

src/cmd/compile/internal/gc/builtin.go

@@ -235,6 +235,9 @@ func racewriterange(addr, size uintptr)
 func msanread(addr, size uintptr)
 func msanwrite(addr, size uintptr)
 
+func checkptrAlignment(unsafe.Pointer, *byte)
+func checkptrArithmetic(unsafe.Pointer, []unsafe.Pointer)
+
 // architecture variants
 var x86HasPOPCNT bool
 var x86HasSSE41 bool

src/cmd/compile/internal/gc/builtin/runtime.go

@@ -135,6 +135,12 @@ func caninl(fn *Node) {
 		return
 	}
 
+	// If marked "go:nocheckptr" and -d checkptr compilation, don't inline.
+	if Debug_checkptr != 0 && fn.Func.Pragma&NoCheckPtr != 0 {
+		reason = "marked go:nocheckptr"
+		return
+	}
+
 	// If marked "go:cgo_unsafe_args", don't inline, since the
 	// function makes assumptions about its argument frame layout.
 	if fn.Func.Pragma&CgoUnsafeArgs != 0 {

src/cmd/compile/internal/gc/inl.go

@@ -35,6 +35,7 @@ const (
 	Norace                       // func must not have race detector annotations
 	Nosplit                      // func should not execute on separate stack
 	Noinline                     // func should not be inlined
+	NoCheckPtr                   // func should not be instrumented by checkptr
 	CgoUnsafeArgs                // treat a pointer to one arg as a pointer to them all
 	UintptrEscapes               // pointers converted to uintptr escape
 
@@ -63,6 +64,8 @@ func pragmaValue(verb string) syntax.Pragma {
 		return Nosplit
 	case "go:noinline":
 		return Noinline
+	case "go:nocheckptr":
+		return NoCheckPtr
 	case "go:systemstack":
 		return Systemstack
 	case "go:nowritebarrier":

src/cmd/compile/internal/gc/lex.go

@@ -40,6 +40,7 @@ var (
 
 var (
 	Debug_append       int
+	Debug_checkptr     int
 	Debug_closure      int
 	Debug_compilelater int
 	debug_dclstack     int
@@ -65,6 +66,7 @@ var debugtab = []struct {
 	val  interface{} // must be *int or *string
 }{
 	{"append", "print information about append compilation", &Debug_append},
+	{"checkptr", "instrument unsafe pointer conversions", &Debug_checkptr},
 	{"closure", "print information about closure compilation", &Debug_closure},
 	{"compilelater", "compile functions as late as possible", &Debug_compilelater},
 	{"disablenil", "disable nil checks", &disable_checknil},
@@ -433,6 +435,11 @@ func Main(archInit func(*Arch)) {
 		}
 	}
 
+	// Runtime can't use -d=checkptr, at least not yet.
+	if compiling_runtime {
+		Debug_checkptr = 0
+	}
+
 	// set via a -d flag
 	Ctxt.Debugpcln = Debug_pctab
 	if flagDWARF {

src/cmd/compile/internal/gc/main.go

@@ -951,6 +951,16 @@ opswitch:
 
 	case OCONV, OCONVNOP:
 		n.Left = walkexpr(n.Left, init)
+		if n.Op == OCONVNOP && Debug_checkptr != 0 && Curfn.Func.Pragma&NoCheckPtr == 0 {
+			if n.Type.IsPtr() && n.Left.Type.Etype == TUNSAFEPTR { // unsafe.Pointer to *T
+				n = walkCheckPtrAlignment(n, init)
+				break
+			}
+			if n.Type.Etype == TUNSAFEPTR && n.Left.Type.Etype == TUINTPTR { // uintptr to unsafe.Pointer
+				n = walkCheckPtrArithmetic(n, init)
+				break
+			}
+		}
 		param, result := rtconvfn(n.Left.Type, n.Type)
 		if param == Txxx {
 			break
@@ -3898,3 +3908,66 @@ func canMergeLoads() bool {
 func isRuneCount(n *Node) bool {
 	return Debug['N'] == 0 && !instrumenting && n.Op == OLEN && n.Left.Op == OSTR2RUNES
 }
+
+func walkCheckPtrAlignment(n *Node, init *Nodes) *Node {
+	if n.Type.Elem().Alignment() == 1 {
+		return n
+	}
+
+	n.Left = cheapexpr(n.Left, init)
+	init.Append(mkcall("checkptrAlignment", nil, init, n.Left, typename(n.Type.Elem())))
+	return n
+}
+
+var walkCheckPtrArithmeticMarker byte
+
+func walkCheckPtrArithmetic(n *Node, init *Nodes) *Node {
+	// Calling cheapexpr(n, init) below leads to a recursive call
+	// to walkexpr, which leads us back here again. Use n.Opt to
+	// prevent infinite loops.
+	if n.Opt() == &walkCheckPtrArithmeticMarker {
+		return n
+	}
+	n.SetOpt(&walkCheckPtrArithmeticMarker)
+	defer n.SetOpt(nil)
+
+	// TODO(mdempsky): Make stricter. We only need to exempt
+	// reflect.Value.Pointer and reflect.Value.UnsafeAddr.
+	switch n.Left.Op {
+	case OCALLFUNC, OCALLMETH, OCALLINTER:
+		return n
+	}
+
+	// Find original unsafe.Pointer operands involved in this
+	// arithmetic expression.
+	//
+	// "It is valid both to add and to subtract offsets from a
+	// pointer in this way. It is also valid to use &^ to round
+	// pointers, usually for alignment."
+	var originals []*Node
+	var walk func(n *Node)
+	walk = func(n *Node) {
+		switch n.Op {
+		case OADD:
+			walk(n.Left)
+			walk(n.Right)
+		case OSUB, OANDNOT:
+			walk(n.Left)
+		case OCONVNOP:
+			if n.Left.Type.Etype == TUNSAFEPTR {
+				n.Left = cheapexpr(n.Left, init)
+				originals = append(originals, n.Left)
+			}
+		}
+	}
+	walk(n.Left)
+
+	n = cheapexpr(n, init)
+
+	slice := mkdotargslice(types.NewSlice(types.Types[TUNSAFEPTR]), originals, init, nil)
+	slice.Esc = EscNone
+	slice.SetTransient(true)
+
+	init.Append(mkcall("checkptrArithmetic", nil, init, n, slice))
+	return n
+}

src/cmd/compile/internal/gc/walk.go

@@ -1407,6 +1407,11 @@ func (v Value) OverflowUint(x uint64) bool {
 	panic(&ValueError{"reflect.Value.OverflowUint", v.kind()})
 }
 
+//go:nocheckptr
+// This prevents inlining Value.Pointer when -d=checkptr is enabled,
+// which ensures cmd/compile can recognize unsafe.Pointer(v.Pointer())
+// and make an exception.
+
 // Pointer returns v's value as a uintptr.
 // It returns uintptr instead of unsafe.Pointer so that
 // code using reflect cannot obtain unsafe.Pointers
@@ -1914,6 +1919,11 @@ func (v Value) Uint() uint64 {
 	panic(&ValueError{"reflect.Value.Uint", v.kind()})
 }
 
+//go:nocheckptr
+// This prevents inlining Value.UnsafeAddr when -d=checkptr is enabled,
+// which ensures cmd/compile can recognize unsafe.Pointer(v.UnsafeAddr())
+// and make an exception.
+
 // UnsafeAddr returns a pointer to v's data.
 // It is for advanced clients that also import the "unsafe" package.
 // It panics if v is not addressable.

src/reflect/value.go

@@ -0,0 +1,50 @@
+// Copyright 2019 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package runtime
+
+import "unsafe"
+
+type ptrAlign struct {
+	ptr   unsafe.Pointer
+	align uintptr
+}
+
+func checkptrAlignment(p unsafe.Pointer, elem *_type) {
+	// TODO(mdempsky): What about fieldAlign?
+	if uintptr(p)&(uintptr(elem.align)-1) != 0 {
+		panic(ptrAlign{p, uintptr(elem.align)})
+	}
+}
+
+type ptrArith struct {
+	ptr       unsafe.Pointer
+	originals []unsafe.Pointer
+}
+
+func checkptrArithmetic(p unsafe.Pointer, originals []unsafe.Pointer) {
+	if 0 < uintptr(p) && uintptr(p) < minLegalPointer {
+		panic(ptrArith{p, originals})
+	}
+
+	base := checkptrBase(p)
+	if base == 0 {
+		return
+	}
+
+	for _, original := range originals {
+		if base == checkptrBase(original) {
+			return
+		}
+	}
+
+	panic(ptrArith{p, originals})
+}
+
+func checkptrBase(p unsafe.Pointer) uintptr {
+	base, _, _ := findObject(uintptr(p), 0, 0)
+	// TODO(mdempsky): If base == 0, then check if p points to the
+	// stack or a global variable.
+	return base
+}

src/runtime/checkptr.go

@@ -24,6 +24,7 @@ type Builder struct {
 // USE CAREFULLY!
 // This was copied from the runtime; see issues 23382 and 7921.
 //go:nosplit
+//go:nocheckptr
 func noescape(p unsafe.Pointer) unsafe.Pointer {
 	x := uintptr(p)
 	return unsafe.Pointer(x ^ 0)