diff --git a/src/pkg/runtime/Makefile b/src/pkg/runtime/Makefile
index df46d0391c2..725c2b07e23 100644
--- a/src/pkg/runtime/Makefile
+++ b/src/pkg/runtime/Makefile
@@ -31,6 +31,7 @@ GOFILES=\
 CLEANFILES+=version.go version_*.go
 
 OFILES_windows=\
+	callback.$O\
 	syscall.$O\
 
 # 386-specific object files
diff --git a/src/pkg/runtime/syscall_windows_test.go b/src/pkg/runtime/syscall_windows_test.go
index aec85ec1676..c270607015a 100644
--- a/src/pkg/runtime/syscall_windows_test.go
+++ b/src/pkg/runtime/syscall_windows_test.go
@@ -59,3 +59,41 @@ func TestCDecl(t *testing.T) {
 		t.Error("cdecl USER32.wsprintfA returns", a, "buf=", buf[:a])
 	}
 }
+
+func TestCallback(t *testing.T) {
+	h, e := syscall.LoadLibrary("user32.dll")
+	if e != 0 {
+		t.Fatal("LoadLibrary(USER32)")
+	}
+	pEnumWindows, e := syscall.GetProcAddress(h, "EnumWindows")
+	if e != 0 {
+		t.Fatal("GetProcAddress(USER32.EnumWindows)")
+	}
+	pIsWindow, e := syscall.GetProcAddress(h, "IsWindow")
+	if e != 0 {
+		t.Fatal("GetProcAddress(USER32.IsWindow)")
+	}
+	counter := 0
+	cb := syscall.NewCallback(func(hwnd syscall.Handle, lparam uintptr) uintptr {
+		if lparam != 888 {
+			t.Error("lparam was not passed to callback")
+		}
+		b, _, _ := syscall.Syscall(uintptr(pIsWindow), 1, uintptr(hwnd), 0, 0)
+		if b == 0 {
+			t.Error("USER32.IsWindow returns FALSE")
+		}
+		counter++
+		return 1 // continue enumeration
+	})
+	a, _, _ := syscall.Syscall(uintptr(pEnumWindows), 2, cb, 888, 0)
+	if a == 0 {
+		t.Error("USER32.EnumWindows returns FALSE")
+	}
+	if counter == 0 {
+		t.Error("Callback has been never called or your have no windows")
+	}
+}
+
+func TestCallbackInAnotherThread(t *testing.T) {
+	// TODO: test a function which calls back in another thread: QueueUserAPC() or CreateThread()
+}
diff --git a/src/pkg/runtime/windows/386/callback.c b/src/pkg/runtime/windows/386/callback.c
new file mode 100644
index 00000000000..11b3d294d9a
--- /dev/null
+++ b/src/pkg/runtime/windows/386/callback.c
@@ -0,0 +1,107 @@
+// Copyright 2009 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.
+
+#include "runtime.h"
+#include "type.h"
+#include "defs.h"
+#include "os.h"
+
+// Will keep all callbacks in a linked list, so they don't get garbage collected.
+typedef	struct	Callback	Callback;
+struct	Callback {
+	Callback*	link;
+	void*		gobody;
+	byte		asmbody;
+};
+
+typedef	struct	Callbacks	Callbacks;
+struct	Callbacks {
+	Lock;
+	Callback*	link;
+	int32		n;
+};
+
+static	Callbacks	cbs;
+
+// Call back from windows dll into go.
+byte *
+runtime·compilecallback(Eface fn, bool cleanstack)
+{
+	FuncType *ft;
+	Type *t;
+	int32 argsize, i, n;
+	byte *p;
+	Callback *c;
+
+	if(fn.type == nil || fn.type->kind != KindFunc)
+		runtime·panicstring("compilecallback: not a function");
+	ft = (FuncType*)fn.type;
+	if(ft->out.len != 1)
+		runtime·panicstring("compilecallback: function must have one output parameter");
+	if(((Type**)ft->out.array)[0]->size != sizeof(uintptr))
+		runtime·panicstring("compilecallback: output parameter size is wrong");
+	argsize = 0;
+	for(i=0; i<ft->in.len; i++) {
+		t = ((Type**)ft->in.array)[i];
+		if(t->size > sizeof(uintptr))
+			runtime·panicstring("compilecallback: input parameter size is wrong");
+		argsize += sizeof(uintptr);
+	}
+
+	// compute size of new fn.
+	// must match code laid out below.
+	n = 1+4;		// MOVL fn, AX
+	n += 1+4;		// MOVL argsize, DX
+	n += 1+4;		// MOVL callbackasm, CX
+	n += 2;			// CALL CX
+	n += 1;			// RET
+	if(cleanstack && argsize!=0)
+		n += 2;		// ... argsize
+
+	runtime·lock(&cbs);
+	for(c = cbs.link; c != nil; c = c->link) {
+		if(c->gobody == fn.data) {
+			runtime·unlock(&cbs);
+			return &c->asmbody;
+		}
+	}
+	if(cbs.n >= 2000)
+		runtime·throw("too many callback functions");
+	c = runtime·mal(sizeof *c + n);
+	c->gobody = fn.data;
+	c->link = cbs.link;
+	cbs.link = c;
+	cbs.n++;
+	runtime·unlock(&cbs);
+
+	p = &c->asmbody;
+
+	// MOVL fn, AX
+	*p++ = 0xb8;
+	*(uint32*)p = (uint32)fn.data;
+	p += 4;
+
+	// MOVL argsize, DX
+	*p++ = 0xba;
+	*(uint32*)p = argsize;
+	p += 4;
+
+	// MOVL callbackasm, CX
+	*p++ = 0xb9;
+	*(uint32*)p = (uint32)runtime·callbackasm;
+	p += 4;
+
+	// CALL CX
+	*p++ = 0xff;
+	*p++ = 0xd1;
+
+	// RET argsize?
+	if(cleanstack && argsize!=0) {
+		*p++ = 0xc2;
+		*(uint16*)p = argsize;
+	} else
+		*p = 0xc3;
+
+	return &c->asmbody;
+}
diff --git a/src/pkg/runtime/windows/amd64/callback.c b/src/pkg/runtime/windows/amd64/callback.c
new file mode 100644
index 00000000000..d53822e2be9
--- /dev/null
+++ b/src/pkg/runtime/windows/amd64/callback.c
@@ -0,0 +1,104 @@
+// Copyright 2009 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.
+
+#include "runtime.h"
+#include "type.h"
+#include "defs.h"
+#include "os.h"
+
+// Will keep all callbacks in a linked list, so they don't get garbage collected.
+typedef	struct	Callback	Callback;
+struct	Callback {
+	Callback*	link;
+	void*		gobody;
+	byte		asmbody;
+};
+
+typedef	struct	Callbacks	Callbacks;
+struct	Callbacks {
+	Lock;
+	Callback*	link;
+	int32		n;
+};
+
+static	Callbacks	cbs;
+
+// Call back from windows dll into go.
+byte *
+runtime·compilecallback(Eface fn, bool /*cleanstack*/)
+{
+	FuncType *ft;
+	Type *t;
+	int32 argsize, i, n;
+	byte *p;
+	Callback *c;
+
+	if(fn.type == nil || fn.type->kind != KindFunc)
+		runtime·panicstring("compilecallback: not a function");
+	ft = (FuncType*)fn.type;
+	if(ft->out.len != 1)
+		runtime·panicstring("compilecallback: function must have one output parameter");
+	if(((Type**)ft->out.array)[0]->size != sizeof(uintptr))
+		runtime·panicstring("compilecallback: output parameter size is wrong");
+	argsize = 0;
+	for(i=0; i<ft->in.len; i++) {
+		t = ((Type**)ft->in.array)[i];
+		if(t->size > sizeof(uintptr))
+			runtime·panicstring("compilecallback: input parameter size is wrong");
+		argsize += sizeof(uintptr);
+	}
+
+	// compute size of new fn.
+	// must match code laid out below.
+	n  = 2+8+1; // MOVQ fn, AX           / PUSHQ AX
+	n += 2+8+1; // MOVQ argsize, AX      / PUSHQ AX
+	n += 2+8;   // MOVQ callbackasm, AX
+	n += 2;     // JMP  AX
+
+	runtime·lock(&cbs);
+	for(c = cbs.link; c != nil; c = c->link) {
+		if(c->gobody == fn.data) {
+			runtime·unlock(&cbs);
+			return &c->asmbody;
+		}
+	}
+	if(cbs.n >= 2000)
+		runtime·throw("too many callback functions");
+	c = runtime·mal(sizeof *c + n);
+	c->gobody = fn.data;
+	c->link = cbs.link;
+	cbs.link = c;
+	cbs.n++;
+	runtime·unlock(&cbs);
+
+	p = &c->asmbody;
+
+	// MOVQ fn, AX
+	*p++ = 0x48;
+	*p++ = 0xb8;
+	*(uint64*)p = (uint64)fn.data;
+	p += 8;
+	// PUSH AX
+	*p++ = 0x50;
+
+	// MOVQ argsize, AX
+	*p++ = 0x48;
+	*p++ = 0xb8;
+	*(uint64*)p = argsize;
+	p += 8;
+	// PUSH AX
+	*p++ = 0x50;
+
+	// MOVQ callbackasm, AX
+	*p++ = 0x48;
+	*p++ = 0xb8;
+	*(uint64*)p = (uint64)runtime·callbackasm;
+	p += 8;
+
+	// JMP AX
+	*p++ = 0xFF;
+	*p++ = 0xE0;
+
+	return &c->asmbody;
+}
diff --git a/src/pkg/runtime/windows/amd64/sys.s b/src/pkg/runtime/windows/amd64/sys.s
index 81659228e2d..eb197d72ef2 100644
--- a/src/pkg/runtime/windows/amd64/sys.s
+++ b/src/pkg/runtime/windows/amd64/sys.s
@@ -106,9 +106,80 @@ TEXT runtime·ctrlhandler(SB),7,$0
 	POPQ	BX
 	POPQ	BP
 	RET
-	
+
+// Continuation of thunk function created for each callback by ../thread.c compilecallback,
+// runs on Windows stack (not Go stack).
+// Thunk code designed to have minimal size for it is copied many (up to thousands) times.
+//
+// thunk:
+//	MOVQ	$fn, AX
+//	PUSHQ	AX
+//	MOVQ	$argsize, AX
+//	PUSHQ	AX
+//	MOVQ	$runtime·callbackasm, AX
+//	JMP	AX
 TEXT runtime·callbackasm(SB),7,$0
-	// TODO
+	// Construct args vector for cgocallback().
+	// By windows/amd64 calling convention first 4 args are in CX, DX, R8, R9
+	// args from the 5th on are on the stack.
+	// In any case, even if function has 0,1,2,3,4 args, there is reserved
+	// but uninitialized "shadow space" for the first 4 args.
+	// The values are in registers.
+  	MOVQ	CX, (24+0)(SP)
+  	MOVQ	DX, (24+8)(SP)
+  	MOVQ	R8, (24+16)(SP)
+  	MOVQ	R9, (24+24)(SP)
+	// 6l does not accept writing POPQs here issuing a warning "unbalanced PUSH/POP"
+  	MOVQ	0(SP), DX	// POPQ DX
+  	MOVQ	8(SP), AX	// POPQ AX
+	ADDQ	$16, SP
+
+	// preserve whatever's at the memory location that
+	// the callback will use to store the return value
+	LEAQ	8(SP), CX       // args vector, skip return address
+	PUSHQ	0(CX)(DX*1)     // store 8 bytes from just after the args array
+	ADDQ	$8, DX          // extend argsize by size of return value
+
+	// DI SI BP BX R12 R13 R14 R15 registers and DF flag are preserved
+	// as required by windows callback convention.
+	// 6l does not allow writing many PUSHQs here issuing a warning "nosplit stack overflow"
+	// the warning has no sense as this code uses os thread stack
+	PUSHFQ
+	SUBQ	$64, SP
+	MOVQ	DI, 56(SP)
+	MOVQ	SI, 48(SP)
+	MOVQ	BP, 40(SP)
+	MOVQ	BX, 32(SP)
+	MOVQ	R12, 24(SP)
+	MOVQ	R13, 16(SP)
+	MOVQ	R14, 8(SP)
+	MOVQ	R15, 0(SP)
+
+	// cgocallback(void (*fn)(void*), void *frame, uintptr framesize)
+	PUSHQ	DX    // uintptr framesize
+	PUSHQ	CX    // void *frame
+	PUSHQ	AX    // void (*fn)(void*)
+	CLD
+	CALL  runtime·cgocallback(SB)
+	POPQ	AX
+	POPQ	CX
+	POPQ	DX
+
+	// restore registers as required for windows callback
+	// 6l does not allow writing many POPs here issuing a warning "nosplit stack overflow"
+	MOVQ	0(SP), R15
+	MOVQ	8(SP), R14
+	MOVQ	16(SP), R13
+	MOVQ	24(SP), R12
+	MOVQ	32(SP), BX
+	MOVQ	40(SP), BP
+	MOVQ	48(SP), SI
+	MOVQ	56(SP), DI
+	ADDQ	$64, SP
+	POPFQ
+
+	MOVL	-8(CX)(DX*1), AX  // return value
+	POPQ	-8(CX)(DX*1)      // restore bytes just after the args
 	RET
 
 // uint32 tstart_stdcall(M *newm);
diff --git a/src/pkg/runtime/windows/thread.c b/src/pkg/runtime/windows/thread.c
index fbcbf871e6b..b76eaac596a 100644
--- a/src/pkg/runtime/windows/thread.c
+++ b/src/pkg/runtime/windows/thread.c
@@ -320,105 +320,6 @@ runtime·ctrlhandler1(uint32 type)
 	return 0;
 }
 
-// Will keep all callbacks in a linked list, so they don't get garbage collected.
-typedef	struct	Callback	Callback;
-struct	Callback {
-	Callback*	link;
-	void*		gobody;
-	byte		asmbody;
-};
-
-typedef	struct	Callbacks	Callbacks;
-struct	Callbacks {
-	Lock;
-	Callback*	link;
-	int32		n;
-};
-
-static	Callbacks	cbs;
-
-// Call back from windows dll into go.
-byte *
-runtime·compilecallback(Eface fn, bool cleanstack)
-{
-	FuncType *ft;
-	Type *t;
-	int32 argsize, i, n;
-	byte *p;
-	Callback *c;
-
-	if(fn.type == nil || fn.type->kind != KindFunc)
-		runtime·panicstring("compilecallback: not a function");
-	ft = (FuncType*)fn.type;
-	if(ft->out.len != 1)
-		runtime·panicstring("compilecallback: function must have one output parameter");
-	if(((Type**)ft->out.array)[0]->size != sizeof(uintptr))
-		runtime·panicstring("compilecallback: output parameter size is wrong");
-	argsize = 0;
-	for(i=0; i<ft->in.len; i++) {
-		t = ((Type**)ft->in.array)[i];
-		if(t->size != sizeof(uintptr))
-			runtime·panicstring("compilecallback: input parameter size is wrong");
-		argsize += t->size;
-	}
-
-	// compute size of new fn.
-	// must match code laid out below.
-	n = 1+4;		// MOVL fn, AX
-	n += 1+4;		// MOVL argsize, DX
-	n += 1+4;		// MOVL callbackasm, CX
-	n += 2;			// CALL CX
-	n += 1;			// RET
-	if(cleanstack)
-		n += 2;		// ... argsize
-
-	runtime·lock(&cbs);
-	for(c = cbs.link; c != nil; c = c->link) {
-		if(c->gobody == fn.data) {
-			runtime·unlock(&cbs);
-			return &c->asmbody;
-		}
-	}
-	if(cbs.n >= 2000)
-		runtime·throw("too many callback functions");
-	c = runtime·mal(sizeof *c + n);
-	c->gobody = fn.data;
-	c->link = cbs.link;
-	cbs.link = c;
-	cbs.n++;
-	runtime·unlock(&cbs);
-
-	p = &c->asmbody;
-
-	// MOVL fn, AX
-	*p++ = 0xb8;
-	*(uint32*)p = (uint32)fn.data;
-	p += 4;
-
-	// MOVL argsize, DX
-	*p++ = 0xba;
-	*(uint32*)p = argsize;
-	p += 4;
-
-	// MOVL callbackasm, CX
-	*p++ = 0xb9;
-	*(uint32*)p = (uint32)runtime·callbackasm;
-	p += 4;
-
-	// CALL CX
-	*p++ = 0xff;
-	*p++ = 0xd1;
-
-	// RET argsize?
-	if(cleanstack) {
-		*p++ = 0xc2;
-		*(uint16*)p = argsize;
-	} else
-		*p = 0xc3;
-
-	return &c->asmbody;
-}
-
 void
 os·sigpipe(void)
 {