diff --git a/src/runtime/export_linux_test.go b/src/runtime/export_linux_test.go
new file mode 100644
index 0000000..c8b9746
--- /dev/null
+++ b/src/runtime/export_linux_test.go
@@ -0,0 +1,9 @@
+// Copyright 2015 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.
+
+// Export guts for testing.
+
+package runtime
+
+var NewOSProc0 = newosproc0
diff --git a/src/runtime/export_test.go b/src/runtime/export_test.go
index 905218b..a2b098d 100644
--- a/src/runtime/export_test.go
+++ b/src/runtime/export_test.go
@@ -22,6 +22,8 @@
 var Exitsyscall = exitsyscall
 var LockedOSThread = lockedOSThread
 
+var FuncPC = funcPC
+
 type LFNode struct {
 	Next    uint64
 	Pushcnt uintptr
diff --git a/src/runtime/norace_linux_test.go b/src/runtime/norace_linux_test.go
new file mode 100644
index 0000000..3f6d4e7
--- /dev/null
+++ b/src/runtime/norace_linux_test.go
@@ -0,0 +1,37 @@
+// Copyright 2015 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.
+
+// The file contains tests that can not run under race detector for some reason.
+// +build !race
+
+package runtime_test
+
+import (
+	"runtime"
+	"testing"
+	"time"
+	"unsafe"
+)
+
+var newOSProcDone bool
+
+//go:nosplit
+func newOSProcCreated() {
+	newOSProcDone = true
+}
+
+func TestNewOSProc0(t *testing.T) {
+	runtime.NewOSProc0(0x800000, unsafe.Pointer(runtime.FuncPC(newOSProcCreated)))
+	check, end := time.Tick(1*time.Second), time.Tick(5*time.Second)
+	for {
+		select {
+		case <-check:
+			if newOSProcDone {
+				return
+			}
+		case <-end:
+			t.Fatalf("couldn't create new OS process")
+		}
+	}
+}
diff --git a/src/runtime/os1_linux.go b/src/runtime/os1_linux.go
index 4214fa7..a286dcd 100644
--- a/src/runtime/os1_linux.go
+++ b/src/runtime/os1_linux.go
@@ -143,16 +143,16 @@
 	}
 }
 
-// Version of newosproc that doesn't require any Go structs to be allocated.
+// Version of newosproc that doesn't require a valid G.
 //go:nosplit
-func newosproc0(stacksize uintptr, fn unsafe.Pointer, fnarg unsafe.Pointer) {
+func newosproc0(stacksize uintptr, fn unsafe.Pointer) {
 	var dummy uint64
 	stack := sysAlloc(stacksize, &dummy)
 	if stack == nil {
 		write(2, unsafe.Pointer(&failallocatestack[0]), int32(len(failallocatestack)))
 		exit(1)
 	}
-	ret := clone0(cloneFlags, unsafe.Pointer(uintptr(stack)+stacksize), fn, fnarg)
+	ret := clone(cloneFlags, unsafe.Pointer(uintptr(stack)+stacksize), nil, nil, fn)
 	if ret < 0 {
 		write(2, unsafe.Pointer(&failthreadcreate[0]), int32(len(failthreadcreate)))
 		exit(1)
diff --git a/src/runtime/os_linux.go b/src/runtime/os_linux.go
index 8e4c05d..abea5d6 100644
--- a/src/runtime/os_linux.go
+++ b/src/runtime/os_linux.go
@@ -13,9 +13,6 @@
 func clone(flags int32, stk, mm, gg, fn unsafe.Pointer) int32
 
 //go:noescape
-func clone0(flags int32, stk, fn, fnarg unsafe.Pointer) int32
-
-//go:noescape
 func rt_sigaction(sig uintptr, new, old *sigactiont, size uintptr) int32
 
 //go:noescape
diff --git a/src/runtime/rt0_linux_386.s b/src/runtime/rt0_linux_386.s
index f98642b..633e806 100644
--- a/src/runtime/rt0_linux_386.s
+++ b/src/runtime/rt0_linux_386.s
@@ -42,7 +42,6 @@
 	MOVL	$0x800000, 0(SP)                    // stacksize = 8192KB
 	MOVL	$_rt0_386_linux_lib_go(SB), AX
 	MOVL	AX, 4(SP)                           // fn
-	MOVL	$0, 8(SP)                           // fnarg
 	MOVL	$runtime·newosproc0(SB), AX
 	CALL	AX
 
diff --git a/src/runtime/rt0_linux_amd64.s b/src/runtime/rt0_linux_amd64.s
index ee1dbc6..cd7c55e 100644
--- a/src/runtime/rt0_linux_amd64.s
+++ b/src/runtime/rt0_linux_amd64.s
@@ -12,13 +12,13 @@
 
 // When building with -buildmode=c-shared, this symbol is called when the shared
 // library is loaded.
-TEXT _rt0_amd64_linux_lib(SB),NOSPLIT,$0x58
-	MOVQ	BX, 0x18(SP)
-	MOVQ	BP, 0x20(SP)
-	MOVQ	R12, 0x28(SP)
-	MOVQ	R13, 0x30(SP)
-	MOVQ	R14, 0x38(SP)
-	MOVQ	R15, 0x40(SP)
+TEXT _rt0_amd64_linux_lib(SB),NOSPLIT,$0x40
+	MOVQ	BX, 0x10(SP)
+	MOVQ	BP, 0x18(SP)
+	MOVQ	R12, 0x20(SP)
+	MOVQ	R13, 0x28(SP)
+	MOVQ	R14, 0x30(SP)
+	MOVQ	R15, 0x38(SP)
 
 	MOVQ	DI, _rt0_amd64_linux_lib_argc<>(SB)
 	MOVQ	SI, _rt0_amd64_linux_lib_argv<>(SB)
@@ -36,17 +36,16 @@
 	MOVQ	$8388608, 0(SP)                    // stacksize
 	MOVQ	$_rt0_amd64_linux_lib_go(SB), AX
 	MOVQ	AX, 8(SP)                          // fn
-	MOVQ	$0, 0x10(SP)                       // fnarg
 	MOVQ	$runtime·newosproc0(SB), AX
 	CALL	AX
 
 restore:
-	MOVQ	0x18(SP), BX
-	MOVQ	0x20(SP), BP
-	MOVQ	0x28(SP), R12
-	MOVQ	0x30(SP), R13
-	MOVQ	0x38(SP), R14
-	MOVQ	0x40(SP), R15
+	MOVQ	0x10(SP), BX
+	MOVQ	0x18(SP), BP
+	MOVQ	0x20(SP), R12
+	MOVQ	0x28(SP), R13
+	MOVQ	0x30(SP), R14
+	MOVQ	0x38(SP), R15
 	RET
 
 TEXT _rt0_amd64_linux_lib_go(SB),NOSPLIT,$0
diff --git a/src/runtime/rt0_linux_arm.s b/src/runtime/rt0_linux_arm.s
index 878a6dd..b71a3f9 100644
--- a/src/runtime/rt0_linux_arm.s
+++ b/src/runtime/rt0_linux_arm.s
@@ -12,15 +12,15 @@
 
 // When building with -buildmode=c-shared, this symbol is called when the shared
 // library is loaded.
-TEXT _rt0_arm_linux_lib(SB),NOSPLIT,$40
+TEXT _rt0_arm_linux_lib(SB),NOSPLIT,$32
 	// Preserve callee-save registers.  Raspberry Pi's dlopen(), for example,
 	// actually cares that R11 is preserved.
-	MOVW	R4, 16(R13)
-	MOVW	R5, 20(R13)
-	MOVW	R6, 24(R13)
-	MOVW	R7, 28(R13)
-	MOVW	R8, 32(R13)
-	MOVW	R11, 36(R13)
+	MOVW	R4, 12(R13)
+	MOVW	R5, 16(R13)
+	MOVW	R6, 20(R13)
+	MOVW	R7, 24(R13)
+	MOVW	R8, 28(R13)
+	MOVW	R11, 32(R13)
 
 	// Save argc/argv.
 	MOVW	R0, _rt0_arm_linux_lib_argc<>(SB)
@@ -37,19 +37,17 @@
 nocgo:
 	MOVW	$0x800000, R0                     // stacksize = 8192KB
 	MOVW	$_rt0_arm_linux_lib_go<>(SB), R1  // fn
-	MOVW	$0, R2                            // fnarg
 	MOVW	R0, 4(R13)
 	MOVW	R1, 8(R13)
-	MOVW	R2, 12(R13)
 	BL	runtime·newosproc0(SB)
 rr:
 	// Restore callee-save registers and return.
-	MOVW	16(R13), R4
-	MOVW	20(R13), R5
-	MOVW	24(R13), R6
-	MOVW	28(R13), R7
-	MOVW	32(R13), R8
-	MOVW	36(R13), R11
+	MOVW	12(R13), R4
+	MOVW	16(R13), R5
+	MOVW	20(R13), R6
+	MOVW	24(R13), R7
+	MOVW	28(R13), R8
+	MOVW	32(R13), R11
 	RET
 
 TEXT _rt0_arm_linux_lib_go<>(SB),NOSPLIT,$8
diff --git a/src/runtime/sys_linux_386.s b/src/runtime/sys_linux_386.s
index d69054f..679a81d 100644
--- a/src/runtime/sys_linux_386.s
+++ b/src/runtime/sys_linux_386.s
@@ -291,18 +291,18 @@
 // int32 clone(int32 flags, void *stack, M *mp, G *gp, void (*fn)(void));
 TEXT runtime·clone(SB),NOSPLIT,$0
 	MOVL	$120, AX	// clone
-	MOVL	flags+4(SP), BX
-	MOVL	stack+8(SP), CX
+	MOVL	flags+0(FP), BX
+	MOVL	stack+4(FP), CX
 	MOVL	$0, DX	// parent tid ptr
 	MOVL	$0, DI	// child tid ptr
 
 	// Copy mp, gp, fn off parent stack for use by child.
 	SUBL	$16, CX
-	MOVL	mm+12(SP), SI
+	MOVL	mm+8(FP), SI
 	MOVL	SI, 0(CX)
-	MOVL	gg+16(SP), SI
+	MOVL	gg+12(FP), SI
 	MOVL	SI, 4(CX)
-	MOVL	fn+20(SP), SI
+	MOVL	fn+16(FP), SI
 	MOVL	SI, 8(CX)
 	MOVL	$1234, 12(CX)
 
@@ -319,7 +319,7 @@
 	RET
 
 	// Paranoia: check that SP is as we expect.
-	MOVL	mm+8(FP), BP
+	MOVL	12(SP), BP
 	CMPL	BP, $1234
 	JEQ	2(PC)
 	INT	$3
@@ -328,10 +328,14 @@
 	MOVL	$224, AX
 	CALL	*runtime·_vdso(SB)
 
-	// In child on new stack.  Reload registers (paranoia).
-	MOVL	0(SP), BX	// m
-	MOVL	flags+0(FP), DX	// g
-	MOVL	stk+4(FP), SI	// fn
+	MOVL	0(SP), BX	    // m
+	MOVL	4(SP), DX	    // g
+	MOVL	8(SP), SI	    // fn
+
+	CMPL	BX, $0
+	JEQ	nog
+	CMPL	DX, $0
+	JEQ	nog
 
 	MOVL	AX, m_procid(BX)	// save tid as m->procid
 
@@ -365,16 +369,11 @@
 	CALL	runtime·emptyfunc(SB)
 	POPAL
 
+nog:
 	CALL	SI	// fn()
 	CALL	runtime·exit1(SB)
 	MOVL	$0x1234, 0x1005
 
-// int32 clone0(int32 flags, void *stack, void* fn, void* fnarg);
-TEXT runtime·clone0(SB),NOSPLIT,$0
-	// TODO(spetrovic): Implement this method.
-	MOVL	$-1, ret+16(FP)
-	RET
-
 TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 	MOVL	$186, AX	// sigaltstack
 	MOVL	new+4(SP), BX
diff --git a/src/runtime/sys_linux_amd64.s b/src/runtime/sys_linux_amd64.s
index 43a65b7..3a0c47f 100644
--- a/src/runtime/sys_linux_amd64.s
+++ b/src/runtime/sys_linux_amd64.s
@@ -302,14 +302,16 @@
 
 // int32 clone(int32 flags, void *stack, M *mp, G *gp, void (*fn)(void));
 TEXT runtime·clone(SB),NOSPLIT,$0
-	MOVL	flags+8(SP), DI
-	MOVQ	stack+16(SP), SI
+	MOVL	flags+0(FP), DI
+	MOVQ	stack+8(FP), SI
+	MOVQ	$0, DX
+	MOVQ	$0, R10
 
 	// Copy mp, gp, fn off parent stack for use by child.
 	// Careful: Linux system call clobbers CX and R11.
-	MOVQ	mm+24(SP), R8
-	MOVQ	gg+32(SP), R9
-	MOVQ	fn+40(SP), R12
+	MOVQ	mp+16(FP), R8
+	MOVQ	gp+24(FP), R9
+	MOVQ	fn+32(FP), R12
 
 	MOVL	$56, AX
 	SYSCALL
@@ -323,6 +325,12 @@
 	// In child, on new stack.
 	MOVQ	SI, SP
 
+	// If g or m are nil, skip Go-related setup.
+	CMPQ	R8, $0    // m
+	JEQ	nog
+	CMPQ	R9, $0    // g
+	JEQ	nog
+
 	// Initialize m->procid to Linux tid
 	MOVL	$186, AX	// gettid
 	SYSCALL
@@ -338,6 +346,7 @@
 	MOVQ	R9, g(CX)
 	CALL	runtime·stackcheck(SB)
 
+nog:
 	// Call fn
 	CALL	R12
 
@@ -347,34 +356,6 @@
 	SYSCALL
 	JMP	-3(PC)	// keep exiting
 
-// int32 clone0(int32 flags, void *stack, void* fn, void* fnarg);
-TEXT runtime·clone0(SB),NOSPLIT,$16-36
-	MOVL	flags+0(FP), DI
-	MOVQ	stack+8(FP), SI
-	MOVQ	fn+16(FP), R12      // used by the child
-	MOVQ	fnarg+24(FP), R13   // used by the child
-	MOVL	$0, DX
-	MOVL	$0, R10
-	MOVL	$56, AX
-	SYSCALL
-
-	CMPQ	AX, $0
-	JEQ	child
-	// In parent, return.
-	MOVL	AX, ret+32(FP)
-	RET
-child:
-	MOVQ	SI, SP
-	MOVQ	R12, AX  // fn
-	MOVQ	R13, DI  // fnarg
-	CALL	AX
-
-	// fn shouldn't return; if it does, exit.
-	MOVL	$111, DI
-	MOVL	$60, AX
-	SYSCALL
-	JMP	-3(PC)	// keep exiting
-
 TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 	MOVQ	new+8(SP), DI
 	MOVQ	old+16(SP), SI
diff --git a/src/runtime/sys_linux_arm.s b/src/runtime/sys_linux_arm.s
index bbd7eb9..9d21eae 100644
--- a/src/runtime/sys_linux_arm.s
+++ b/src/runtime/sys_linux_arm.s
@@ -241,7 +241,6 @@
 	MOVW	R0, ret+24(FP)
 	RET
 
-
 // int32 clone(int32 flags, void *stack, M *mp, G *gp, void (*fn)(void));
 TEXT runtime·clone(SB),NOSPLIT,$0
 	MOVW	flags+0(FP), R0
@@ -279,8 +278,15 @@
 	BEQ	2(PC)
 	BL	runtime·abort(SB)
 
-	MOVW	4(R13), g
-	MOVW	0(R13), R8
+	MOVW	0(R13), R8    // m
+	MOVW	4(R13), R0    // g
+
+	CMP	$0, R8
+	BEQ	nog
+	CMP	$0, R0
+	BEQ	nog
+
+	MOVW	R0, g
 	MOVW	R8, g_m(g)
 
 	// paranoia; check they are not nil
@@ -295,54 +301,17 @@
 	MOVW	g_m(g), R8
 	MOVW	R0, m_procid(R8)
 
+nog:
 	// Call fn
 	MOVW	8(R13), R0
 	MOVW	$16(R13), R13
 	BL	(R0)
 
+	// It shouldn't return
 	MOVW	$0, R0
 	MOVW	R0, 4(R13)
 	BL	runtime·exit1(SB)
 
-	// It shouldn't return
-	MOVW	$1234, R0
-	MOVW	$1005, R1
-	MOVW	R0, (R1)
-
-// int32 clone0(int32 flags, void *stack, void* fn, void* fnarg);
-TEXT runtime·clone0(SB),NOSPLIT,$0-20
-	MOVW	flags+0(FP), R0
-	MOVW	stack+4(FP), R1
-	// Update child's future stack and save fn and fnarg on it.
-	MOVW	$-8(R1), R1
-	MOVW	fn+8(FP), R6
-	MOVW	R6, 0(R1)
-	MOVW	fnarg+12(FP), R6
-	MOVW	R6, 4(R1)
-	MOVW	$0, R2	// parent tid ptr
-	MOVW	$0, R3	// tls_val
-	MOVW	$0, R4	// child tid ptr
-	MOVW	$0, R5
-	MOVW	$SYS_clone, R7
-	SWI	$0
-
-	// In parent, return.
-	CMP	$0, R0
-	BEQ	3(PC)
-	MOVW	R0, ret+16(FP)
-	RET
-
-	// In child.
-	MOVW	0(R13), R6   // fn
-	MOVW	4(R13), R0   // fnarg
-	MOVW	$8(R13), R13
-	BL	(R6)
-
-	MOVW	$0, R0
-	MOVW	R0, 4(R13)
-	BL	runtime·exit1(SB)
-
-	// It shouldn't return
 	MOVW	$1234, R0
 	MOVW	$1005, R1
 	MOVW	R0, (R1)
diff --git a/src/runtime/sys_linux_arm64.s b/src/runtime/sys_linux_arm64.s
index 52b34e8..ea8520c 100644
--- a/src/runtime/sys_linux_arm64.s
+++ b/src/runtime/sys_linux_arm64.s
@@ -333,14 +333,19 @@
 	MOVD	$0, R0
 	MOVD	R0, (R0)	// crash
 
-	// Initialize m->procid to Linux tid
 good:
+	// Initialize m->procid to Linux tid
 	MOVD	$SYS_gettid, R8
 	SVC
 
-	MOVD	-24(RSP), R12
-	MOVD	-16(RSP), R11
-	MOVD	-8(RSP), R10
+	MOVD	-24(RSP), R12     // fn
+	MOVD	-16(RSP), R11     // g
+	MOVD	-8(RSP), R10      // m
+
+	CMP	$0, R10
+	BEQ	nog
+	CMP	$0, R11
+	BEQ	nog
 
 	MOVD	R0, m_procid(R10)
 
@@ -351,6 +356,7 @@
 	MOVD	R11, g
 	//CALL	runtime·stackcheck(SB)
 
+nog:
 	// Call fn
 	MOVD	R12, R0
 	BL	(R0)
@@ -362,13 +368,6 @@
 	SVC
 	B	again	// keep exiting
 
-// int32 clone0(int32 flags, void *stack, void* fn, void* fnarg);
-TEXT runtime·clone0(SB),NOSPLIT,$0
-	// TODO(spetrovic): Implement this method.
-	MOVW	$-1, R0
-	MOVW	R0, ret+32(FP)
-	RET
-
 TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 	MOVD	new+0(FP), R0
 	MOVD	old+8(FP), R1
diff --git a/src/runtime/sys_linux_ppc64x.s b/src/runtime/sys_linux_ppc64x.s
index fd7ce4f..ff397f0 100644
--- a/src/runtime/sys_linux_ppc64x.s
+++ b/src/runtime/sys_linux_ppc64x.s
@@ -323,9 +323,14 @@
 	// Initialize m->procid to Linux tid
 	SYSCALL $SYS_gettid
 
-	MOVD	-24(R1), R12
-	MOVD	-16(R1), R8
-	MOVD	-8(R1), R7
+	MOVD	-24(R1), R12       // fn
+	MOVD	-16(R1), R8        // g
+	MOVD	-8(R1), R7         // m
+
+	CMP	R7, $0
+	BEQ	nog
+	CMP	R8, $0
+	BEQ	nog
 
 	MOVD	R3, m_procid(R7)
 
@@ -336,6 +341,7 @@
 	MOVD	R8, g
 	//CALL	runtime·stackcheck(SB)
 
+nog:
 	// Call fn
 	MOVD	R12, CTR
 	BL	(CTR)
@@ -345,13 +351,6 @@
 	SYSCALL $SYS_exit_group
 	BR	-2(PC)	// keep exiting
 
-// int32 clone0(int32 flags, void *stack, void* fn, void* fnarg);
-TEXT runtime·clone0(SB),NOSPLIT,$0
-	// TODO(spetrovic): Implement this method.
-	MOVW	$-1, R3
-	MOVW	R3, ret+32(FP)
-	RETURN
-
 TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 	MOVD	new+0(FP), R3
 	MOVD	old+8(FP), R4