runtime: initialize shared library at library-load time

This is Part 2 of the change, see Part 1 here: in https://go-review.googlesource.com/#/c/7692/

Suggested by iant@, we use the library initialization entry point to:
    - create a new OS thread and run the "regular" runtime init stack on
      that thread
    - return immediately from the main (i.e., loader) thread
    - at the first CGO invocation, we wait for the runtime initialization
      to complete.

The above mechanism is implemented only on linux_amd64.  Next step is to
support it on linux_arm.  Other platforms don't yet support shared library
compiling/linking, but we intend to use the same strategy there as well.

Change-Id: Ib2c81b1b83bee837134084b75a3beecfb8de6bf4
Reviewed-on: https://go-review.googlesource.com/8094
Run-TryBot: Srdjan Petrovic <spetrovic@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>

src/cmd/cgo/doc.go

@@ -428,6 +428,7 @@ file compiled by gcc, the file x.cgo2.c:
 	void
 	_cgo_be59f0f25121_Cfunc_puts(void *v)
 	{
+		_cgo_wait_runtime_init_done();
 		struct {
 			char* p0;
 			int r;
@@ -436,7 +437,8 @@ file compiled by gcc, the file x.cgo2.c:
 		a->r = puts((void*)a->p0);
 	}
 
-It extracts the arguments from the pointer to _Cfunc_puts's argument
+It waits for Go runtime to be initialized (required for shared libraries),
+extracts the arguments from the pointer to _Cfunc_puts's argument
 frame, invokes the system C function (in this case, puts), stores the
 result in the frame, and returns.
 
@@ -455,6 +457,7 @@ _cgo_main.c:
 
 	int main() { return 0; }
 	void crosscall2(void(*fn)(void*, int), void *a, int c) { }
+	void _cgo_wait_runtime_init_done() { }
 	void _cgo_allocate(void *a, int c) { }
 	void _cgo_panic(void *a, int c) { }
 

src/cmd/cgo/out.go

@@ -52,11 +52,13 @@ func (p *Package) writeDefs() {
 	fmt.Fprintf(fm, "int main() { return 0; }\n")
 	if *importRuntimeCgo {
 		fmt.Fprintf(fm, "void crosscall2(void(*fn)(void*, int), void *a, int c) { }\n")
+		fmt.Fprintf(fm, "void _cgo_wait_runtime_init_done() { }\n")
 		fmt.Fprintf(fm, "char* _cgo_topofstack(void) { return (char*)0; }\n")
 	} else {
 		// If we're not importing runtime/cgo, we *are* runtime/cgo,
-		// which provides crosscall2.  We just need a prototype.
+		// which provides these functions.  We just need a prototype.
 		fmt.Fprintf(fm, "void crosscall2(void(*fn)(void*, int), void *a, int c);\n")
+		fmt.Fprintf(fm, "void _cgo_wait_runtime_init_done();\n")
 	}
 	fmt.Fprintf(fm, "void _cgo_allocate(void *a, int c) { }\n")
 	fmt.Fprintf(fm, "void _cgo_panic(void *a, int c) { }\n")
@@ -641,9 +643,10 @@ func (p *Package) writeExports(fgo2, fm io.Writer) {
 	fmt.Fprintf(fgcch, "%s\n", p.gccExportHeaderProlog())
 
 	fmt.Fprintf(fgcc, "/* Created by cgo - DO NOT EDIT. */\n")
-	fmt.Fprintf(fgcc, "#include \"_cgo_export.h\"\n")
+	fmt.Fprintf(fgcc, "#include \"_cgo_export.h\"\n\n")
 
-	fmt.Fprintf(fgcc, "\nextern void crosscall2(void (*fn)(void *, int), void *, int);\n\n")
+	fmt.Fprintf(fgcc, "extern void crosscall2(void (*fn)(void *, int), void *, int);\n")
+	fmt.Fprintf(fgcc, "extern void _cgo_wait_runtime_init_done();\n\n")
 
 	for _, exp := range p.ExpFunc {
 		fn := exp.Func
@@ -739,6 +742,7 @@ func (p *Package) writeExports(fgo2, fm io.Writer) {
 		fmt.Fprintf(fgcc, "extern void _cgoexp%s_%s(void *, int);\n", cPrefix, exp.ExpName)
 		fmt.Fprintf(fgcc, "\n%s\n", s)
 		fmt.Fprintf(fgcc, "{\n")
+		fmt.Fprintf(fgcc, "\t_cgo_wait_runtime_init_done();\n")
 		fmt.Fprintf(fgcc, "\t%s %v a;\n", ctype, p.packedAttribute())
 		if gccResult != "void" && (len(fntype.Results.List) > 1 || len(fntype.Results.List[0].Names) > 1) {
 			fmt.Fprintf(fgcc, "\t%s r;\n", gccResult)

src/runtime/cgo.go

@@ -14,12 +14,16 @@ import "unsafe"
 //go:linkname _cgo_malloc _cgo_malloc
 //go:linkname _cgo_free _cgo_free
 //go:linkname _cgo_thread_start _cgo_thread_start
+//go:linkname _cgo_sys_thread_create _cgo_sys_thread_create
+//go:linkname _cgo_notify_runtime_init_done _cgo_notify_runtime_init_done
 
 var (
 	_cgo_init                     unsafe.Pointer
 	_cgo_malloc                   unsafe.Pointer
 	_cgo_free                     unsafe.Pointer
 	_cgo_thread_start             unsafe.Pointer
+	_cgo_sys_thread_create        unsafe.Pointer
+	_cgo_notify_runtime_init_done unsafe.Pointer
 )
 
 // iscgo is set to true by the runtime/cgo package

src/runtime/cgo/callbacks.go

@@ -91,5 +91,31 @@ var _cgo_free = &x_cgo_free
 var x_cgo_thread_start byte
 var _cgo_thread_start = &x_cgo_thread_start
 
+// Creates a new system thread without updating any Go state.
+//
+// This method is invoked during shared library loading to create a new OS
+// thread to perform the runtime initialization.  This method is similar to
+// _cgo_sys_thread_start except that it doesn't update any Go state.
+
+//go:cgo_import_static x_cgo_sys_thread_create
+//go:linkname x_cgo_sys_thread_create x_cgo_sys_thread_create
+//go:linkname _cgo_sys_thread_create _cgo_sys_thread_create
+var x_cgo_sys_thread_create byte
+var _cgo_sys_thread_create = &x_cgo_sys_thread_create
+
+// Notifies that the runtime has been intialized.
+//
+// We currently block at every CGO entry point (via _cgo_wait_runtime_init_done)
+// to ensure that the runtime has been initialized before the CGO call is
+// executed.  This is necessary for shared libraries where we kickoff runtime
+// initialization in a separate thread and return without waiting for this
+// thread to complete the init.
+
+//go:cgo_import_static x_cgo_notify_runtime_init_done
+//go:linkname x_cgo_notify_runtime_init_done x_cgo_notify_runtime_init_done
+//go:linkname _cgo_notify_runtime_init_done _cgo_notify_runtime_init_done
+var x_cgo_notify_runtime_init_done byte
+var _cgo_notify_runtime_init_done = &x_cgo_notify_runtime_init_done
+
 //go:cgo_export_static _cgo_topofstack
 //go:cgo_export_dynamic _cgo_topofstack

src/runtime/cgo/gcc_libinit.c

@@ -0,0 +1,41 @@
+// 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.
+
+// +build darwin dragonfly freebsd linux netbsd
+
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h> // strerror
+
+static pthread_cond_t runtime_init_cond;
+static pthread_mutex_t runtime_init_mu;
+static int runtime_init_done;
+
+void
+x_cgo_sys_thread_create(void* (*func)(void*), void* arg) {
+	pthread_t p;
+	int err = pthread_create(&p, NULL, func, arg);
+	if (err != 0) {
+		fprintf(stderr, "pthread_create failed: %s", strerror(err));
+		abort();
+	}
+}
+
+void
+_cgo_wait_runtime_init_done() {
+	pthread_mutex_lock(&runtime_init_mu);
+	while (runtime_init_done == 0) {
+		pthread_cond_wait(&runtime_init_cond, &runtime_init_mu);
+	}
+	pthread_mutex_unlock(&runtime_init_mu);
+}
+
+void
+x_cgo_notify_runtime_init_done(void* dummy) {
+	pthread_mutex_lock(&runtime_init_mu);
+	runtime_init_done = 1;
+	pthread_cond_broadcast(&runtime_init_cond);
+	pthread_mutex_unlock(&runtime_init_mu);
+}

src/runtime/cgo/gcc_libinit_openbsd.c

@@ -0,0 +1,22 @@
+// 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.
+
+#include <stdio.h>
+#include <stdlib.h>
+
+void
+x_cgo_sys_thread_create(void* (*func)(void*), void* arg) {
+	fprintf(stderr, "x_cgo_sys_thread_create not implemented");
+	abort();
+}
+
+void
+_cgo_wait_runtime_init_done() {
+	// TODO(spetrovic): implement this method.
+}
+
+void
+x_cgo_notify_runtime_init_done(void* dummy) {
+	// TODO(spetrovic): implement this method.
+}

src/runtime/cgo/gcc_libinit_windows.c

@@ -0,0 +1,22 @@
+// 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.
+
+#include <stdio.h>
+#include <stdlib.h>
+
+void
+x_cgo_sys_thread_create(void* (*func)(void*), void* arg) {
+	fprintf(stderr, "x_cgo_sys_thread_create not implemented");
+	abort();
+}
+
+void
+_cgo_wait_runtime_init_done() {
+	// TODO(spetrovic): implement this method.
+}
+
+void
+x_cgo_notify_runtime_init_done(void* dummy) {
+	// TODO(spetrovic): implement this method.
+}

src/runtime/cgo/libcgo.h

@@ -44,11 +44,22 @@ struct ThreadStart
  */
 extern void (*_cgo_thread_start)(ThreadStart *ts);
 
+/*
+ * Creates a new operating system thread without updating any Go state
+ * (OS dependent).
+ */
+extern void (*_cgo_sys_thread_create)(void* (*func)(void*), void* arg);
+
 /*
  * Creates the new operating system thread (OS, arch dependent).
  */
 void _cgo_sys_thread_start(ThreadStart *ts);
 
+/*
+ * Waits for the Go runtime to be initialized (OS dependent).
+ */
+void _cgo_wait_runtime_init_done();
+
 /*
  * Call fn in the 6c world.
  */

src/runtime/os1_linux.go

@@ -111,6 +111,12 @@ const (
 	_CLONE_STOPPED        = 0x2000000
 	_CLONE_NEWUTS         = 0x4000000
 	_CLONE_NEWIPC         = 0x8000000
+
+	cloneFlags = _CLONE_VM | /* share memory */
+		_CLONE_FS | /* share cwd, etc */
+		_CLONE_FILES | /* share fd table */
+		_CLONE_SIGHAND | /* share sig handler table */
+		_CLONE_THREAD /* revisit - okay for now */
 )
 
 // May run with m.p==nil, so write barriers are not allowed.
@@ -119,12 +125,6 @@ func newosproc(mp *m, stk unsafe.Pointer) {
 	/*
 	 * note: strace gets confused if we use CLONE_PTRACE here.
 	 */
-	var flags int32 = _CLONE_VM | /* share memory */
-		_CLONE_FS | /* share cwd, etc */
-		_CLONE_FILES | /* share fd table */
-		_CLONE_SIGHAND | /* share sig handler table */
-		_CLONE_THREAD /* revisit - okay for now */
-
 	mp.tls[0] = uintptr(mp.id) // so 386 asm can find it
 	if false {
 		print("newosproc stk=", stk, " m=", mp, " g=", mp.g0, " clone=", funcPC(clone), " id=", mp.id, "/", mp.tls[0], " ostk=", &mp, "\n")
@@ -134,7 +134,7 @@ func newosproc(mp *m, stk unsafe.Pointer) {
 	// with signals disabled.  It will enable them in minit.
 	var oset sigset
 	rtsigprocmask(_SIG_SETMASK, &sigset_all, &oset, int32(unsafe.Sizeof(oset)))
-	ret := clone(flags, stk, unsafe.Pointer(mp), unsafe.Pointer(mp.g0), unsafe.Pointer(funcPC(mstart)))
+	ret := clone(cloneFlags, stk, unsafe.Pointer(mp), unsafe.Pointer(mp.g0), unsafe.Pointer(funcPC(mstart)))
 	rtsigprocmask(_SIG_SETMASK, &oset, nil, int32(unsafe.Sizeof(oset)))
 
 	if ret < 0 {
@@ -143,6 +143,25 @@ func newosproc(mp *m, stk unsafe.Pointer) {
 	}
 }
 
+// Version of newosproc that doesn't require any Go structs to be allocated.
+//go:nosplit
+func newosproc0(stacksize uintptr, fn unsafe.Pointer, fnarg 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)
+	if ret < 0 {
+		write(2, unsafe.Pointer(&failthreadcreate[0]), int32(len(failthreadcreate)))
+		exit(1)
+	}
+}
+
+var failallocatestack = []byte("runtime: failed to allocate stack for the new OS thread\n")
+var failthreadcreate = []byte("runtime: failed to create new OS thread\n")
+
 func osinit() {
 	ncpu = getproccount()
 }

src/runtime/os_linux.go

@@ -12,6 +12,9 @@ func futex(addr unsafe.Pointer, op int32, val uint32, ts, addr2 unsafe.Pointer,
 //go:noescape
 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
 

src/runtime/proc.go

@@ -66,6 +66,10 @@ func main() {
 
 	gcenable()
 
+	if islibrary {
+		// Allocate new M as main_main() is expected to block forever.
+		systemstack(newextram)
+	}
 	if iscgo {
 		if _cgo_thread_start == nil {
 			throw("_cgo_thread_start missing")
@@ -84,6 +88,10 @@ func main() {
 				throw("_cgo_unsetenv missing")
 			}
 		}
+		if _cgo_notify_runtime_init_done == nil {
+			throw("_cgo_notify_runtime_init_done missing")
+		}
+		cgocall(_cgo_notify_runtime_init_done, nil)
 	}
 
 	main_init()

src/runtime/rt0_linux_amd64.s

@@ -12,10 +12,37 @@ TEXT _rt0_amd64_linux(SB),NOSPLIT,$-8
 
 // When linking with -shared, this symbol is called when the shared library
 // is loaded.
-TEXT _rt0_amd64_linux_lib(SB),NOSPLIT,$0
-	// TODO(spetrovic): Do something useful, like calling $main.  (Note that
-	// this has to be done in a separate thread, as main is expected to block.)
+TEXT _rt0_amd64_linux_lib(SB),NOSPLIT,$40
+	MOVQ	DI, _rt0_amd64_linux_lib_argc<>(SB)
+	MOVQ	SI, _rt0_amd64_linux_lib_argv<>(SB)
+
+	// Create a new thread to do the runtime initialization and return.
+	MOVQ	_cgo_sys_thread_create(SB), AX
+	TESTQ	AX, AX
+	JZ	nocgo
+	MOVQ	$_rt0_amd64_linux_lib_go(SB), DI
+	MOVQ	$0, SI
+	CALL	AX
 	RET
+nocgo:
+	MOVQ	$8388608, 0(SP)                    // stacksize
+	MOVQ	$_rt0_amd64_linux_lib_go(SB), AX
+	MOVQ	AX, 8(SP)                          // fn
+	MOVQ	$0, 16(SP)                         // fnarg
+	MOVQ	$runtime·newosproc0(SB), AX
+	CALL	AX
+	RET
+
+TEXT _rt0_amd64_linux_lib_go(SB),NOSPLIT,$0
+	MOVQ	_rt0_amd64_linux_lib_argc<>(SB), DI
+	MOVQ	_rt0_amd64_linux_lib_argv<>(SB), SI
+	MOVQ	$runtime·rt0_go(SB), AX
+	JMP	AX
+
+DATA _rt0_amd64_linux_lib_argc<>(SB)/8, $0
+GLOBL _rt0_amd64_linux_lib_argc<>(SB),NOPTR, $8
+DATA _rt0_amd64_linux_lib_argv<>(SB)/8, $0
+GLOBL _rt0_amd64_linux_lib_argv<>(SB),NOPTR, $8
 
 TEXT main(SB),NOSPLIT,$-8
 	MOVQ	$runtime·rt0_go(SB), AX

src/runtime/runtime2.go

@@ -621,6 +621,9 @@ var (
 	cpuid_ecx         uint32
 	cpuid_edx         uint32
 	lfenceBeforeRdtsc bool
+
+	// Set by the linker when linking with -shared.
+	islibrary bool
 )
 
 /*

src/runtime/sys_linux_386.s

@@ -369,6 +369,12 @@ TEXT runtime·clone(SB),NOSPLIT,$0
 	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

src/runtime/sys_linux_amd64.s

@@ -347,6 +347,34 @@ TEXT runtime·clone(SB),NOSPLIT,$0
 	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

src/runtime/sys_linux_arm.s

@@ -309,6 +309,12 @@ TEXT runtime·clone(SB),NOSPLIT,$0
 	MOVW	$1005, R1
 	MOVW	R0, (R1)
 
+// int32 clone0(int32 flags, void *stack, void* fn, void* fnarg);
+TEXT runtime·clone0(SB),NOSPLIT,$0
+	// TODO(spetrovic): Implement this method.
+	MOVW	$-1, ret+16(FP)
+	RET
+
 TEXT runtime·sigaltstack(SB),NOSPLIT,$0
 	MOVW	new+0(FP), R0
 	MOVW	old+4(FP), R1

src/runtime/sys_linux_arm64.s

@@ -356,6 +356,12 @@ again:
 	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	$-1, ret+32(FP)
+	RET
+
 TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 	MOVD	new+0(FP), R0
 	MOVD	old+8(FP), R1

src/runtime/sys_linux_ppc64x.s

@@ -345,6 +345,12 @@ TEXT runtime·clone(SB),NOSPLIT,$-8
 	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	$-1, ret+32(FP)
+	RETURN
+
 TEXT runtime·sigaltstack(SB),NOSPLIT,$-8
 	MOVD	new+0(FP), R3
 	MOVD	old+8(FP), R4