runtime: improve cpu profiles for GC/syscalls/cgo

Current "System->etext" is not very informative.
Add parent "GC" frame.
Replace un-unwindable syscall/cgo frames with Go stack that leads to the call.

LGTM=rsc
R=rsc, alex.brainman, ality
CC=golang-codereviews
https://golang.org/cl/61270043

src/pkg/runtime/os_windows.c

@@ -291,7 +291,14 @@ runtime·stdcall(void *fn, int32 count, ...)
 	m->libcall.fn = fn;
 	m->libcall.n = count;
 	m->libcall.args = (uintptr*)&count + 1;
+	if(m->profilehz != 0) {
+		// leave pc/sp for cpu profiler
+		m->libcallpc = (uintptr)runtime·getcallerpc(&fn);
+		m->libcallsp = (uintptr)runtime·getcallersp(&fn);
+		m->libcallg = g;
+	}
 	runtime·asmcgocall(runtime·asmstdcall, &m->libcall);
+	m->libcallsp = 0;
 	return (void*)m->libcall.r1;
 }
 

src/pkg/runtime/proc.c

@@ -2104,10 +2104,10 @@ static struct {
 	uintptr pcbuf[100];
 } prof;
 
-static void
-System(void)
-{
-}
+static void System(void) {}
+static void ExternalCode(void) {}
+static void GC(void) {}
+extern byte etext[];
 
 // Called if we receive a SIGPROF signal.
 void
@@ -2221,9 +2221,35 @@ runtime·sigprof(uint8 *pc, uint8 *sp, uint8 *lr, G *gp, M *mp)
 	if(traceback)
 		n = runtime·gentraceback((uintptr)pc, (uintptr)sp, (uintptr)lr, gp, 0, prof.pcbuf, nelem(prof.pcbuf), nil, nil, false);
 	if(!traceback || n <= 0) {
-		n = 2;
-		prof.pcbuf[0] = (uintptr)pc;
-		prof.pcbuf[1] = (uintptr)System + 1;
+		// Normal traceback is impossible or has failed.
+		// See if it falls into several common cases.
+		n = 0;
+		if(mp->ncgo > 0 && mp->curg != nil &&
+			mp->curg->syscallpc != 0 && mp->curg->syscallsp != 0) {
+			// Cgo, we can't unwind and symbolize arbitrary C code,
+			// so instead collect Go stack that leads to the cgo call.
+			// This is especially important on windows, since all syscalls are cgo calls.
+			n = runtime·gentraceback(mp->curg->syscallpc, mp->curg->syscallsp, 0, mp->curg, 0, prof.pcbuf, nelem(prof.pcbuf), nil, nil, false);
+		}
+#ifdef GOOS_windows
+		if(n == 0 && mp->libcallg != nil && mp->libcallpc != 0 && mp->libcallsp != 0) {
+			// Libcall, i.e. runtime syscall on windows.
+			// Collect Go stack that leads to the call.
+			n = runtime·gentraceback(mp->libcallpc, mp->libcallsp, 0, mp->libcallg, 0, prof.pcbuf, nelem(prof.pcbuf), nil, nil, false);
+		}
+#endif
+		if(n == 0) {
+			// If all of the above has failed, account it against abstract "System" or "GC".
+			n = 2;
+			// "ExternalCode" is better than "etext".
+			if((uintptr)pc > (uintptr)etext)
+				pc = (byte*)ExternalCode + PCQuantum;
+			prof.pcbuf[0] = (uintptr)pc;
+			if(mp->gcing || mp->helpgc)
+				prof.pcbuf[1] = (uintptr)GC + PCQuantum;
+			else
+				prof.pcbuf[1] = (uintptr)System + PCQuantum;
+		}
 	}
 	prof.fn(prof.pcbuf, n);
 	runtime·unlock(&prof);

src/pkg/runtime/runtime.h

@@ -350,6 +350,9 @@ struct	M
 	// these are here because they are too large to be on the stack
 	// of low-level NOSPLIT functions.
 	LibCall	libcall;
+	uintptr	libcallpc;	// for cpu profiler
+	uintptr	libcallsp;
+	G*	libcallg;
 #endif
 #ifdef GOOS_solaris
 	int32*	perrno; 	// pointer to TLS errno

src/pkg/runtime/sys_windows_386.s

@@ -343,19 +343,34 @@ TEXT runtime·usleep1(SB),NOSPLIT,$0
 	RET
 
 	MOVL	m(CX), BP
+
+	// leave pc/sp for cpu profiler
+	MOVL	(SP), SI
+	MOVL	SI, m_libcallpc(BP)
+	LEAL	4(SP), SI
+	MOVL	SI, m_libcallsp(BP)
+	MOVL	g(BP), SI
+	MOVL	SI, m_libcallg(BP)
+
 	MOVL	m_g0(BP), SI
 	CMPL	g(CX), SI
-	JNE	3(PC)
+	JNE	usleep1_switch
 	// executing on m->g0 already
 	CALL	AX
-	RET
+	JMP	usleep1_ret
 
+usleep1_switch:
 	// Switch to m->g0 stack and back.
 	MOVL	(g_sched+gobuf_sp)(SI), SI
 	MOVL	SP, -4(SI)
 	LEAL	-4(SI), SP
 	CALL	AX
 	MOVL	0(SP), SP
+
+usleep1_ret:
+	get_tls(CX)
+	MOVL	m(CX), BP
+	MOVL	$0, m_libcallsp(BP)
 	RET
 
 // Runs on OS stack. duration (in 100ns units) is in BX.

src/pkg/runtime/sys_windows_amd64.s

@@ -337,20 +337,33 @@ TEXT runtime·usleep1(SB),NOSPLIT,$0
 	CALL	AX
 	RET
 
-	MOVQ	m(R15), R14
-	MOVQ	m_g0(R14), R14
+	MOVQ	m(R15), R13
+
+	// leave pc/sp for cpu profiler
+	MOVQ	(SP), R12
+	MOVQ	R12, m_libcallpc(R13)
+	LEAQ	8(SP), R12
+	MOVQ	R12, m_libcallsp(R13)
+	MOVQ	g(R13), R12
+	MOVQ	R12, m_libcallg(R13)
+
+	MOVQ	m_g0(R13), R14
 	CMPQ	g(R15), R14
-	JNE	3(PC)
+	JNE	usleep1_switch
 	// executing on m->g0 already
 	CALL	AX
-	RET
+	JMP	usleep1_ret
 
+usleep1_switch:
 	// Switch to m->g0 stack and back.
 	MOVQ	(g_sched+gobuf_sp)(R14), R14
 	MOVQ	SP, -8(R14)
 	LEAQ	-8(R14), SP
 	CALL	AX
 	MOVQ	0(SP), SP
+
+usleep1_ret:
+	MOVQ	$0, m_libcallsp(R13)
 	RET
 
 // Runs on OS stack. duration (in 100ns units) is in BX.