runtime: implement pprof support for windows

Credit to jp for proof of concept.

R=alex.brainman, jp, rsc, dvyukov
CC=golang-dev
https://golang.org/cl/4960057

src/pkg/runtime/runtime.h

@@ -212,6 +212,7 @@ struct	G
 	uintptr	sigcode1;
 	uintptr	sigpc;
 	uintptr	gopc;	// pc of go statement that created this goroutine
+	uintptr	end[];
 };
 struct	M
 {
@@ -253,9 +254,11 @@ struct	M
 	uint32	fflag;		// floating point compare flags
 
 #ifdef __WINDOWS__
+	void*	thread;		// thread handle
 	void*	event;		// event for signalling
 	M*	nextwaitm;	// next M waiting for lock
 #endif
+	uintptr	end[];
 };
 
 struct	Stktop

src/pkg/runtime/windows/386/defs.h

@@ -10,9 +10,13 @@ enum {
 	PROT_EXEC = 0x4,
 	MAP_ANON = 0x1,
 	MAP_PRIVATE = 0x2,
+	DUPLICATE_SAME_ACCESS = 0x2,
+	THREAD_PRIORITY_HIGHEST = 0x2,
 	SIGINT = 0x2,
 	CTRL_C_EVENT = 0,
 	CTRL_BREAK_EVENT = 0x1,
+	CONTEXT_CONTROL = 0x10001,
+	CONTEXT_FULL = 0x10007,
 	EXCEPTION_ACCESS_VIOLATION = 0xc0000005,
 	EXCEPTION_BREAKPOINT = 0x80000003,
 	EXCEPTION_FLT_DENORMAL_OPERAND = 0xc000008d,

src/pkg/runtime/windows/386/signal.c

@@ -90,9 +90,7 @@ runtime·sighandler(ExceptionRecord *info, void *frame, Context *r)
 }
 
 void
-runtime·resetcpuprofiler(int32 hz)
+runtime·dosigprof(Context *r, G *gp)
 {
-	// TODO: Enable profiling interrupts.
-	
-	m->profilehz = hz;
+	runtime·sigprof((uint8*)r->Eip, (uint8*)r->Esp, nil, gp);
 }

src/pkg/runtime/windows/386/sys.s

@@ -96,31 +96,52 @@ TEXT runtime·sigtramp1(SB),0,$16-40
 sigdone:
 	RET
 
-// Windows runs the ctrl handler in a new thread.
 TEXT runtime·ctrlhandler(SB),7,$0
+	PUSHL	$runtime·ctrlhandler1(SB)
+	CALL	runtime·externalthreadhandler(SB)
+	MOVL	4(SP), CX
+	ADDL	$12, SP
+	JMP	CX
+
+TEXT runtime·profileloop(SB),7,$0
+	PUSHL	$runtime·profileloop1(SB)
+	CALL	runtime·externalthreadhandler(SB)
+	MOVL	4(SP), CX
+	ADDL	$12, SP
+	JMP	CX
+
+TEXT runtime·externalthreadhandler(SB),7,$0
 	PUSHL	BP
 	MOVL	SP, BP
 	PUSHL	BX
 	PUSHL	SI
 	PUSHL	DI
 	PUSHL	0x2c(FS)
-	MOVL	SP, BX
+	MOVL	SP, DX
 
 	// setup dummy m, g
-	SUBL	$(m_fflag+4), SP	// at least space for m_fflag
+	SUBL	$m_end, SP		// space for M
+	MOVL	SP, 0(SP)
+	MOVL	$m_end, 4(SP)
+	CALL	runtime·memclr(SB)	// smashes AX,BX,CX
+
 	LEAL	m_tls(SP), CX
 	MOVL	CX, 0x2c(FS)
 	MOVL	SP, m(CX)
-	MOVL	SP, DX
-	SUBL	$8, SP			// space for g_stack{guard,base}
+	MOVL	SP, BX
+	SUBL	$g_end, SP		// space for G
 	MOVL	SP, g(CX)
-	MOVL	SP, m_g0(DX)
+	MOVL	SP, m_g0(BX)
+
+	MOVL	SP, 0(SP)
+	MOVL	$g_end, 4(SP)
+	CALL	runtime·memclr(SB)	// smashes AX,BX,CX
 	LEAL	-4096(SP), CX
 	MOVL	CX, g_stackguard(SP)
-	MOVL	BX, g_stackbase(SP)
+	MOVL	DX, g_stackbase(SP)
 
-	PUSHL	8(BP)
-	CALL	runtime·ctrlhandler1(SB)
+	PUSHL	16(BP)			// arg for handler
+	CALL	8(BP)
 	POPL	CX
 
 	get_tls(CX)
@@ -131,9 +152,7 @@ TEXT runtime·ctrlhandler(SB),7,$0
 	POPL	SI
 	POPL	BX
 	POPL	BP
-	MOVL	0(SP), CX
-	ADDL	$8, SP
-	JMP	CX
+	RET
 
 // Called from dynamic function created by ../thread.c compilecallback,
 // running on Windows stack (not Go stack).

src/pkg/runtime/windows/amd64/defs.h

@@ -10,9 +10,13 @@ enum {
 	PROT_EXEC = 0x4,
 	MAP_ANON = 0x1,
 	MAP_PRIVATE = 0x2,
+	DUPLICATE_SAME_ACCESS = 0x2,
+	THREAD_PRIORITY_HIGHEST = 0x2,
 	SIGINT = 0x2,
 	CTRL_C_EVENT = 0,
 	CTRL_BREAK_EVENT = 0x1,
+	CONTEXT_CONTROL = 0x100001,
+	CONTEXT_FULL = 0x10000b,
 	EXCEPTION_ACCESS_VIOLATION = 0xc0000005,
 	EXCEPTION_BREAKPOINT = 0x80000003,
 	EXCEPTION_FLT_DENORMAL_OPERAND = 0xc000008d,

src/pkg/runtime/windows/amd64/signal.c

@@ -100,9 +100,7 @@ runtime·sighandler(ExceptionRecord *info, Context *r, G *gp)
 }
 
 void
-runtime·resetcpuprofiler(int32 hz)
+runtime·dosigprof(Context *r, G *gp)
 {
-	// TODO: Enable profiling interrupts.
-	
-	m->profilehz = hz;
+	runtime·sigprof((uint8*)r->Rip, (uint8*)r->Rsp, nil, gp);
 }

src/pkg/runtime/windows/amd64/sys.s

@@ -100,31 +100,51 @@ TEXT runtime·sigtramp(SB),7,$56
 sigdone:
 	RET
 
-// Windows runs the ctrl handler in a new thread.
-TEXT runtime·ctrlhandler(SB),7,$0
+TEXT runtime·ctrlhandler(SB),7,$8
+	MOVQ	CX, 16(SP)		// spill
+	MOVQ	$runtime·ctrlhandler1(SB), CX
+	MOVQ	CX, 0(SP)
+	CALL	runtime·externalthreadhandler(SB)
+	RET
+
+TEXT runtime·profileloop(SB),7,$8
+	MOVQ	$runtime·profileloop1(SB), CX
+	MOVQ	CX, 0(SP)
+	CALL	runtime·externalthreadhandler(SB)
+	RET
+
+TEXT runtime·externalthreadhandler(SB),7,$0
 	PUSHQ	BP
 	MOVQ	SP, BP
 	PUSHQ	BX
 	PUSHQ	SI
 	PUSHQ	DI
 	PUSHQ	0x58(GS)
-	MOVQ	SP, BX
+	MOVQ	SP, DX
 
 	// setup dummy m, g
-	SUBQ	$(m_fflag+4), SP	// at least space for m_fflag
+	SUBQ	$m_end, SP		// space for M
+	MOVQ	SP, 0(SP)
+	MOVQ	$m_end, 8(SP)
+	CALL	runtime·memclr(SB)	// smashes AX,BX,CX
+
 	LEAQ	m_tls(SP), CX
 	MOVQ	CX, 0x58(GS)
 	MOVQ	SP, m(CX)
-	MOVQ	SP, DX
-	SUBQ	$16, SP			// space for g_stack{guard,base}
+	MOVQ	SP, BX
+	SUBQ	$g_end, SP		// space for G
 	MOVQ	SP, g(CX)
-	MOVQ	SP, m_g0(DX)
+	MOVQ	SP, m_g0(BX)
+
+	MOVQ	SP, 0(SP)
+	MOVQ	$g_end, 8(SP)
+	CALL	runtime·memclr(SB)	// smashes AX,BX,CX
 	LEAQ	-8192(SP), CX
 	MOVQ	CX, g_stackguard(SP)
-	MOVQ	BX, g_stackbase(SP)
+	MOVQ	DX, g_stackbase(SP)
 
-	PUSHQ	16(BP)
-	CALL	runtime·ctrlhandler1(SB)
+	PUSHQ	32(BP)			// arg for handler
+	CALL	16(BP)
 	POPQ	CX
 
 	get_tls(CX)

src/pkg/runtime/windows/defs.c

@@ -17,10 +17,16 @@ enum {
 	$MAP_ANON = 1,
 	$MAP_PRIVATE = 2,
 
+	$DUPLICATE_SAME_ACCESS = DUPLICATE_SAME_ACCESS,
+	$THREAD_PRIORITY_HIGHEST = THREAD_PRIORITY_HIGHEST,
+
 	$SIGINT = SIGINT,
 	$CTRL_C_EVENT = CTRL_C_EVENT,
 	$CTRL_BREAK_EVENT = CTRL_BREAK_EVENT,
 
+	$CONTEXT_CONTROL = CONTEXT_CONTROL,
+	$CONTEXT_FULL = CONTEXT_FULL,
+
 	$EXCEPTION_ACCESS_VIOLATION = STATUS_ACCESS_VIOLATION,
 	$EXCEPTION_BREAKPOINT = STATUS_BREAKPOINT,
 	$EXCEPTION_FLT_DENORMAL_OPERAND = STATUS_FLOAT_DENORMAL_OPERAND,

src/pkg/runtime/windows/os.h

@@ -13,8 +13,8 @@ extern void *runtime·GetProcAddress;
 void runtime·asmstdcall(void *c);
 void *runtime·stdcall(void *fn, int32 count, ...);
 
-uintptr runtime·getlasterror(void);
-void runtime·setlasterror(uintptr err);
+uint32 runtime·getlasterror(void);
+void runtime·setlasterror(uint32 err);
 
 // Function to be called by windows CreateThread
 // to start new os thread.

src/pkg/runtime/windows/thread.c

@@ -10,32 +10,48 @@
 #pragma dynimport runtime·CloseHandle CloseHandle "kernel32.dll"
 #pragma dynimport runtime·CreateEvent CreateEventA "kernel32.dll"
 #pragma dynimport runtime·CreateThread CreateThread "kernel32.dll"
+#pragma dynimport runtime·CreateWaitableTimer CreateWaitableTimerA "kernel32.dll"
+#pragma dynimport runtime·DuplicateHandle DuplicateHandle "kernel32.dll"
 #pragma dynimport runtime·ExitProcess ExitProcess "kernel32.dll"
 #pragma dynimport runtime·FreeEnvironmentStringsW FreeEnvironmentStringsW "kernel32.dll"
 #pragma dynimport runtime·GetEnvironmentStringsW GetEnvironmentStringsW "kernel32.dll"
 #pragma dynimport runtime·GetProcAddress GetProcAddress "kernel32.dll"
 #pragma dynimport runtime·GetStdHandle GetStdHandle "kernel32.dll"
+#pragma dynimport runtime·GetThreadContext GetThreadContext "kernel32.dll"
 #pragma dynimport runtime·LoadLibraryEx LoadLibraryExA "kernel32.dll"
 #pragma dynimport runtime·QueryPerformanceCounter QueryPerformanceCounter "kernel32.dll"
 #pragma dynimport runtime·QueryPerformanceFrequency QueryPerformanceFrequency "kernel32.dll"
+#pragma dynimport runtime·ResumeThread ResumeThread "kernel32.dll"
 #pragma dynimport runtime·SetConsoleCtrlHandler SetConsoleCtrlHandler "kernel32.dll"
 #pragma dynimport runtime·SetEvent SetEvent "kernel32.dll"
+#pragma dynimport runtime·SetThreadPriority SetThreadPriority "kernel32.dll"
+#pragma dynimport runtime·SetWaitableTimer SetWaitableTimer "kernel32.dll"
+#pragma dynimport runtime·SuspendThread SuspendThread "kernel32.dll"
+#pragma dynimport runtime·timeBeginPeriod timeBeginPeriod "winmm.dll"
 #pragma dynimport runtime·WaitForSingleObject WaitForSingleObject "kernel32.dll"
 #pragma dynimport runtime·WriteFile WriteFile "kernel32.dll"
 
 extern void *runtime·CloseHandle;
 extern void *runtime·CreateEvent;
 extern void *runtime·CreateThread;
+extern void *runtime·CreateWaitableTimer;
+extern void *runtime·DuplicateHandle;
 extern void *runtime·ExitProcess;
 extern void *runtime·FreeEnvironmentStringsW;
 extern void *runtime·GetEnvironmentStringsW;
 extern void *runtime·GetProcAddress;
 extern void *runtime·GetStdHandle;
+extern void *runtime·GetThreadContext;
 extern void *runtime·LoadLibraryEx;
 extern void *runtime·QueryPerformanceCounter;
 extern void *runtime·QueryPerformanceFrequency;
+extern void *runtime·ResumeThread;
 extern void *runtime·SetConsoleCtrlHandler;
 extern void *runtime·SetEvent;
+extern void *runtime·SetThreadPriority;
+extern void *runtime·SetWaitableTimer;
+extern void *runtime·SuspendThread;
+extern void *runtime·timeBeginPeriod;
 extern void *runtime·WaitForSingleObject;
 extern void *runtime·WriteFile;
 
@@ -44,8 +60,13 @@ static int64 timerfreq;
 void
 runtime·osinit(void)
 {
+	// -1 = current process, -2 = current thread
+	runtime·stdcall(runtime·DuplicateHandle, 7,
+		(uintptr)-1, (uintptr)-2, (uintptr)-1, &m->thread,
+		(uintptr)0, (uintptr)0, (uintptr)DUPLICATE_SAME_ACCESS);
 	runtime·stdcall(runtime·QueryPerformanceFrequency, 1, &timerfreq);
 	runtime·stdcall(runtime·SetConsoleCtrlHandler, 2, runtime·ctrlhandler, (uintptr)1);
+	runtime·stdcall(runtime·timeBeginPeriod, 1, (uintptr)1);
 }
 
 void
@@ -211,11 +232,13 @@ runtime·newosproc(M *m, G *g, void *stk, void (*fn)(void))
 	USED(g);	// assuming g = m->g0
 	USED(fn);	// assuming fn = mstart
 
-	thandle = runtime·stdcall(runtime·CreateThread, 6, (uintptr)0, (uintptr)0, runtime·tstart_stdcall, m, (uintptr)0, (uintptr)0);
-	if(thandle == 0) {
+	thandle = runtime·stdcall(runtime·CreateThread, 6,
+		nil, nil, runtime·tstart_stdcall, m, nil, nil);
+	if(thandle == nil) {
 		runtime·printf("runtime: failed to create new OS thread (have %d already; errno=%d)\n", runtime·mcount(), runtime·getlasterror());
 		runtime·throw("runtime.newosproc");
 	}
+	runtime·atomicstorep(&m->thread, thandle);
 }
 
 // Called to initialize a new m (including the bootstrap m).
@@ -324,6 +347,89 @@ runtime·ctrlhandler1(uint32 type)
 	return 0;
 }
 
+extern void runtime·dosigprof(Context *r, G *gp);
+extern void runtime·profileloop(void);
+static void *profiletimer;
+
+static void
+profilem(M *mp)
+{
+	extern M runtime·m0;
+	extern uint32 runtime·tls0[];
+	byte rbuf[sizeof(Context)+15];
+	Context *r;
+	void *tls;
+	G *gp;
+
+	tls = mp->tls;
+	if(mp == &runtime·m0)
+		tls = runtime·tls0;
+	gp = *(G**)tls;
+
+	if(gp != nil && gp != mp->g0 && gp->status != Gsyscall) {
+		// align Context to 16 bytes
+		r = (Context*)((uintptr)(&rbuf[15]) & ~15);
+		r->ContextFlags = CONTEXT_CONTROL;
+		runtime·stdcall(runtime·GetThreadContext, 2, mp->thread, r);
+		runtime·dosigprof(r, gp);
+	}
+}
+
+void
+runtime·profileloop1(void)
+{
+	M *mp, *allm;
+	void *thread;
+
+	runtime·stdcall(runtime·SetThreadPriority, 2,
+		(uintptr)-2, (uintptr)THREAD_PRIORITY_HIGHEST);
+
+	for(;;) {
+		runtime·stdcall(runtime·WaitForSingleObject, 2, profiletimer, (uintptr)-1);
+		allm = runtime·atomicloadp(&runtime·allm);
+		for(mp = allm; mp != nil; mp = mp->alllink) {
+			thread = runtime·atomicloadp(&mp->thread);
+			if(thread == nil)
+				continue;
+			runtime·stdcall(runtime·SuspendThread, 1, thread);
+			if(mp->profilehz != 0)
+				profilem(mp);
+			runtime·stdcall(runtime·ResumeThread, 1, thread);
+		}
+	}
+}
+
+void
+runtime·resetcpuprofiler(int32 hz)
+{
+	static Lock lock;
+	void *timer, *thread;
+	int32 ms;
+	int64 due;
+
+	runtime·lock(&lock);
+	if(profiletimer == nil) {
+		timer = runtime·stdcall(runtime·CreateWaitableTimer, 3, nil, nil, nil);
+		runtime·atomicstorep(&profiletimer, timer);
+		thread = runtime·stdcall(runtime·CreateThread, 6,
+			nil, nil, runtime·profileloop, nil, nil, nil);
+		runtime·stdcall(runtime·CloseHandle, 1, thread);
+	}
+	runtime·unlock(&lock);
+
+	ms = 0;
+	due = 1LL<<63;
+	if(hz > 0) {
+		ms = 1000 / hz;
+		if(ms == 0)
+			ms = 1;
+		due = ms * -10000;
+	}
+	runtime·stdcall(runtime·SetWaitableTimer, 6,
+		profiletimer, &due, (uintptr)ms, nil, nil, nil);
+	runtime·atomicstore((uint32*)&m->profilehz, hz);
+}
+
 void
 os·sigpipe(void)
 {