runtime: start all threads with runtime.mstart

Putting the M initialization in multiple places will not scale.
Various code assumes mstart is the start already. Make it so.

R=golang-dev, devon.odell
CC=golang-dev
https://golang.org/cl/7420048

src/pkg/runtime/proc.c

@@ -84,7 +84,7 @@ static void schedule(void);
 static void procresize(int32);
 static void acquirep(P*);
 static P* releasep(void);
-static void newm(void(*)(void), P*, bool, bool);
+static void newm(void(*)(void), P*);
 static void goidle(void);
 static void stopm(void);
 static void startm(P*, bool);
@@ -161,7 +161,7 @@ static FuncVal scavenger = {runtime·MHeap_Scavenger};
 void
 runtime·main(void)
 {
-	newm(sysmon, nil, false, false);
+	newm(sysmon, nil);
 
 	// Lock the main goroutine onto this, the main OS thread,
 	// during initialization.  Most programs won't care, but a few
@@ -381,6 +381,12 @@ runtime·stoptheworld(void)
 	}
 }
 
+static void
+mhelpgc(void)
+{
+	m->helpgc = 1;
+}
+
 void
 runtime·starttheworld(void)
 {
@@ -428,7 +434,7 @@ runtime·starttheworld(void)
 		// coordinate.  This lazy approach works out in practice:
 		// we don't mind if the first couple gc rounds don't have quite
 		// the maximum number of procs.
-		newm(runtime·mstart, nil, true, false);
+		newm(mhelpgc, nil);
 	}
 }
 
@@ -460,6 +466,9 @@ runtime·mstart(void)
 		if(runtime·iscgo)
 			runtime·newextram();
 	}
+	
+	if(m->mstartfn)
+		m->mstartfn();
 
 	if(m->helpgc) {
 		m->helpgc = false;
@@ -726,16 +735,15 @@ unlockextra(M *mp)
 }
 
 
-// Create a new m.  It will start off with a call to fn.
+// Create a new m.  It will start off with a call to fn, or else the scheduler.
 static void
-newm(void(*fn)(void), P *p, bool helpgc, bool spinning)
+newm(void(*fn)(void), P *p)
 {
 	M *mp;
 
 	mp = runtime·allocm(p);
 	mp->nextp = p;
-	mp->helpgc = helpgc;
-	mp->spinning = spinning;
+	mp->mstartfn = fn;
 
 	if(runtime·iscgo) {
 		CgoThreadStart ts;
@@ -744,11 +752,11 @@ newm(void(*fn)(void), P *p, bool helpgc, bool spinning)
 			runtime·throw("_cgo_thread_start missing");
 		ts.m = mp;
 		ts.g = mp->g0;
-		ts.fn = fn;
+		ts.fn = runtime·mstart;
 		runtime·asmcgocall(_cgo_thread_start, &ts);
 		return;
 	}
-	runtime·newosproc(mp, mp->g0, (byte*)mp->g0->stackbase, fn);
+	runtime·newosproc(mp, (byte*)mp->g0->stackbase);
 }
 
 // Stops execution of the current m until new work is available.
@@ -781,12 +789,19 @@ retry:
 	m->nextp = nil;
 }
 
+static void
+mspinning(void)
+{
+	m->spinning = true;
+}
+
 // Schedules some M to run the p (creates an M if necessary).
 // If p==nil, tries to get an idle P, if no idle P's returns false.
 static void
 startm(P *p, bool spinning)
 {
 	M *mp;
+	void (*fn)(void);
 
 	runtime·lock(&runtime·sched);
 	if(p == nil) {
@@ -801,7 +816,10 @@ startm(P *p, bool spinning)
 	mp = mget();
 	runtime·unlock(&runtime·sched);
 	if(mp == nil) {
-		newm(runtime·mstart, p, false, spinning);
+		fn = nil;
+		if(spinning)
+			fn = mspinning;
+		newm(fn, p);
 		return;
 	}
 	if(mp->spinning)
@@ -1887,11 +1905,6 @@ sysmon(void)
 	uint32 idle, delay;
 	uint32 ticks[MaxGomaxprocs];
 
-	// This is a special dedicated thread that retakes P's from blocking syscalls.
-	// It works w/o mcache nor stackalloc, it may work concurrently with GC.
-	runtime·asminit();
-	runtime·minit();
-
 	idle = 0;  // how many cycles in succession we had not wokeup somebody
 	delay = 0;
 	for(;;) {

src/pkg/runtime/runtime.h

@@ -702,7 +702,8 @@ void	runtime·exit1(int32);
 void	runtime·ready(G*);
 byte*	runtime·getenv(int8*);
 int32	runtime·atoi(byte*);
-void	runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void));
+void	runtime·newosproc(M *mp, void *stk);
+void	runtime·mstart(void);
 G*	runtime·malg(int32);
 void	runtime·asminit(void);
 void	runtime·mpreinit(M*);

src/pkg/runtime/sys_freebsd_386.s

@@ -38,12 +38,7 @@ TEXT runtime·thr_start(SB),7,$0
 	
 	MOVL	AX, m(CX)
 	CALL	runtime·stackcheck(SB)		// smashes AX
-
-	// newosproc left the function we should call in mp->mstartfn.
-	get_tls(CX)
-	MOVL	m(CX), AX
-	MOVL	m_mstartfn(AX), AX
-	CALL	AX
+	CALL	runtime·mstart(SB)
 
 	MOVL	0, AX			// crash (not reached)
 

src/pkg/runtime/sys_freebsd_amd64.s

@@ -38,13 +38,8 @@ TEXT runtime·thr_start(SB),7,$0
 	MOVQ	m_g0(R13), DI
 	MOVQ	DI, g(CX)
 
-	CALL runtime·stackcheck(SB)
-	
-	// newosproc left the function we should call in mp->mstartfn.
-	get_tls(CX)
-	MOVQ	m(CX), AX
-	MOVQ	m_mstartfn(AX), AX
-	CALL	AX
+	CALL	runtime·stackcheck(SB)
+	CALL	runtime·mstart(SB)
 
 	MOVQ 0, AX			// crash (not reached)
 

src/pkg/runtime/sys_freebsd_arm.s

@@ -33,10 +33,7 @@ TEXT runtime·thr_start(SB),7,$0
 	// set up g
 	MOVW m_g0(R9), R10
 	BL runtime·emptyfunc(SB) // fault if stack check is wrong
-
-	// newosproc left the function we should call in mp->tls[2] for us.
-	MOVW	(m_tls+8)(m), R0
-	BL	(R0)
+	BL runtime·mstart(SB)
 
 	MOVW $2, R9  // crash (not reached)
 	MOVW R9, (R9)

src/pkg/runtime/sys_windows_386.s

@@ -259,12 +259,7 @@ TEXT runtime·tstart(SB),7,$0
 	CLD
 
 	CALL	runtime·stackcheck(SB)	// clobbers AX,CX
-
-	// newosproc left the function we should call in mp->mstartfn.
-	get_tls(CX)
-	MOVL	m(CX), AX
-	MOVL	m_mstartfn(AX), AX
-	CALL	AX
+	CALL	runtime·mstart(SB)
 
 	RET
 

src/pkg/runtime/sys_windows_amd64.s

@@ -329,11 +329,7 @@ TEXT runtime·tstart_stdcall(SB),7,$0
 	CLD
 
 	CALL	runtime·stackcheck(SB)	// clobbers AX,CX
-
-	get_tls(CX)
-	MOVQ	m(CX), AX
-	MOVQ	m_mstartfn(AX), AX
-	CALL	AX
+	CALL	runtime·mstart(SB)
 
 	XORL	AX, AX			// return 0 == success
 	RET

src/pkg/runtime/thread_darwin.c

@@ -87,19 +87,19 @@ runtime·goenvs(void)
 }
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	int32 errno;
 	Sigset oset;
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
 	if(0){
-		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
+		runtime·printf("newosproc stk=%p m=%p g=%p id=%d/%d ostk=%p\n",
+			stk, mp, mp->g0, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	runtime·sigprocmask(SIG_SETMASK, &sigset_all, &oset);
-	errno = runtime·bsdthread_create(stk, mp, gp, fn);
+	errno = runtime·bsdthread_create(stk, mp, mp->g0, runtime·mstart);
 	runtime·sigprocmask(SIG_SETMASK, &oset, nil);
 
 	if(errno < 0) {

src/pkg/runtime/thread_freebsd.c

@@ -77,18 +77,14 @@ runtime·futexwakeup(uint32 *addr, uint32 cnt)
 void runtime·thr_start(void*);
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	ThrParam param;
 	Sigset oset;
 
-	// thr_start assumes gp == mp->g0
-	if(gp != mp->g0)
-		runtime·throw("invalid newosproc gp");
-
 	if(0){
-		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
+		runtime·printf("newosproc stk=%p m=%p g=%p id=%d/%d ostk=%p\n",
+			stk, mp, mp->g0, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	runtime·sigprocmask(&sigset_all, &oset);
@@ -96,15 +92,18 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 
 	param.start_func = runtime·thr_start;
 	param.arg = (byte*)mp;
-	param.stack_base = (void*)gp->stackbase;
-	param.stack_size = (byte*)stk - (byte*)gp->stackbase;
+	
+	// NOTE(rsc): This code is confused. stackbase is the top of the stack
+	// and is equal to stk. However, it's working, so I'm not changing it.
+	param.stack_base = (void*)mp->g0->stackbase;
+	param.stack_size = (byte*)stk - (byte*)mp->g0->stackbase;
+
 	param.child_tid = (intptr*)&mp->procid;
 	param.parent_tid = nil;
 	param.tls_base = (void*)&mp->tls[0];
 	param.tls_size = sizeof mp->tls;
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
-	mp->mstartfn = fn;
 
 	runtime·thr_new(&param, sizeof param);
 	runtime·sigprocmask(&oset, nil);

src/pkg/runtime/thread_linux.c

@@ -124,7 +124,7 @@ enum
 };
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	int32 ret;
 	int32 flags;
@@ -142,14 +142,14 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
 	if(0){
-		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p clone=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, runtime·clone, mp->id, (int32)mp->tls[0], &mp);
+		runtime·printf("newosproc stk=%p m=%p g=%p clone=%p id=%d/%d ostk=%p\n",
+			stk, mp, mp->g0, runtime·clone, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	// Disable signals during clone, so that the new thread starts
 	// with signals disabled.  It will enable them in minit.
 	runtime·rtsigprocmask(SIG_SETMASK, &sigset_all, &oset, sizeof oset);
-	ret = runtime·clone(flags, stk, mp, gp, fn);
+	ret = runtime·clone(flags, stk, mp, mp->g0, runtime·mstart);
 	runtime·rtsigprocmask(SIG_SETMASK, &oset, nil, sizeof oset);
 
 	if(ret < 0) {

src/pkg/runtime/thread_netbsd.c

@@ -145,15 +145,15 @@ runtime·semawakeup(M *mp)
 }
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	UcontextT uc;
 	int32 ret;
 
 	if(0) {
 		runtime·printf(
-			"newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
+			"newosproc stk=%p m=%p g=%p id=%d/%d ostk=%p\n",
+			stk, mp, mp->g0, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
@@ -164,7 +164,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	uc.uc_link = nil;
 	uc.uc_sigmask = sigset_all;
 
-	runtime·lwp_mcontext_init(&uc.uc_mcontext, stk, mp, gp, fn);
+	runtime·lwp_mcontext_init(&uc.uc_mcontext, stk, mp, mp->g0, runtime·mstart);
 
 	ret = runtime·lwp_create(&uc, 0, &mp->procid);
 

src/pkg/runtime/thread_openbsd.c

@@ -123,7 +123,7 @@ runtime·semawakeup(M *mp)
 }
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	Tfork param;
 	Sigset oset;
@@ -132,7 +132,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	if(0) {
 		runtime·printf(
 			"newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
+			stk, mp, mp->g0, fn, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
@@ -142,7 +142,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	param.tf_stack = stk;
 
 	oset = runtime·sigprocmask(SIG_SETMASK, sigset_all);
-	ret = runtime·tfork((byte*)&param, sizeof(param), mp, gp, fn);
+	ret = runtime·tfork((byte*)&param, sizeof(param), mp, mp->g0, runtime·mstart);
 	runtime·sigprocmask(SIG_SETMASK, oset);
 
 	if(ret < 0) {

src/pkg/runtime/thread_plan9.c

@@ -223,15 +223,15 @@ runtime·exit(int32 e)
 }
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	mp->tls[0] = mp->id;	// so 386 asm can find it
 	if(0){
 		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p rfork=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, runtime·rfork, mp->id, (int32)mp->tls[0], &mp);
+			stk, mp, mp->g0, fn, runtime·rfork, mp->id, (int32)mp->tls[0], &mp);
 	}
 
-	if(runtime·rfork(RFPROC|RFMEM|RFNOWAIT, stk, mp, gp, fn) < 0)
+	if(runtime·rfork(RFPROC|RFMEM|RFNOWAIT, stk, mp, mp->g0, runtime·mstart) < 0)
 		runtime·throw("newosproc: rfork failed");
 }
 

src/pkg/runtime/thread_windows.c

@@ -187,18 +187,12 @@ runtime·semacreate(void)
 #define STACK_SIZE_PARAM_IS_A_RESERVATION ((uintptr)0x00010000)
 
 void
-runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
+runtime·newosproc(M *mp, void *stk)
 {
 	void *thandle;
 
 	USED(stk);
 
-	// assume gp == mp->g0
-	if(gp != mp->g0)
-		runtime·throw("invalid newosproc gp");
-
-	mp->mstartfn = fn;
-
 	thandle = runtime·stdcall(runtime·CreateThread, 6,
 		nil, (uintptr)0x20000, runtime·tstart_stdcall, mp,
 		STACK_SIZE_PARAM_IS_A_RESERVATION, nil);