go threads for OS X

R=r
OCL=14944
CL=15013

src/runtime/amd64_darwin.h

@@ -22,6 +22,11 @@ struct timespec {
 	int64 tv_nsec;
 };
 
+struct timeval {
+	time_t tv_sec;
+	int64 tv_usec;
+};
+
 struct stat {	// really a stat64
 	dev_t st_dev;
 	mode_t st_mode;
@@ -43,3 +48,19 @@ struct stat {	// really a stat64
 };
 
 #define	O_CREAT	0x0200
+
+void bsdthread_create(void*, M*, G*, void(*)(void));
+void bsdthread_register(void);
+int64 select(int32, void*, void*, void*, struct timeval*);
+
+
+// Mach calls
+
+typedef int32 kern_return_t;
+typedef uint32 mach_port_t;
+
+mach_port_t semcreate(void);
+void semacquire(mach_port_t);
+void semrelease(mach_port_t);
+void semreset(mach_port_t);
+void semdestroy(mach_port_t);

src/runtime/amd64_linux.h

@@ -49,5 +49,5 @@ struct stat {
 // Linux-specific system calls
 int64	futex(uint32*, int32, uint32, struct timespec*, uint32*, uint32);
 int64	clone(int32, void*, M*, G*, void(*)(void));
-int64	select(int32, void*, void*, void*, void*);
+int64	select(int32, void*, void*, void*, struct timeval*);
 

src/runtime/chan.c

@@ -4,9 +4,8 @@
 
 #include "runtime.h"
 
-// TODO locking of select
-
 static	int32	debug	= 0;
+static	Lock		chanlock;
 
 typedef	struct	Hchan	Hchan;
 typedef	struct	Link	Link;
@@ -32,7 +31,6 @@ struct	WaitQ
 
 struct	Hchan
 {
-	Lock;
 	uint32	elemsize;
 	uint32	dataqsiz;		// size of the circular q
 	uint32	qcount;			// total data in the q
@@ -162,7 +160,7 @@ sendchan(Hchan *c, byte *ep, bool *pres)
 		prints("\n");
 	}
 
-	lock(c);
+	lock(&chanlock);
 	if(c->dataqsiz > 0)
 		goto asynch;
 
@@ -173,7 +171,7 @@ sendchan(Hchan *c, byte *ep, bool *pres)
 
 		gp = sg->g;
 		gp->param = sg;
-		unlock(c);
+		unlock(&chanlock);
 		ready(gp);
 
 		if(pres != nil)
@@ -182,7 +180,7 @@ sendchan(Hchan *c, byte *ep, bool *pres)
 	}
 
 	if(pres != nil) {
-		unlock(c);
+		unlock(&chanlock);
 		*pres = false;
 		return;
 	}
@@ -193,13 +191,13 @@ sendchan(Hchan *c, byte *ep, bool *pres)
 	g->param = nil;
 	g->status = Gwaiting;
 	enqueue(&c->sendq, sg);
-	unlock(c);
+	unlock(&chanlock);
 	sys·gosched();
 
-	lock(c);
+	lock(&chanlock);
 	sg = g->param;
 	freesg(c, sg);
-	unlock(c);
+	unlock(&chanlock);
 	return;
 
 asynch:
@@ -208,9 +206,9 @@ asynch:
 		sg = allocsg(c);
 		g->status = Gwaiting;
 		enqueue(&c->sendq, sg);
-		unlock(c);
+		unlock(&chanlock);
 		sys·gosched();
-		lock(c);
+		lock(&chanlock);
 	}
 	if(ep != nil)
 		c->elemalg->copy(c->elemsize, c->senddataq->elem, ep);
@@ -221,10 +219,10 @@ asynch:
 	if(sg != nil) {
 		gp = sg->g;
 		freesg(c, sg);
-		unlock(c);
+		unlock(&chanlock);
 		ready(gp);
 	}else
-		unlock(c);
+		unlock(&chanlock);
 }
 
 static void
@@ -239,7 +237,7 @@ chanrecv(Hchan* c, byte *ep, bool* pres)
 		prints("\n");
 	}
 
-	lock(c);
+	lock(&chanlock);
 	if(c->dataqsiz > 0)
 		goto asynch;
 
@@ -249,7 +247,7 @@ chanrecv(Hchan* c, byte *ep, bool* pres)
 
 		gp = sg->g;
 		gp->param = sg;
-		unlock(c);
+		unlock(&chanlock);
 		ready(gp);
 
 		if(pres != nil)
@@ -258,7 +256,7 @@ chanrecv(Hchan* c, byte *ep, bool* pres)
 	}
 
 	if(pres != nil) {
-		unlock(c);
+		unlock(&chanlock);
 		*pres = false;
 		return;
 	}
@@ -267,14 +265,14 @@ chanrecv(Hchan* c, byte *ep, bool* pres)
 	g->param = nil;
 	g->status = Gwaiting;
 	enqueue(&c->recvq, sg);
-	unlock(c);
+	unlock(&chanlock);
 	sys·gosched();
 
-	lock(c);
+	lock(&chanlock);
 	sg = g->param;
 	c->elemalg->copy(c->elemsize, ep, sg->elem);
 	freesg(c, sg);
-	unlock(c);
+	unlock(&chanlock);
 	return;
 
 asynch:
@@ -282,9 +280,9 @@ asynch:
 		sg = allocsg(c);
 		g->status = Gwaiting;
 		enqueue(&c->recvq, sg);
-		unlock(c);
+		unlock(&chanlock);
 		sys·gosched();
-		lock(c);
+		lock(&chanlock);
 	}
 	c->elemalg->copy(c->elemsize, ep, c->recvdataq->elem);
 	c->recvdataq = c->recvdataq->link;
@@ -293,10 +291,10 @@ asynch:
 	if(sg != nil) {
 		gp = sg->g;
 		freesg(c, sg);
-		unlock(c);
+		unlock(&chanlock);
 		ready(gp);
 	}else
-		unlock(c);
+		unlock(&chanlock);
 }
 
 // chansend1(hchan *chan any, elem any);
@@ -371,12 +369,14 @@ sys·newselect(int32 size, Select *sel)
 	if(size > 1)
 		n = size-1;
 
+	lock(&chanlock);
 	sel = nil;
 	if(size >= 1 && size < nelem(selfree)) {
 		sel = selfree[size];
 		if(sel != nil)
 			selfree[size] = sel->link;
 	}
+	unlock(&chanlock);
 	if(sel == nil)
 		sel = mal(sizeof(*sel) + n*sizeof(sel->scase[0]));
 
@@ -517,6 +517,8 @@ sys·selectgo(Select *sel)
 	p %= sel->ncase;
 	o %= sel->ncase;
 
+	lock(&chanlock);
+
 	// pass 1 - look for something already waiting
 	for(i=0; i<sel->ncase; i++) {
 		cas = &sel->scase[o];
@@ -598,8 +600,10 @@ sys·selectgo(Select *sel)
 	// (rsc) not correct to set Gwaiting after queueing;
 	// might already have been readied.
 	g->status = Gwaiting;
+	unlock(&chanlock);
 	sys·gosched();
 
+	lock(&chanlock);
 	sg = g->param;
 	o = sg->offset;
 	cas = &sel->scase[o];
@@ -629,6 +633,7 @@ sys·selectgo(Select *sel)
 
 asynr:
 asyns:
+	unlock(&chanlock);
 	throw("asyn");
 	return;	// compiler doesn't know throw doesn't return
 
@@ -671,6 +676,7 @@ retc:
 		sel->link = selfree[sel->ncase];
 		selfree[sel->ncase] = sel;
 	}
+	unlock(&chanlock);
 
 	sys·setcallerpc(&sel, cas->pc);
 	as = (byte*)&sel + cas->so;

src/runtime/proc.c

@@ -341,8 +341,6 @@ scheduler(void)
 {
 	G* gp;
 
-	// Initialization.
-	m->procid = getprocid();
 	lock(&sched);
 
 	if(gosave(&m->sched)){
@@ -472,7 +470,7 @@ oldstack(void)
 		mcpy(top->oldsp+16, sp, siz2);
 	}
 
-	// call  no more functions after this point - limit register disagrees with R15
+	// call  no more functions after this point - stackguard disagrees with SP
 	m->curg->stackbase = top->oldbase;
 	m->curg->stackguard = top->oldguard;
 	m->morestack.SP = top->oldsp+8;

src/runtime/rt0_amd64.s

@@ -33,6 +33,7 @@ TEXT	_rt0_amd64(SB),7,$-8
 	MOVQ	24(SP), AX		// copy argv
 	MOVQ	AX, 8(SP)
 	CALL	args(SB)
+	CALL	osinit(SB)
 	CALL	schedinit(SB)
 	CALL	main·init_function(SB) // initialization
 

src/runtime/rt1_amd64_darwin.c

@@ -3,6 +3,7 @@
 // license that can be found in the LICENSE file.
 
 #include "runtime.h"
+#include "amd64_darwin.h"
 #include "signals.h"
 
 typedef uint64 __uint64_t;
@@ -185,53 +186,494 @@ unimplemented(int8 *name)
 void
 sys·sleep(int64 ms)
 {
-	unimplemented("sleep");
+	struct timeval tv;
+
+	tv.tv_sec = ms/1000;
+	tv.tv_usec = ms%1000 * 1000;
+	select(0, nil, nil, nil, &tv);
+}
+
+// Thread-safe allocation of a semaphore.
+// Psema points at a kernel semaphore key.
+// It starts out zero, meaning no semaphore.
+// Fill it in, being careful of others calling initsema
+// simultaneously.
+static void
+initsema(uint32 *psema)
+{
+	uint32 sema;
+
+	if(*psema != 0)	// already have one
+		return;
+
+	sema = semcreate();
+	if(!cas(psema, 0, sema)){
+		// Someone else filled it in.  Use theirs.
+		semdestroy(sema);
+		return;
+	}
+}
+
+
+// Atomic add and return new value.
+static uint32
+xadd(uint32 volatile *val, int32 delta)
+{
+	uint32 oval, nval;
+
+	for(;;){
+		oval = *val;
+		nval = oval + delta;
+		if(cas(val, oval, nval))
+			return nval;
+	}
 }
 
+
+// Blocking locks.
+
+// Implement Locks, using semaphores.
+// l->key is the number of threads who want the lock.
+// In a race, one thread increments l->key from 0 to 1
+// and the others increment it from >0 to >1.  The thread
+// who does the 0->1 increment gets the lock, and the
+// others wait on the semaphore.  When the 0->1 thread
+// releases the lock by decrementing l->key, l->key will
+// be >0, so it will increment the semaphore to wake up
+// one of the others.  This is the same algorithm used
+// in Plan 9's user-space locks.
+//
+// Note that semaphores are never destroyed (the kernel
+// will clean up when the process exits).  We assume for now
+// that Locks are only used for long-lived structures like M and G.
+
 void
 lock(Lock *l)
 {
-	if(cas(&l->key, 0, 1))
-		return;
-	unimplemented("lock wait");
+	// Allocate semaphore if needed.
+	if(l->sema == 0)
+		initsema(&l->sema);
+
+	if(xadd(&l->key, 1) > 1)	// someone else has it; wait
+		semacquire(l->sema);
 }
 
 void
 unlock(Lock *l)
 {
-	if(cas(&l->key, 1, 0))
-		return;
-	unimplemented("unlock wakeup");
+	if(xadd(&l->key, -1) > 0)	// someone else is waiting
+		semrelease(l->sema);
 }
 
+
+// Event notifications.
 void
 noteclear(Note *n)
 {
-	n->lock.key = 0;
-	lock(&n->lock);
+	n->wakeup = 0;
 }
 
 void
 notesleep(Note *n)
 {
-	lock(&n->lock);
-	unlock(&n->lock);
+	if(n->sema == 0)
+		initsema(&n->sema);
+	while(!n->wakeup)
+		semacquire(n->sema);
 }
 
 void
 notewakeup(Note *n)
 {
-	unlock(&n->lock);
+	if(n->sema == 0)
+		initsema(&n->sema);
+	n->wakeup = 1;
+	semrelease(n->sema);
 }
 
+
+// BSD interface for threading.
 void
-newosproc(M *mm, G *gg, void *stk, void (*fn)(void))
+osinit(void)
+{
+	// Register our thread-creation callback (see sys_amd64_darwin.s).
+	bsdthread_register();
+}
+
+void
+newosproc(M *m, G *g, void *stk, void (*fn)(void))
+{
+	bsdthread_create(stk, m, g, fn);
+}
+
+
+// Mach IPC, to get at semaphores
+// Definitions are in /usr/include/mach on a Mac.
+
+static void
+macherror(kern_return_t r, int8 *fn)
 {
-	unimplemented("newosproc");
+	prints("mach error ");
+	prints(fn);
+	prints(": ");
+	sys·printint(r);
+	prints("\n");
+	throw("mach error");
 }
 
-int32
-getprocid(void)
+enum
+{
+	DebugMach = 0
+};
+
+typedef int32 mach_msg_option_t;
+typedef uint32 mach_msg_bits_t;
+typedef uint32 mach_msg_id_t;
+typedef uint32 mach_msg_size_t;
+typedef uint32 mach_msg_timeout_t;
+typedef uint32 mach_port_name_t;
+typedef uint64 mach_vm_address_t;
+
+typedef struct mach_msg_header_t mach_msg_header_t;
+typedef struct mach_msg_body_t mach_msg_body_t;
+typedef struct mach_msg_port_descriptor_t mach_msg_port_descriptor_t;
+typedef struct NDR_record_t NDR_record_t;
+
+enum
+{
+	MACH_MSG_TYPE_MOVE_RECEIVE = 16,
+	MACH_MSG_TYPE_MOVE_SEND = 17,
+	MACH_MSG_TYPE_MOVE_SEND_ONCE = 18,
+	MACH_MSG_TYPE_COPY_SEND = 19,
+	MACH_MSG_TYPE_MAKE_SEND = 20,
+	MACH_MSG_TYPE_MAKE_SEND_ONCE = 21,
+	MACH_MSG_TYPE_COPY_RECEIVE = 22,
+
+	MACH_MSG_PORT_DESCRIPTOR = 0,
+	MACH_MSG_OOL_DESCRIPTOR = 1,
+	MACH_MSG_OOL_PORTS_DESCRIPTOR = 2,
+	MACH_MSG_OOL_VOLATILE_DESCRIPTOR = 3,
+
+	MACH_MSGH_BITS_COMPLEX = 0x80000000,
+
+	MACH_SEND_MSG = 1,
+	MACH_RCV_MSG = 2,
+	MACH_RCV_LARGE = 4,
+
+	MACH_SEND_TIMEOUT = 0x10,
+	MACH_SEND_INTERRUPT = 0x40,
+	MACH_SEND_CANCEL = 0x80,
+	MACH_SEND_ALWAYS = 0x10000,
+	MACH_SEND_TRAILER = 0x20000,
+	MACH_RCV_TIMEOUT = 0x100,
+	MACH_RCV_NOTIFY = 0x200,
+	MACH_RCV_INTERRUPT = 0x400,
+	MACH_RCV_OVERWRITE = 0x1000,
+};
+
+mach_port_t mach_task_self(void);
+mach_port_t mach_thread_self(void);
+
+#pragma pack on
+struct mach_msg_header_t
+{
+	mach_msg_bits_t bits;
+	mach_msg_size_t size;
+	mach_port_t remote_port;
+	mach_port_t local_port;
+	mach_msg_size_t reserved;
+	mach_msg_id_t id;
+};
+
+struct mach_msg_body_t
+{
+	uint32 descriptor_count;
+};
+
+struct mach_msg_port_descriptor_t
+{
+	mach_port_t name;
+	uint32 pad1;
+	uint16 pad2;
+	uint8 disposition;
+	uint8 type;
+};
+
+enum
+{
+	NDR_PROTOCOL_2_0 = 0,
+	NDR_INT_BIG_ENDIAN = 0,
+	NDR_INT_LITTLE_ENDIAN = 1,
+	NDR_FLOAT_IEEE = 0,
+	NDR_CHAR_ASCII = 0
+};
+
+struct NDR_record_t
+{
+	uint8 mig_vers;
+	uint8 if_vers;
+	uint8 reserved1;
+	uint8 mig_encoding;
+	uint8 int_rep;
+	uint8 char_rep;
+	uint8 float_rep;
+	uint8 reserved2;
+};
+#pragma pack off
+
+static NDR_record_t zerondr;
+
+#define MACH_MSGH_BITS(a, b) ((a) | ((b)<<8))
+
+// Mach system calls (in sys_amd64_darwin.s)
+kern_return_t mach_msg_trap(mach_msg_header_t*,
+	mach_msg_option_t, mach_msg_size_t, mach_msg_size_t,
+	mach_port_name_t, mach_msg_timeout_t, mach_port_name_t);
+mach_port_t mach_reply_port(void);
+mach_port_t mach_task_self(void);
+mach_port_t mach_thread_self(void);
+
+static kern_return_t
+mach_msg(mach_msg_header_t *h,
+	mach_msg_option_t op,
+	mach_msg_size_t send_size,
+	mach_msg_size_t rcv_size,
+	mach_port_name_t rcv_name,
+	mach_msg_timeout_t timeout,
+	mach_port_name_t notify)
+{
+	// TODO: Loop on interrupt.
+	return mach_msg_trap(h, op, send_size, rcv_size, rcv_name, timeout, notify);
+}
+
+
+// Mach RPC (MIG)
+// I'm not using the Mach names anymore.  They're too long.
+
+enum
+{
+	MinMachMsg = 48,
+	Reply = 100,
+};
+
+#pragma pack on
+typedef struct CodeMsg CodeMsg;
+struct CodeMsg
 {
+	mach_msg_header_t h;
+	NDR_record_t NDR;
+	kern_return_t code;
+};
+#pragma pack off
+
+static kern_return_t
+machcall(mach_msg_header_t *h, int32 maxsize, int32 rxsize)
+{
+	uint32 *p;
+	int32 i, ret, id;
+	mach_port_t port;
+	CodeMsg *c;
+
+	if((port = m->machport) == 0){
+		port = mach_reply_port();
+		m->machport = port;
+	}
+
+	h->bits |= MACH_MSGH_BITS(MACH_MSG_TYPE_COPY_SEND, MACH_MSG_TYPE_MAKE_SEND_ONCE);
+	h->local_port = port;
+	h->reserved = 0;
+	id = h->id;
+
+	if(DebugMach){
+		p = (uint32*)h;
+		prints("send:\t");
+		for(i=0; i<h->size/sizeof(p[0]); i++){
+			prints(" ");
+			sys·printpointer((void*)p[i]);
+			if(i%8 == 7)
+				prints("\n\t");
+		}
+		if(i%8)
+			prints("\n");
+	}
+
+	ret = mach_msg(h, MACH_SEND_MSG|MACH_RCV_MSG,
+		h->size, maxsize, port, 0, 0);
+	if(ret != 0){
+		if(DebugMach){
+			prints("mach_msg error ");
+			sys·printint(ret);
+			prints("\n");
+		}
+		return ret;
+	}
+
+	if(DebugMach){
+		p = (uint32*)h;
+		prints("recv:\t");
+		for(i=0; i<h->size/sizeof(p[0]); i++){
+			prints(" ");
+			sys·printpointer((void*)p[i]);
+			if(i%8 == 7)
+				prints("\n\t");
+		}
+		if(i%8)
+			prints("\n");
+	}
+
+	if(h->id != id+Reply){
+		if(DebugMach){
+			prints("mach_msg reply id mismatch ");
+			sys·printint(h->id);
+			prints(" != ");
+			sys·printint(id+Reply);
+			prints("\n");
+		}
+		return -303;	// MIG_REPLY_MISMATCH
+	}
+
+	// Look for a response giving the return value.
+	// Any call can send this back with an error,
+	// and some calls only have return values so they
+	// send it back on success too.  I don't quite see how
+	// you know it's one of these and not the full response
+	// format, so just look if the message is right.
+	c = (CodeMsg*)h;
+	if(h->size == sizeof(CodeMsg)
+	&& !(h->bits & MACH_MSGH_BITS_COMPLEX)){
+		if(DebugMach){
+			prints("mig result ");
+			sys·printint(c->code);
+			prints("\n");
+		}
+		return c->code;
+	}
+
+	if(h->size != rxsize){
+		if(DebugMach){
+			prints("mach_msg reply size mismatch ");
+			sys·printint(h->size);
+			prints(" != ");
+			sys·printint(rxsize);
+			prints("\n");
+		}
+		return -307;	// MIG_ARRAY_TOO_LARGE
+	}
+
 	return 0;
 }
+
+
+// Semaphores!
+
+enum
+{
+	Tsemcreate = 3418,
+	Rsemcreate = Tsemcreate + Reply,
+
+	Tsemdestroy = 3419,
+	Rsemdestroy = Tsemdestroy + Reply,
+};
+
+typedef struct TsemcreateMsg TsemcreateMsg;
+typedef struct RsemcreateMsg RsemcreateMsg;
+typedef struct TsemdestroyMsg TsemdestroyMsg;
+// RsemdestroyMsg = CodeMsg
+
+#pragma pack on
+struct TsemcreateMsg
+{
+	mach_msg_header_t h;
+	NDR_record_t ndr;
+	int32 policy;
+	int32 value;
+};
+
+struct RsemcreateMsg
+{
+	mach_msg_header_t h;
+	mach_msg_body_t body;
+	mach_msg_port_descriptor_t semaphore;
+};
+
+struct TsemdestroyMsg
+{
+	mach_msg_header_t h;
+	mach_msg_body_t body;
+	mach_msg_port_descriptor_t semaphore;
+};
+#pragma pack off
+
+mach_port_t
+semcreate(void)
+{
+	union {
+		TsemcreateMsg tx;
+		RsemcreateMsg rx;
+		uint8 pad[MinMachMsg];
+	} m;
+	kern_return_t r;
+
+	m.tx.h.bits = 0;
+	m.tx.h.size = sizeof(m.tx);
+	m.tx.h.remote_port = mach_task_self();
+	m.tx.h.id = Tsemcreate;
+	m.tx.ndr = zerondr;
+
+	m.tx.policy = 0;	// 0 = SYNC_POLICY_FIFO
+	m.tx.value = 0;
+
+	if((r = machcall(&m.tx.h, sizeof m, sizeof(m.rx))) != 0)
+		macherror(r, "semaphore_create");
+	if(m.rx.body.descriptor_count != 1)
+		unimplemented("semcreate desc count");
+	return m.rx.semaphore.name;
+}
+
+void
+semdestroy(mach_port_t sem)
+{
+	union {
+		TsemdestroyMsg tx;
+		uint8 pad[MinMachMsg];
+	} m;
+	kern_return_t r;
+
+	m.tx.h.bits = MACH_MSGH_BITS_COMPLEX;
+	m.tx.h.size = sizeof(m.tx);
+	m.tx.h.remote_port = mach_task_self();
+	m.tx.h.id = Tsemdestroy;
+	m.tx.body.descriptor_count = 1;
+	m.tx.semaphore.name = sem;
+	m.tx.semaphore.disposition = MACH_MSG_TYPE_MOVE_SEND;
+	m.tx.semaphore.type = 0;
+
+	if((r = machcall(&m.tx.h, sizeof m, 0)) != 0)
+		macherror(r, "semaphore_destroy");
+}
+
+// The other calls have simple system call traps
+// in sys_amd64_darwin.s
+kern_return_t mach_semaphore_wait(uint32 sema);
+kern_return_t mach_semaphore_timedwait(uint32 sema, uint32 sec, uint32 nsec);
+kern_return_t mach_semaphore_signal(uint32 sema);
+kern_return_t mach_semaphore_signal_all(uint32 sema);
+
+void
+semacquire(mach_port_t sem)
+{
+	kern_return_t r;
+
+	if((r = mach_semaphore_wait(sem)) != 0)
+		macherror(r, "semaphore_wait");
+}
+
+void
+semrelease(mach_port_t sem)
+{
+	kern_return_t r;
+
+	if((r = mach_semaphore_signal(sem)) != 0)
+		macherror(r, "semaphore_signal");
+}
+

src/runtime/rt1_amd64_linux.c

@@ -427,3 +427,7 @@ sys·sleep(int64 ms)
 	select(0, nil, nil, nil, &tv);
 }
 
+void
+osinit(void)
+{
+}

src/runtime/runtime.h

@@ -44,7 +44,7 @@ typedef	struct	M		M;
 typedef	struct	Stktop		Stktop;
 typedef	struct	Alg		Alg;
 typedef	struct	Lock		Lock;
-typedef	struct	Note	Note;
+typedef	union	Note	Note;
 typedef	struct	Mem		Mem;
 
 /*
@@ -78,10 +78,17 @@ enum
 struct	Lock
 {
 	uint32	key;
+	uint32	sema;	// for OS X
 };
-struct	Note
+union	Note
 {
-	Lock	lock;
+	struct {	// Linux
+		Lock	lock;
+	};
+	struct {	// OS X
+		int32	wakeup;
+		uint32	sema;
+	};
 };
 struct String
 {
@@ -149,6 +156,7 @@ struct	M
 	G*	g0;		// g0 w interrupt stack - must not move
 	uint64	morearg;	// arg to morestack - must not move
 	uint64	cret;		// return value from C - must not move
+	uint64	procid;	// for debuggers - must not move
 	G*	curg;		// current running goroutine
 	G*	lastg;		// last running goroutine - to emulate fifo
 	Gobuf	sched;
@@ -159,8 +167,8 @@ struct	M
 	Note	havenextg;
 	G*	nextg;
 	M*	schedlink;
-	int32	procid;	// for debuggers
 	Mem	mem;
+	uint32	machport;	// Return address for Mach IPC (OS X)
 };
 struct	Stktop
 {
@@ -239,7 +247,6 @@ void	ready(G*);
 byte*	getenv(int8*);
 int32	atoi(byte*);
 void	newosproc(M *m, G *g, void *stk, void (*fn)(void));
-int32	getprocid(void);
 
 /*
  * mutual exclusion locks.  in the uncontended case,

src/runtime/sys_amd64_darwin.s

@@ -4,12 +4,11 @@
 
 //
 // System calls and other sys.stuff for AMD64, Darwin
+// See http://fxr.watson.org/fxr/source/bsd/kern/syscalls.c?v=xnu-1228
+// or /usr/include/sys/syscall.h (on a Mac) for system call numbers.
 //
 
-// TODO(rsc): Either sys·exit or exit1 is wrong!
-// It looks like sys·exit is correct (exits the entire program)
-// and exit1 should be mimicking the OS X library routine
-// __bsdthread_terminate.
+// Exit the entire program (like C exit)
 TEXT	sys·exit(SB),7,$-8
 	MOVL	8(SP), DI		// arg 1 exit status
 	MOVL	$(0x2000000+1), AX	// syscall entry
@@ -17,9 +16,11 @@ TEXT	sys·exit(SB),7,$-8
 	CALL	notok(SB)
 	RET
 
+// Exit this OS thread (like pthread_exit, which eventually
+// calls __bsdthread_terminate).
 TEXT	exit1(SB),7,$-8
 	MOVL	8(SP), DI		// arg 1 exit status
-	MOVL	$(0x2000000+1), AX	// syscall entry
+	MOVL	$(0x2000000+361), AX	// syscall entry
 	SYSCALL
 	CALL	notok(SB)
 	RET
@@ -130,3 +131,127 @@ TEXT	sys·setcallerpc+0(SB),7,$0
 	MOVQ	x+8(FP), BX
 	MOVQ	BX, -8(AX)		// set calling pc
 	RET
+
+// void bsdthread_create(void *stk, M *m, G *g, void (*fn)(void))
+TEXT bsdthread_create(SB),7,$-8
+	// Set up arguments to bsdthread_create system call.
+	// The ones in quotes pass through to the thread callback
+	// uninterpreted, so we can put whatever we want there.
+	MOVQ	fn+32(SP), DI	// "func"
+	MOVQ	m+16(SP), SI	// "arg"
+	MOVQ	stk+8(SP), DX	// stack
+	MOVQ	g+24(SP), R10	// "pthread"
+	MOVQ	$0, R10	// flags
+	MOVQ	$(0x2000000+360), AX	// bsdthread_create
+	SYSCALL
+	JCC 2(PC)
+	CALL	notok(SB)
+	RET
+
+// The thread that bsdthread_create creates starts executing here,
+// because we registered this function using bsdthread_register
+// at startup.
+//	DI = "pthread" (= g)
+//	SI = mach thread port
+//	DX = "func" (= fn)
+//	CX = "arg" (= m)
+//	R8 = stack
+//	R9 = flags (= 0)
+//	SP = stack - C_64_REDZONE_LEN (= stack - 128)
+TEXT bsdthread_start(SB),7,$-8
+	MOVQ	CX, R14	// m
+	MOVQ	DI, R15	// g
+	MOVQ	SI, 24(R14)	// thread port is m->procid
+	CALL	DX	// fn
+	CALL	exit1(SB)
+	RET
+
+// void bsdthread_register(void)
+// registers callbacks for threadstart (see bsdthread_create above
+// and wqthread and pthsize (not used).  returns 0 on success.
+TEXT bsdthread_register(SB),7,$-8
+	MOVQ	$bsdthread_start(SB), DI	// threadstart
+	MOVQ	$0, SI	// wqthread, not used by us
+	MOVQ	$0, DX	// pthsize, not used by us
+	MOVQ	$(0x2000000+366), AX	// bsdthread_register
+	SYSCALL
+	JCC 2(PC)
+	CALL	notok(SB)
+	RET
+
+// int64 select(int32, void*, void*, void*, void*)
+TEXT select(SB),7,$0
+	MOVL	8(SP), DI
+	MOVQ	16(SP), SI
+	MOVQ	24(SP), DX
+	MOVQ	32(SP), R10
+	MOVQ	40(SP), R8
+	MOVL	$(0x2000000+407), AX	// select_nocancel
+	SYSCALL
+	RET
+
+// Mach system calls use 0x1000000 instead of the BSD's 0x2000000.
+
+// uint32 mach_msg_trap(void*, uint32, uint32, uint32, uint32, uint32, uint32)
+TEXT mach_msg_trap(SB),7,$0
+	MOVQ	8(SP), DI
+	MOVL	16(SP), SI
+	MOVL	20(SP), DX
+	MOVL	24(SP), R10
+	MOVL	28(SP), R8
+	MOVL	32(SP), R9
+	MOVL	36(SP), R11
+	PUSHQ	R11	// seventh arg, on stack
+	MOVL	$(0x1000000+31), AX	// mach_msg_trap
+	SYSCALL
+	POPQ	R11
+	RET
+
+TEXT mach_task_self(SB),7,$0
+	MOVL	$(0x1000000+28), AX	// task_self_trap
+	SYSCALL
+	RET
+
+TEXT mach_thread_self(SB),7,$0
+	MOVL	$(0x1000000+27), AX	// thread_self_trap
+	SYSCALL
+	RET
+
+TEXT mach_reply_port(SB),7,$0
+	MOVL	$(0x1000000+26), AX	// mach_reply_port
+	SYSCALL
+	RET
+
+// Mach provides trap versions of the semaphore ops,
+// instead of requiring the use of RPC.
+
+// uint32 mach_semaphore_wait(uint32)
+TEXT mach_semaphore_wait(SB),7,$0
+	MOVL	8(SP), DI
+	MOVL	$(0x1000000+36), AX	// semaphore_wait_trap
+	SYSCALL
+	RET
+
+// uint32 mach_semaphore_timedwait(uint32, uint32, uint32)
+TEXT mach_semaphore_timedwait(SB),7,$0
+	MOVL	8(SP), DI
+	MOVL	12(SP), SI
+	MOVL	16(SP), DX
+	MOVL	$(0x1000000+38), AX	// semaphore_timedwait_trap
+	SYSCALL
+	RET
+
+// uint32 mach_semaphore_signal(uint32)
+TEXT mach_semaphore_signal(SB),7,$0
+	MOVL	8(SP), DI
+	MOVL	$(0x1000000+33), AX	// semaphore_signal_trap
+	SYSCALL
+	RET
+
+// uint32 mach_semaphore_signal_all(uint32)
+TEXT mach_semaphore_signal_all(SB),7,$0
+	MOVL	8(SP), DI
+	MOVL	$(0x1000000+34), AX	// semaphore_signal_all_trap
+	SYSCALL
+	RET
+

src/runtime/sys_amd64_linux.s

@@ -162,10 +162,17 @@ TEXT clone(SB),7,$0
 	JEQ	2(PC)
 	RET
 
-	// In child, call fn on new stack
+	// In child, set up new stack
 	MOVQ	SI, SP
 	MOVQ	R8, R14	// m
 	MOVQ	R9, R15	// g
+	
+	// Initialize m->procid to Linux tid
+	MOVL	$186, AX	// gettid
+	SYSCALL
+	MOVQ	AX, 24(R14)
+	
+	// Call fn
 	CALL	R12
 
 	// It shouldn't return.  If it does, exi
@@ -174,7 +181,7 @@ TEXT clone(SB),7,$0
 	SYSCALL
 	JMP	-3(PC)	// keep exiting
 
-// int64 select(int32, void*, void*, void*, void*)
+// int64 select(int32, void*, void*, void*, struct timeval*)
 TEXT select(SB),7,$0
 	MOVL	8(SP), DI
 	MOVQ	16(SP), SI
@@ -185,16 +192,3 @@ TEXT select(SB),7,$0
 	SYSCALL
 	RET
 
-// Linux allocates each thread its own pid, like Plan 9.
-// But the getpid() system call returns the pid of the
-// original thread (the one that exec started with),
-// no matter which thread asks.  This system call,
-// which Linux calls gettid, returns the actual pid of
-// the calling thread, not the fake one.
-//
-// int32 getprocid(void)
-TEXT getprocid(SB),7,$0
-	MOVL	$186, AX
-	SYSCALL
-	RET
-