add support for ref counts to memory allocator.

mark and sweep, stop the world garbage collector
(intermediate step in the way to ref counting).
can run pretty with an explicit gc after each file.

R=r
DELTA=502  (346 added, 143 deleted, 13 changed)
OCL=20630
CL=20635

src/runtime/Makefile

@@ -22,11 +22,11 @@ LIBOFILES=\
 	chan.$O\
 	iface.$O\
 	array.$O\
+	mem.$O\
 	print.$O\
 	rune.$O\
 	proc.$O\
 	sema.$O\
-	stack.$O\
 	string.$O\
 	symtab.$O\
 	sys_file.$O\

src/runtime/mem.c

+// Copyright 2009 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 "runtime.h"
+
+// Stubs for memory management.
+// In a separate file so they can be overridden during testing of gc.
+
+enum
+{
+	NHUNK		= 20<<20,
+
+	PROT_NONE	= 0x00,
+	PROT_READ	= 0x01,
+	PROT_WRITE	= 0x02,
+	PROT_EXEC	= 0x04,
+
+	MAP_FILE	= 0x0000,
+	MAP_SHARED	= 0x0001,
+	MAP_PRIVATE	= 0x0002,
+	MAP_FIXED	= 0x0010,
+	MAP_ANON	= 0x1000,	// not on Linux - TODO(rsc)
+};
+
+void*
+stackalloc(uint32 n)
+{
+	return mal(n);
+}
+
+void
+stackfree(void*)
+{
+}
+
+// Convenient wrapper around mmap.
+static void*
+brk(uint32 n)
+{
+	byte *v;
+
+	v = sys·mmap(nil, n, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 0, 0);
+	m->mem.nmmap += n;
+	return v;
+}
+
+// Allocate n bytes of memory.  Note that this gets used
+// to allocate new stack segments, so at each call to a function
+// you have to ask yourself "would it be okay to call mal recursively
+// right here?"  The answer is yes unless we're in the middle of
+// editing the malloc state in m->mem.
+void*
+mal(uint32 n)
+{
+	byte* v;
+
+	// round to keep everything 64-bit aligned
+	n = rnd(n, 8);
+
+	// be careful.  calling any function might invoke
+	// mal to allocate more stack.
+	if(n > NHUNK) {
+		v = brk(n);
+	} else {
+		// allocate a new hunk if this one is too small
+		if(n > m->mem.nhunk) {
+			// here we're in the middle of editing m->mem
+			// (we're about to overwrite m->mem.hunk),
+			// so we can't call brk - it might call mal to grow the
+			// stack, and the recursive call would allocate a new
+			// hunk, and then once brk returned we'd immediately
+			// overwrite that hunk with our own.
+			// (the net result would be a memory leak, not a crash.)
+			// so we have to call sys·mmap directly - it is written
+			// in assembly and tagged not to grow the stack.
+			m->mem.hunk =
+				sys·mmap(nil, NHUNK, PROT_READ|PROT_WRITE,
+					MAP_ANON|MAP_PRIVATE, 0, 0);
+			m->mem.nhunk = NHUNK;
+			m->mem.nmmap += NHUNK;
+		}
+		v = m->mem.hunk;
+		m->mem.hunk += n;
+		m->mem.nhunk -= n;
+	}
+	m->mem.nmal += n;
+	return v;
+}
+
+void
+sys·mal(uint32 n, uint8 *ret)
+{
+	ret = mal(n);
+	FLUSH(&ret);
+}

src/runtime/proc.c

@@ -53,9 +53,12 @@ struct Sched {
 	int32 mcount;	// number of ms that have been created
 	int32 mcpu;	// number of ms executing on cpu
 	int32 mcpumax;	// max number of ms allowed on cpu
+	int32 gomaxprocs;
 	int32 msyscall;	// number of ms in system calls
 
 	int32 predawn;	// running initialization, don't run new gs.
+
+	Note	stopped;	// one g can wait here for ms to stop
 };
 
 Sched sched;
@@ -91,10 +94,11 @@ schedinit(void)
 	int32 n;
 	byte *p;
 
-	sched.mcpumax = 1;
+	sched.gomaxprocs = 1;
 	p = getenv("GOMAXPROCS");
 	if(p != nil && (n = atoi(p)) != 0)
-		sched.mcpumax = n;
+		sched.gomaxprocs = n;
+	sched.mcpumax = sched.gomaxprocs;
 	sched.mcount = 1;
 	sched.predawn = 1;
 }
@@ -134,7 +138,7 @@ malg(int32 stacksize)
 
 	// 160 is the slop amount known to the stack growth code
 	g = mal(sizeof(G));
-	stk = mal(160 + stacksize);
+	stk = stackalloc(160 + stacksize);
 	g->stack0 = stk;
 	g->stackguard = stk + 160;
 	g->stackbase = stk + 160 + stacksize;
@@ -348,6 +352,7 @@ nextgandunlock(void)
 		throw("all goroutines are asleep - deadlock!");
 	m->nextg = nil;
 	noteclear(&m->havenextg);
+	notewakeup(&sched.stopped);
 	unlock(&sched);
 
 	notesleep(&m->havenextg);
@@ -362,6 +367,33 @@ nextgandunlock(void)
 	return gp;
 }
 
+// TODO(rsc): Remove. This is only temporary,
+// for the mark and sweep collector.
+void
+stoptheworld(void)
+{
+	lock(&sched);
+	sched.mcpumax = 1;
+	while(sched.mcpu > 1) {
+		noteclear(&sched.stopped);
+		unlock(&sched);
+		notesleep(&sched.stopped);
+		lock(&sched);
+	}
+	unlock(&sched);
+}
+
+// TODO(rsc): Remove. This is only temporary,
+// for the mark and sweep collector.
+void
+starttheworld(void)
+{
+	lock(&sched);
+	sched.mcpumax = sched.gomaxprocs;
+	matchmg();
+	unlock(&sched);
+}
+
 // Called to start an M.
 void
 mstart(void)
@@ -500,11 +532,15 @@ sys·entersyscall(uint64 callerpc, int64 trap)
 		unlock(&debuglock);
 	}
 	lock(&sched);
+	g->status = Gsyscall;
 	sched.mcpu--;
 	sched.msyscall++;
 	if(sched.gwait != 0)
 		matchmg();
 	unlock(&sched);
+	// leave SP around for gc; poison PC to make sure it's not used
+	g->sched.SP = (byte*)&callerpc;
+	g->sched.PC = (byte*)0xdeadbeef;
 }
 
 // The goroutine g exited its system call.
@@ -521,6 +557,7 @@ sys·exitsyscall(void)
 	}
 
 	lock(&sched);
+	g->status = Grunning;
 	sched.msyscall--;
 	sched.mcpu++;
 	// Fast path - if there's room for this m, we're done.

src/runtime/runtime.c

@@ -48,22 +48,6 @@ sys·throwreturn(void)
 	throw("no return at end of a typed function");
 }
 
-enum
-{
-	NHUNK		= 20<<20,
-
-	PROT_NONE	= 0x00,
-	PROT_READ	= 0x01,
-	PROT_WRITE	= 0x02,
-	PROT_EXEC	= 0x04,
-
-	MAP_FILE	= 0x0000,
-	MAP_SHARED	= 0x0001,
-	MAP_PRIVATE	= 0x0002,
-	MAP_FIXED	= 0x0010,
-	MAP_ANON	= 0x1000,	// not on Linux - TODO(rsc)
-};
-
 void
 throw(int8 *s)
 {
@@ -129,67 +113,6 @@ rnd(uint32 n, uint32 m)
 	return n;
 }
 
-// Convenient wrapper around mmap.
-static void*
-brk(uint32 n)
-{
-	byte *v;
-
-	v = sys·mmap(nil, n, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 0, 0);
-	m->mem.nmmap += n;
-	return v;
-}
-
-// Allocate n bytes of memory.  Note that this gets used
-// to allocate new stack segments, so at each call to a function
-// you have to ask yourself "would it be okay to call mal recursively
-// right here?"  The answer is yes unless we're in the middle of
-// editing the malloc state in m->mem.
-void*
-mal(uint32 n)
-{
-	byte* v;
-
-	// round to keep everything 64-bit aligned
-	n = rnd(n, 8);
-
-	// be careful.  calling any function might invoke
-	// mal to allocate more stack.
-	if(n > NHUNK) {
-		v = brk(n);
-	} else {
-		// allocate a new hunk if this one is too small
-		if(n > m->mem.nhunk) {
-			// here we're in the middle of editing m->mem
-			// (we're about to overwrite m->mem.hunk),
-			// so we can't call brk - it might call mal to grow the
-			// stack, and the recursive call would allocate a new
-			// hunk, and then once brk returned we'd immediately
-			// overwrite that hunk with our own.
-			// (the net result would be a memory leak, not a crash.)
-			// so we have to call sys·mmap directly - it is written
-			// in assembly and tagged not to grow the stack.
-			m->mem.hunk =
-				sys·mmap(nil, NHUNK, PROT_READ|PROT_WRITE,
-					MAP_ANON|MAP_PRIVATE, 0, 0);
-			m->mem.nhunk = NHUNK;
-			m->mem.nmmap += NHUNK;
-		}
-		v = m->mem.hunk;
-		m->mem.hunk += n;
-		m->mem.nhunk -= n;
-	}
-	m->mem.nmal += n;
-	return v;
-}
-
-void
-sys·mal(uint32 n, uint8 *ret)
-{
-	ret = mal(n);
-	FLUSH(&ret);
-}
-
 static	uint64	uvnan		= 0x7FF0000000000001ULL;
 static	uint64	uvinf		= 0x7FF0000000000000ULL;
 static	uint64	uvneginf	= 0xFFF0000000000000ULL;

src/runtime/runtime.h

@@ -65,6 +65,7 @@ enum
 	Gidle,
 	Grunnable,
 	Grunning,
+	Gsyscall,
 	Gwaiting,
 	Gmoribund,
 	Gdead,
@@ -280,6 +281,11 @@ int32	funcline(Func*, uint64);
 void*	stackalloc(uint32);
 void	stackfree(void*);
 
+// TODO(rsc): Remove. These are only temporary,
+// for the mark and sweep collector.
+void	stoptheworld(void);
+void	starttheworld(void);
+
 /*
  * mutual exclusion locks.  in the uncontended case,
  * as fast as spin locks (just a few user-level instructions),
@@ -340,3 +346,5 @@ bool	isNaN(float64);
 void	sys·readfile(string, string, bool);
 void	sys·bytestorune(byte*, int32, int32, int32, int32);
 void	sys·stringtorune(string, int32, int32, int32);
+void	sys·semacquire(uint32*);
+void	sys·semrelease(uint32*);

src/runtime/stack.c

-// Copyright 2009 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 "runtime.h"
-
-// Stubs for stack management.
-// In a separate file so they can be overridden during testing of gc.
-
-void*
-stackalloc(uint32 n)
-{
-	return mal(n);
-}
-
-void
-stackfree(void*)
-{
-}

usr/rsc/mem/Makefile

@@ -20,6 +20,8 @@ default: $(TARG)
 OFILES=\
 	allocator.$O\
 	malloc.$O\
+	mem.$O\
+	ms.$O\
 	pagemap.$O\
 	triv.$O\
 
@@ -32,6 +34,5 @@ test%: test%.$O $(OFILES)
 clean:
 	rm -f *.$O $(TARG)
 
-runtime: allocator.$O malloc.$O pagemap.$O triv.$O stack.$O
+runtime: $(OFILES)
 	6ar grc $(GOROOT)/lib/lib_$(GOARCH)_$(GOOS).a $^
-

usr/rsc/mem/allocator.go

@@ -11,3 +11,5 @@ export var footprint int64
 export var frozen bool
 export func testsizetoclass()
 export var allocated int64
+export func find(uint64) (obj *byte, size int64, ref *int32, ok bool)
+export func gc()

usr/rsc/mem/malloc.c

@@ -18,36 +18,8 @@
 
 #include "malloc.h"
 
-typedef struct Span Span;
-typedef struct Central Central;
-
-// A Span contains metadata about a range of pages.
-enum {
-	SpanInUse = 0,	// span has been handed out by allocator
-	SpanFree = 1,	// span is in central free list
-};
-struct Span
-{
-	Span *next;	// in free lists
-	byte *base;	// first byte in span
-	uintptr length;	// number of pages in span
-	int32 cl;
-	int32 state;	// state (enum above)
-//	int ref;	// reference count if state == SpanInUse (for GC)
-//	void *type;	// object type if state == SpanInUse (for GC)
-};
-
-// The Central cache contains a list of free spans,
-// as well as free lists of small blocks.
-struct Central
-{
-	Lock;
-	Span *free[256];
-	Span *large;	// free spans >= MaxPage pages
-};
-
-static Central central;
-static PageMap spanmap;
+Central central;
+PageMap spanmap;
 
 // Insert a new span into the map.
 static void
@@ -86,6 +58,39 @@ spanofptr(void *v)
 
 static void freespan(Span*);
 
+// Linked list of spans.
+// TODO(rsc): Remove - should be able to walk pagemap.
+Span *spanfirst;
+Span *spanlast;
+static void
+addtolist(Span *s)
+{
+	if(spanlast) {
+		s->aprev = spanlast;
+		s->aprev->anext = s;
+	} else {
+		s->aprev = nil;
+		spanfirst = s;
+	}
+	s->anext = nil;
+	spanlast = s;
+}
+
+/*
+static void
+delfromlist(Span *s)
+{
+	if(s->aprev)
+		s->aprev->anext = s->anext;
+	else
+		spanfirst = s->anext;
+	if(s->anext)
+		s->anext->aprev = s->aprev;
+	else
+		spanlast = s->aprev;
+}
+*/
+
 // Allocate a span of at least n pages.
 static Span*
 allocspan(int32 npage)
@@ -122,6 +127,7 @@ allocspan(int32 npage)
 //printf("New span %d for %d\n", allocnpage, npage);
 	s->base = trivalloc(allocnpage<<PageShift);
 	insertspan(s);
+	addtolist(s);
 
 havespan:
 	// If span is bigger than needed, redistribute the remainder.
@@ -131,6 +137,7 @@ havespan:
 		s1->length = s->length - npage;
 		shrinkspan(s, npage);
 		insertspan(s1);
+		addtolist(s1);
 		freespan(s1);
 	}
 	s->state = SpanInUse;
@@ -138,6 +145,7 @@ havespan:
 }
 
 // Free a span.
+// TODO(rsc): Coalesce adjacent free spans.
 static void
 freespan(Span *s)
 {
@@ -161,7 +169,7 @@ freespan(Span *s)
 
 // Small objects are kept on per-size free lists in the M.
 // There are SmallFreeClasses (defined in runtime.h) different lists.
-static int32 classtosize[SmallFreeClasses] = {
+int32 classtosize[SmallFreeClasses] = {
 	/*
 	seq 8 8 127 | sed 's/$/,/' | fmt
 	seq 128 16 255 | sed 's/$/,/' | fmt
@@ -257,16 +265,24 @@ centralgrab(int32 cl, int32 *pn)
 	chunk = (chunk+PageMask) & ~PageMask;
 	s = allocspan(chunk>>PageShift);
 //printf("New class %d\n", cl);
+
 	s->state = SpanInUse;
 	s->cl = cl;
 	siz = classtosize[cl];
-	n = chunk/siz;
+	n = chunk/(siz+sizeof(s->refbase[0]));
 	p = s->base;
 //printf("centralgrab cl=%d siz=%d n=%d\n", cl, siz, n);
-	for(i=0; i<n-1; i++) {
+	for(i=0; i<n; i++) {
+		if(i < n-1)
 			*(void**)p = p+siz;
 		p += siz;
 	}
+	s->refbase = (int32*)p;
+
+	// TODO(rsc): Remove - only for mark/sweep
+	for(i=0; i<n; i++)
+		s->refbase[i] = RefFree;
+
 	*pn = n;
 	return s->base;
 }
@@ -292,10 +308,21 @@ allocsmall(int32 cl)
 		unlock(&central);
 	}
 
-//printf("alloc from cl %d\n", cl);
+//printf("alloc from cl %d %p\n", cl, p);
 	// advance linked list.
 	m->freelist[cl] = *p;
 
+	// TODO(rsc): If cl > 0, can store ref ptr in *(p+1),
+	// avoiding call to findobj.
+	// Or could get rid of RefFree, which is only truly
+	// necessary for mark/sweep.
+	int32 *ref;
+	if(!findobj(p, nil, nil, &ref))
+		throw("bad findobj");
+	if(*ref != RefFree)
+		throw("double alloc");
+	*ref = 0;
+
 	// Blocks on free list are zeroed except for
 	// the linked list pointer that we just used.  Zero it.
 	*p = 0;
@@ -315,6 +342,7 @@ alloclarge(int32 np)
 	unlock(&central);
 	s->state = SpanInUse;
 	s->cl = -1;
+	s->ref = 0;
 	return s->base;
 }
 
@@ -347,13 +375,62 @@ allocator·malloc(int32 n, byte *out)
 	FLUSH(&out);
 }
 
+// Check whether v points into a known memory block.
+// If so, return true with
+//	*obj = base pointer of object (can pass to free)
+//	*size = size of object
+//	*ref = pointer to ref count for object
+// Object might already be freed, in which case *ref == RefFree.
+bool
+findobj(void *v, void **obj, int64 *size, int32 **ref)
+{
+	Span *s;
+	int32 siz, off, indx;
+
+	s = spanofptr(v);
+	if(s == nil || s->state != SpanInUse)
+		return false;
+
+	// Big object
+	if(s->cl < 0) {
+		if(obj)
+			*obj = s->base;
+		if(size)
+			*size = s->length<<PageShift;
+		if(ref)
+			*ref = &s->ref;
+		return true;
+	}
+
+	// Small object
+	if((byte*)v >= (byte*)s->refbase)
+		return false;
+	siz = classtosize[s->cl];
+	off = (byte*)v - (byte*)s->base;
+	indx = off/siz;
+	if(obj)
+		*obj = s->base + indx*siz;
+	if(size)
+		*size = siz;
+	if(ref)
+		*ref = s->refbase + indx;
+	return true;
+}
+
+void
+allocator·find(uint64 ptr, byte *obj, int64 siz, int32 *ref, bool ok)
+{
+	ok = findobj((void*)ptr, &obj, &siz, &ref);
+	FLUSH(&ok);
+}
+
 // Free object with base pointer v.
 void
 free(void *v)
 {
 	void **p;
 	Span *s;
-	int32 siz, off;
+	int32 siz, off, n;
 
 	s = spanofptr(v);
 	if(s->state != SpanInUse)
@@ -365,6 +442,9 @@ free(void *v)
 			throw("free - invalid pointer2");
 		// TODO: For large spans, maybe just return the
 		// memory to the operating system and let it zero it.
+		if(s->ref != 0 && s->ref != RefManual && s->ref != RefStack)
+			throw("free - bad ref count");
+		s->ref = RefFree;
 		sys·memclr(s->base, s->length << PageShift);
 //printf("Free big %D\n", s->length);
 		allocator·allocated -= s->length << PageShift;
@@ -375,10 +455,17 @@ free(void *v)
 	}
 
 	// Small object should be aligned properly.
+	if((byte*)v >= (byte*)s->refbase)
+		throw("free - invalid pointer4");
+
 	siz = classtosize[s->cl];
 	off = (byte*)v - (byte*)s->base;
 	if(off%siz)
 		throw("free - invalid pointer3");
+	n = off/siz;
+	if(s->refbase[n] != 0 && s->refbase[n] != RefManual && s->refbase[n] != RefStack)
+		throw("free - bad ref count1");
+	s->refbase[n] = RefFree;
 
 	// Zero and add to free list.
 	sys·memclr(v, siz);

usr/rsc/mem/malloc.h

@@ -12,6 +12,10 @@ enum
 	PageMask = (1<<PageShift) - 1,
 };
 
+#define RefFree	0xffffffffU
+#define RefManual	0xfffffffeU
+#define RefStack	0xfffffffdU
+
 enum {
 	PMBits = 64 - PageShift,
 	PMLevels = 4,
@@ -24,6 +28,40 @@ struct PageMap
 	void *level0[PMLevelSize];
 };
 
+typedef struct Span Span;
+typedef struct Central Central;
+
+// A Span contains metadata about a range of pages.
+enum {
+	SpanInUse = 0,	// span has been handed out by allocator
+	SpanFree = 1,	// span is in central free list
+};
+struct Span
+{
+	Span *aprev;	// in list of all spans
+	Span *anext;
+
+	Span *next;	// in free lists
+	byte *base;	// first byte in span
+	uintptr length;	// number of pages in span
+	int32 cl;
+	int32 state;	// state (enum above)
+	union {
+		int32 ref;	// reference count if state == SpanInUse (for GC)
+		int32 *refbase;	// ptr to packed ref counts
+	};
+//	void *type;	// object type if state == SpanInUse (for GC)
+};
+
+// The Central cache contains a list of free spans,
+// as well as free lists of small blocks.
+struct Central
+{
+	Lock;
+	Span *free[256];
+	Span *large;	// free spans >= MaxPage pages
+};
+
 extern int64 allocator·allocated;
 extern int64 allocator·footprint;
 extern bool allocator·frozen;
@@ -34,3 +72,9 @@ void* pminsert(PageMap*, uintptr, void*);
 
 void*	alloc(int32);
 void	free(void*);
+bool	findobj(void*, void**, int64*, int32**);
+
+extern Central central;
+extern PageMap spanmap;
+extern int32 classtosize[SmallFreeClasses];
+extern Span *spanfirst, *spanlast;

usr/rsc/mem/stack.c → usr/rsc/mem/mem.c

@@ -8,9 +8,13 @@ void*
 stackalloc(uint32 n)
 {
 	void *v;
+	int32 *ref;
 
 	v = alloc(n);
 //printf("stackalloc %d = %p\n", n, v);
+	ref = nil;
+	findobj(v, nil, nil, &ref);
+	*ref = RefStack;
 	return v;
 }
 
@@ -20,3 +24,16 @@ stackfree(void *v)
 //printf("stackfree %p\n", v);
 	free(v);
 }
+
+void*
+mal(uint32 n)
+{
+	return alloc(n);
+}
+
+void
+sys·mal(uint32 n, uint8 *ret)
+{
+	ret = alloc(n);
+	FLUSH(&ret);
+}

usr/rsc/mem/testrandom.go

@@ -6,7 +6,8 @@ package main
 
 import (
 	"allocator";
-	"rand"
+	"rand";
+	"syscall"
 )
 
 var footprint int64;
@@ -52,6 +53,11 @@ func main() {
 		}
 		siz := rand.rand() >> (11 + rand.urand32() % 20);
 		base := allocator.malloc(siz);
+		ptr := uint64(syscall.BytePtr(base))+uint64(siz/2);
+		obj, size, ref, ok := allocator.find(ptr);
+		if obj != base || *ref != 0 || !ok {
+			panicln("find", siz, obj, ref, ok);
+		}
 		blocks[b].base = base;
 		blocks[b].siz = siz;
 		allocated += int64(siz);