runtime: parallelize garbage collector mark + sweep

Running test/garbage/parser.out.

On a 4-core Lenovo X201s (Linux):
31.12u 0.60s 31.74r 	 1 cpu, no atomics
32.27u 0.58s 32.86r 	 1 cpu, atomic instructions
33.04u 0.83s 27.47r 	 2 cpu

On a 16-core Xeon (Linux):
33.08u 0.65s 33.80r 	 1 cpu, no atomics
34.87u 1.12s 29.60r 	 2 cpu
36.00u 1.87s 28.43r 	 3 cpu
36.46u 2.34s 27.10r 	 4 cpu
38.28u 3.85s 26.92r 	 5 cpu
37.72u 5.25s 26.73r	 6 cpu
39.63u 7.11s 26.95r	 7 cpu
39.67u 8.10s 26.68r	 8 cpu

On a 2-core MacBook Pro Core 2 Duo 2.26 (circa 2009, MacBookPro5,5):
39.43u 1.45s 41.27r 	 1 cpu, no atomics
43.98u 2.95s 38.69r 	 2 cpu

On a 2-core Mac Mini Core 2 Duo 1.83 (circa 2008; Macmini2,1):
48.81u 2.12s 51.76r 	 1 cpu, no atomics
57.15u 4.72s 51.54r 	 2 cpu

The handoff algorithm is really only good for two cores.
Beyond that we will need to so something more sophisticated,
like have each core hand off to the next one, around a circle.
Even so, the code is a good checkpoint; for now we'll limit the
number of gc procs to at most 2.

R=dvyukov
CC=golang-dev
https://golang.org/cl/4641082

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

@@ -138,6 +138,26 @@ TEXT runtime·sigaltstack(SB),7,$0
 	CALL	runtime·notok(SB)
 	RET
 
+TEXT runtime·usleep(SB),7,$32
+	MOVL	$0, DX
+	MOVL	usec+0(FP), AX
+	MOVL	$1000000, CX
+	DIVL	CX
+	MOVL	AX, 24(SP)  // sec
+	MOVL	DX, 28(SP)  // usec
+
+	// select(0, 0, 0, 0, &tv)
+	MOVL	$0, 0(SP)  // "return PC" - ignored
+	MOVL	$0, 4(SP)
+	MOVL	$0, 8(SP)
+	MOVL	$0, 12(SP)
+	MOVL	$0, 16(SP)
+	LEAL	24(SP), AX
+	MOVL	AX, 20(SP)
+	MOVL	$93, AX
+	INT	$0x80
+	RET
+
 // void bsdthread_create(void *stk, M *m, G *g, void (*fn)(void))
 // System call args are: func arg stack pthread flags.
 TEXT runtime·bsdthread_create(SB),7,$32
@@ -309,3 +329,12 @@ TEXT runtime·setldt(SB),7,$32
 	XORL	AX, AX
 	MOVW	GS, AX
 	RET
+
+TEXT runtime·sysctl(SB),7,$0
+	MOVL	$202, AX
+	INT	$0x80
+	JAE	3(PC)
+	NEGL	AX
+	RET
+	MOVL	$0, AX
+	RET

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

@@ -146,6 +146,24 @@ TEXT runtime·sigaltstack(SB),7,$0
 	CALL	runtime·notok(SB)
 	RET
 
+TEXT runtime·usleep(SB),7,$16
+	MOVL	$0, DX
+	MOVL	usec+0(FP), AX
+	MOVL	$1000000, CX
+	DIVL	CX
+	MOVQ	AX, 0(SP)  // sec
+	MOVL	DX, 8(SP)  // usec
+
+	// select(0, 0, 0, 0, &tv)
+	MOVL	$0, DI
+	MOVL	$0, SI
+	MOVL	$0, DX
+	MOVL	$0, R10
+	MOVQ	SP, R8
+	MOVL	$(0x2000000+23), AX
+	SYSCALL
+	RET
+
 // void bsdthread_create(void *stk, M *m, G *g, void (*fn)(void))
 TEXT runtime·bsdthread_create(SB),7,$0
 	// Set up arguments to bsdthread_create system call.
@@ -293,3 +311,18 @@ TEXT runtime·settls(SB),7,$32
 	MOVL	$(0x3000000+3), AX	// thread_fast_set_cthread_self - machdep call #3
 	SYSCALL
 	RET
+
+TEXT runtime·sysctl(SB),7,$0
+	MOVQ	8(SP), DI
+	MOVL	16(SP), SI
+	MOVQ	24(SP), DX
+	MOVQ	32(SP), R10
+	MOVQ	40(SP), R8
+	MOVQ	48(SP), R9
+	MOVL	$(0x2000000+202), AX	// syscall entry
+	SYSCALL
+	JCC 3(PC)
+	NEGL	AX
+	RET
+	MOVL	$0, AX
+	RET

src/pkg/runtime/darwin/os.h

@@ -18,6 +18,7 @@ uint32	runtime·mach_task_self(void);
 uint32	runtime·mach_task_self(void);
 uint32	runtime·mach_thread_self(void);
 uint32	runtime·mach_thread_self(void);
+int32	runtime·sysctl(uint32*, uint32, byte*, uintptr*, byte*, uintptr);
 
 struct Sigaction;
 void	runtime·sigaction(uintptr, struct Sigaction*, struct Sigaction*);

src/pkg/runtime/darwin/thread.c

@@ -148,6 +148,20 @@ runtime·osinit(void)
 	if(!runtime·iscgo)
 		runtime·bsdthread_register();
 	runtime·destroylock = destroylock;
+
+	// Use sysctl to fetch hw.ncpu.
+	uint32 mib[2];
+	uint32 out;
+	int32 ret;
+	uintptr nout;
+
+	mib[0] = 6;
+	mib[1] = 3;
+	nout = sizeof out;
+	out = 0;
+	ret = runtime·sysctl(mib, 2, (byte*)&out, &nout, nil, 0);
+	if(ret >= 0)
+		runtime·ncpu = out;
 }
 
 void

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

@@ -52,6 +52,25 @@ TEXT runtime·read(SB),7,$0
 	CALL	*runtime·_vdso(SB)
 	RET
 
+TEXT runtime·usleep(SB),7,$28
+	MOVL	$0, DX
+	MOVL	usec+0(FP), AX
+	MOVL	$1000000, CX
+	DIVL	CX
+	MOVL	AX, 20(SP)
+	MOVL	DX, 24(SP)
+
+	// select(0, 0, 0, 0, &tv)
+	MOVL	$0, 0(SP)
+	MOVL	$0, 4(SP)
+	MOVL	$0, 8(SP)
+	MOVL	$0, 12(SP)
+	LEAL	20(SP), AX
+	MOVL	AX, 16(SP)
+	MOVL	$82, AX
+	SYSCALL
+	RET
+
 TEXT runtime·raisesigpipe(SB),7,$12
 	MOVL	$224, AX	// syscall - gettid
 	CALL	*runtime·_vdso(SB)

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

@@ -50,6 +50,24 @@ TEXT runtime·read(SB),7,$0-24
 	SYSCALL
 	RET
 
+TEXT runtime·usleep(SB),7,$16
+	MOVL	$0, DX
+	MOVL	usec+0(FP), AX
+	MOVL	$1000000, CX
+	DIVL	CX
+	MOVQ	AX, 0(SP)
+	MOVQ	DX, 8(SP)
+
+	// select(0, 0, 0, 0, &tv)
+	MOVL	$0, DI
+	MOVL	$0, SI
+	MOVL	$0, DX
+	MOVL	$0, R10
+	MOVQ	SP, R8
+	MOVL	$23, AX
+	SYSCALL
+	RET
+
 TEXT runtime·raisesigpipe(SB),7,$12
 	MOVL	$186, AX	// syscall - gettid
 	SYSCALL

src/pkg/runtime/linux/arm/sys.s

@@ -33,6 +33,7 @@
 #define SYS_gettid (SYS_BASE + 224)
 #define SYS_tkill (SYS_BASE + 238)
 #define SYS_sched_yield (SYS_BASE + 158)
+#define SYS_select (SYS_BASE + 82)
 
 #define ARM_BASE (SYS_BASE + 0x0f0000)
 #define SYS_ARM_cacheflush (ARM_BASE + 2)
@@ -285,6 +286,23 @@ TEXT runtime·sigreturn(SB),7,$0
 	SWI	$0
 	RET
 
+TEXT runtime·usleep(SB),7,$12
+	MOVW	usec+0(FP), R0
+	MOVW	R0, R1
+	MOVW	$1000000, R2
+	DIV	R1, R0
+	MOD	R2, R0
+	MOVW	R1, 4(SP)
+	MOVW	R2, 8(SP)
+	MOVW	$0, R0
+	MOVW	$0, R1
+	MOVW	$0, R2
+	MOVW	$0, R3
+	MOVW	$4(SP), R4
+	MOVW	$SYS_select, R7
+	SWI	$0
+	RET
+
 // Use kernel version instead of native armcas in ../../arm.s.
 // See ../../../sync/atomic/asm_linux_arm.s for details.
 TEXT cas<>(SB),7,$0

src/pkg/runtime/linux/thread.c

@@ -8,7 +8,6 @@
 #include "stack.h"
 
 extern SigTab runtime·sigtab[];
-static int32 proccount;
 
 int32 runtime·open(uint8*, int32, int32);
 int32 runtime·close(int32);
@@ -136,13 +135,10 @@ futexlock(Lock *l)
 	// its wakeup call.
 	wait = v;
 
-	if(proccount == 0)
-		proccount = getproccount();
-
 	// On uniprocessor's, no point spinning.
 	// On multiprocessors, spin for ACTIVE_SPIN attempts.
 	spin = 0;
-	if(proccount > 1)
+	if(runtime·ncpu > 1)
 		spin = ACTIVE_SPIN;
 
 	for(;;) {
@@ -276,6 +272,7 @@ runtime·newosproc(M *m, G *g, void *stk, void (*fn)(void))
 void
 runtime·osinit(void)
 {
+	runtime·ncpu = getproccount();
 }
 
 void

src/pkg/runtime/malloc.h

@@ -120,6 +120,13 @@ enum
 #else
 	MHeapMap_Bits = 20,
 #endif
+
+	// Max number of threads to run garbage collection.
+	// 2, 3, and 4 are all plausible maximums depending
+	// on the hardware details of the machine.  The second
+	// proc is the one that helps the most (after the first),
+	// so start with just 2 for now.
+	MaxGcproc = 2,
 };
 
 // A generic linked list of blocks.  (Typically the block is bigger than sizeof(MLink).)
@@ -400,6 +407,8 @@ enum
 
 void	runtime·MProf_Malloc(void*, uintptr);
 void	runtime·MProf_Free(void*, uintptr);
+int32	runtime·helpgc(void);
+void	runtime·gchelper(void);
 
 // Malloc profiling settings.
 // Must match definition in extern.go.

src/pkg/runtime/mgc0.c

@@ -10,8 +10,8 @@
 
 enum {
 	Debug = 0,
-	UseCas = 1,
 	PtrSize = sizeof(void*),
+	DebugMark = 0,  // run second pass to check mark
 
 	// Four bits per word (see #defines below).
 	wordsPerBitmapWord = sizeof(void*)*8/4,
@@ -51,17 +51,20 @@ enum {
 
 #define bitMask (bitBlockBoundary | bitAllocated | bitMarked | bitSpecial)
 
+// TODO: Make these per-M.
 static uint64 nlookup;
 static uint64 nsizelookup;
 static uint64 naddrlookup;
+static uint64 nhandoff;
+
 static int32 gctrace;
 
 typedef struct Workbuf Workbuf;
 struct Workbuf
 {
 	Workbuf *next;
-	uintptr nw;
-	byte *w[2048-2];
+	uintptr nobj;
+	byte *obj[512-2];
 };
 
 extern byte data[];
@@ -75,6 +78,26 @@ static int32 fingwait;
 static void runfinq(void);
 static Workbuf* getempty(Workbuf*);
 static Workbuf* getfull(Workbuf*);
+static void	putempty(Workbuf*);
+static Workbuf* handoff(Workbuf*);
+
+static struct {
+	Lock fmu;
+	Workbuf	*full;
+	Lock emu;
+	Workbuf	*empty;
+	uint32	nproc;
+	volatile uint32	nwait;
+	volatile uint32	ndone;
+	Note	alldone;
+	Lock	markgate;
+	Lock	sweepgate;
+	MSpan	*spans;
+
+	Lock;
+	byte	*chunk;
+	uintptr	nchunk;
+} work;
 
 // scanblock scans a block of n bytes starting at pointer b for references
 // to other objects, scanning any it finds recursively until there are no
@@ -85,13 +108,14 @@ static Workbuf* getfull(Workbuf*);
 static void
 scanblock(byte *b, int64 n)
 {
-	byte *obj, *arena_start, *p;
+	byte *obj, *arena_start, *arena_used, *p;
 	void **vp;
-	uintptr size, *bitp, bits, shift, i, j, x, xbits, off;
+	uintptr size, *bitp, bits, shift, i, j, x, xbits, off, nobj, nproc;
 	MSpan *s;
 	PageID k;
-	void **bw, **w, **ew;
+	void **wp;
 	Workbuf *wbuf;
+	bool keepworking;
 
 	if((int64)(uintptr)n != n || n < 0) {
 		runtime·printf("scanblock %p %D\n", b, n);
@@ -100,11 +124,19 @@ scanblock(byte *b, int64 n)
 
 	// Memory arena parameters.
 	arena_start = runtime·mheap.arena_start;
+	arena_used = runtime·mheap.arena_used;
+	nproc = work.nproc;
 
 	wbuf = nil;  // current work buffer
-	ew = nil;  // end of work buffer
-	bw = nil;  // beginning of work buffer
-	w = nil;  // current pointer into work buffer
+	wp = nil;  // storage for next queued pointer (write pointer)
+	nobj = 0;  // number of queued objects
+
+	// Scanblock helpers pass b==nil.
+	// The main proc needs to return to make more
+	// calls to scanblock.  But if work.nproc==1 then
+	// might as well process blocks as soon as we
+	// have them.
+	keepworking = b == nil || work.nproc == 1;
 
 	// Align b to a word boundary.
 	off = (uintptr)b & (PtrSize-1);
@@ -120,12 +152,12 @@ scanblock(byte *b, int64 n)
 			runtime·printf("scanblock %p %D\n", b, n);
 
 		vp = (void**)b;
-		n /= PtrSize;
+		n >>= (2+PtrSize/8);  /* n /= PtrSize (4 or 8) */
 		for(i=0; i<n; i++) {
 			obj = (byte*)vp[i];
 
 			// Words outside the arena cannot be pointers.
-			if((byte*)obj < arena_start || (byte*)obj >= runtime·mheap.arena_used)
+			if((byte*)obj < arena_start || (byte*)obj >= arena_used)
 				continue;
 
 			// obj may be a pointer to a live object.
@@ -188,39 +220,64 @@ scanblock(byte *b, int64 n)
 		found:
 			// Now we have bits, bitp, and shift correct for
 			// obj pointing at the base of the object.
-			// If not allocated or already marked, done.
-			if((bits & bitAllocated) == 0 || (bits & bitMarked) != 0)
+			// Only care about allocated and not marked.
+			if((bits & (bitAllocated|bitMarked)) != bitAllocated)
 				continue;
+			if(nproc == 1)
 				*bitp |= bitMarked<<shift;
+			else {
+				for(;;) {
+					x = *bitp;
+					if(x & (bitMarked<<shift))
+						goto continue_obj;
+					if(runtime·casp((void**)bitp, (void*)x, (void*)(x|(bitMarked<<shift))))
+						break;
+				}
+			}
 
 			// If object has no pointers, don't need to scan further.
 			if((bits & bitNoPointers) != 0)
 				continue;
 
+			// If another proc wants a pointer, give it some.
+			if(nobj > 4 && work.nwait > 0 && work.full == nil) {
+				wbuf->nobj = nobj;
+				wbuf = handoff(wbuf);
+				nobj = wbuf->nobj;
+				wp = wbuf->obj + nobj;
+			}
+
 			// If buffer is full, get a new one.
-			if(w >= ew) {
+			if(wbuf == nil || nobj >= nelem(wbuf->obj)) {
+				if(wbuf != nil)
+					wbuf->nobj = nobj;
 				wbuf = getempty(wbuf);
-				bw = wbuf->w;
-				w = bw;
-				ew = bw + nelem(wbuf->w);
+				wp = wbuf->obj;
+				nobj = 0;
 			}
-			*w++ = obj;
+			*wp++ = obj;
+			nobj++;
+		continue_obj:;
 		}
 
 		// Done scanning [b, b+n).  Prepare for the next iteration of
 		// the loop by setting b and n to the parameters for the next block.
 
-		// Fetch b from the work buffers.
-		if(w <= bw) {
+		// Fetch b from the work buffer.
+		if(nobj == 0) {
+			if(!keepworking) {
+				putempty(wbuf);
+				return;
+			}
 			// Emptied our buffer: refill.
 			wbuf = getfull(wbuf);
 			if(wbuf == nil)
-				break;
-			bw = wbuf->w;
-			ew = wbuf->w + nelem(wbuf->w);
-			w = bw+wbuf->nw;
+				return;
+			nobj = wbuf->nobj;
+			wp = wbuf->obj + wbuf->nobj;
 		}
-		b = *--w;
+		b = *--wp;
+		nobj--;
 
 		// Figure out n = size of b.  Start by loading bits for b.
 		off = (uintptr*)b - (uintptr*)arena_start;
@@ -277,29 +334,123 @@ scanblock(byte *b, int64 n)
 	}
 }
 
-static struct {
-	Workbuf	*full;
-	Workbuf	*empty;
-	byte	*chunk;
-	uintptr	nchunk;
-} work;
+// debug_scanblock is the debug copy of scanblock.
+// it is simpler, slower, single-threaded, recursive,
+// and uses bitSpecial as the mark bit.
+static void
+debug_scanblock(byte *b, int64 n)
+{
+	byte *obj, *p;
+	void **vp;
+	uintptr size, *bitp, bits, shift, i, xbits, off;
+	MSpan *s;
+
+	if(!DebugMark)
+		runtime·throw("debug_scanblock without DebugMark");
+
+	if((int64)(uintptr)n != n || n < 0) {
+		runtime·printf("debug_scanblock %p %D\n", b, n);
+		runtime·throw("debug_scanblock");
+	}
+
+	// Align b to a word boundary.
+	off = (uintptr)b & (PtrSize-1);
+	if(off != 0) {
+		b += PtrSize - off;
+		n -= PtrSize - off;
+	}
+
+	vp = (void**)b;
+	n /= PtrSize;
+	for(i=0; i<n; i++) {
+		obj = (byte*)vp[i];
+
+		// Words outside the arena cannot be pointers.
+		if((byte*)obj < runtime·mheap.arena_start || (byte*)obj >= runtime·mheap.arena_used)
+			continue;
+
+		// Round down to word boundary.
+		obj = (void*)((uintptr)obj & ~((uintptr)PtrSize-1));
+
+		// Consult span table to find beginning.
+		s = runtime·MHeap_LookupMaybe(&runtime·mheap, obj);
+		if(s == nil)
+			continue;
+
+
+		p =  (byte*)((uintptr)s->start<<PageShift);
+		if(s->sizeclass == 0) {
+			obj = p;
+			size = (uintptr)s->npages<<PageShift;
+		} else {
+			if((byte*)obj >= (byte*)s->limit)
+				continue;
+			size = runtime·class_to_size[s->sizeclass];
+			int32 i = ((byte*)obj - p)/size;
+			obj = p+i*size;
+		}
+
+		// Now that we know the object header, reload bits.
+		off = (uintptr*)obj - (uintptr*)runtime·mheap.arena_start;
+		bitp = (uintptr*)runtime·mheap.arena_start - off/wordsPerBitmapWord - 1;
+		shift = off % wordsPerBitmapWord;
+		xbits = *bitp;
+		bits = xbits >> shift;
+
+		// Now we have bits, bitp, and shift correct for
+		// obj pointing at the base of the object.
+		// If not allocated or already marked, done.
+		if((bits & bitAllocated) == 0 || (bits & bitSpecial) != 0)  // NOTE: bitSpecial not bitMarked
+			continue;
+		*bitp |= bitSpecial<<shift;
+		if(!(bits & bitMarked))
+			runtime·printf("found unmarked block %p in %p\n", obj, vp+i);
+
+		// If object has no pointers, don't need to scan further.
+		if((bits & bitNoPointers) != 0)
+			continue;
+
+		debug_scanblock(obj, size);
+	}
+}
 
 // Get an empty work buffer off the work.empty list,
 // allocating new buffers as needed.
 static Workbuf*
 getempty(Workbuf *b)
 {
+	if(work.nproc == 1) {
+		// Put b on full list.
 		if(b != nil) {
-		b->nw = nelem(b->w);
 			b->next = work.full;
 			work.full = b;
 		}
+		// Grab from empty list if possible.
 		b = work.empty;
 		if(b != nil) {
 			work.empty = b->next;
-		return b;
+			goto haveb;
+		}
+	} else {
+		// Put b on full list.
+		if(b != nil) {
+			runtime·lock(&work.fmu);
+			b->next = work.full;
+			work.full = b;
+			runtime·unlock(&work.fmu);
+		}
+		// Grab from empty list if possible.
+		runtime·lock(&work.emu);
+		b = work.empty;
+		if(b != nil)
+			work.empty = b->next;
+		runtime·unlock(&work.emu);
+		if(b != nil)
+			goto haveb;
 	}
 
+	// Need to allocate.
+	runtime·lock(&work);
 	if(work.nchunk < sizeof *b) {
 		work.nchunk = 1<<20;
 		work.chunk = runtime·SysAlloc(work.nchunk);
@@ -307,27 +458,122 @@ getempty(Workbuf *b)
 	b = (Workbuf*)work.chunk;
 	work.chunk += sizeof *b;
 	work.nchunk -= sizeof *b;
+	runtime·unlock(&work);
+
+haveb:
+	b->nobj = 0;
 	return b;
 }
 
+static void
+putempty(Workbuf *b)
+{
+	if(b == nil)
+		return;
+
+	if(work.nproc == 1) {
+		b->next = work.empty;
+		work.empty = b;
+		return;
+	}
+
+	runtime·lock(&work.emu);
+	b->next = work.empty;
+	work.empty = b->next;
+	runtime·unlock(&work.emu);
+}
+
 // Get a full work buffer off the work.full list, or return nil.
 static Workbuf*
 getfull(Workbuf *b)
 {
+	int32 i;
+	Workbuf *b1;
+
+	if(work.nproc == 1) {
+		// Put b on empty list.
 		if(b != nil) {
-		b->nw = 0;
 			b->next = work.empty;
 			work.empty = b;
 		}
+		// Grab from full list if possible.
+		// Since work.nproc==1, no one else is
+		// going to give us work.
 		b = work.full;
 		if(b != nil)
 			work.full = b->next;
 		return b;
+	}
+
+	putempty(b);
+
+	// Grab buffer from full list if possible.
+	for(;;) {
+		b1 = work.full;
+		if(b1 == nil)
+			break;
+		runtime·lock(&work.fmu);
+		if(work.full != nil) {
+			b1 = work.full;
+			work.full = b1->next;
+			runtime·unlock(&work.fmu);
+			return b1;
+		}
+		runtime·unlock(&work.fmu);
+	}
+
+	runtime·xadd(&work.nwait, +1);
+	for(i=0;; i++) {
+		b1 = work.full;
+		if(b1 != nil) {
+			runtime·lock(&work.fmu);
+			if(work.full != nil) {
+				runtime·xadd(&work.nwait, -1);
+				b1 = work.full;
+				work.full = b1->next;
+				runtime·unlock(&work.fmu);
+				return b1;
+			}
+			runtime·unlock(&work.fmu);
+			continue;
+		}
+		if(work.nwait == work.nproc)
+			return nil;
+		if(i < 10)
+			runtime·procyield(20);
+		else if(i < 20)
+			runtime·osyield();
+		else
+			runtime·usleep(100);
+	}
+}
+
+static Workbuf*
+handoff(Workbuf *b)
+{
+	int32 n;
+	Workbuf *b1;
+
+	// Make new buffer with half of b's pointers.
+	b1 = getempty(nil);
+	n = b->nobj/2;
+	b->nobj -= n;
+	b1->nobj = n;
+	runtime·memmove(b1->obj, b->obj+b->nobj, n*sizeof b1->obj[0]);
+	nhandoff += n;
+
+	// Put b on full list - let first half of b get stolen.
+	runtime·lock(&work.fmu);
+	b->next = work.full;
+	work.full = b;
+	runtime·unlock(&work.fmu);
+
+	return b1;
 }
 
 // Scanstack calls scanblock on each of gp's stack segments.
 static void
-scanstack(G *gp)
+scanstack(void (*scanblock)(byte*, int64), G *gp)
 {
 	int32 n;
 	Stktop *stk;
@@ -339,6 +585,9 @@ scanstack(G *gp)
 	if(gp == g) {
 		// Scanning our own stack: start at &gp.
 		sp = (byte*)&gp;
+	} else if(gp->m != nil && gp->m->helpgc) {
+		// Gc helper scans its own stack.
+		return;
 	} else {
 		// Scanning another goroutine's stack.
 		// The goroutine is usually asleep (the world is stopped).
@@ -387,17 +636,27 @@ markfin(void *v)
 	scanblock(v, size);
 }
 
+static void
+debug_markfin(void *v)
+{
+	uintptr size;
+
+	if(!runtime·mlookup(v, &v, &size, nil))
+		runtime·throw("debug_mark - finalizer inconsistency");
+	debug_scanblock(v, size);
+}
+
 // Mark
 static void
-mark(void)
+mark(void (*scan)(byte*, int64))
 {
 	G *gp;
 
 	// mark data+bss.
 	// skip runtime·mheap itself, which has no interesting pointers
 	// and is mostly zeroed and would not otherwise be paged in.
-	scanblock(data, (byte*)&runtime·mheap - data);
-	scanblock((byte*)(&runtime·mheap+1), end - (byte*)(&runtime·mheap+1));
+	scan(data, (byte*)&runtime·mheap - data);
+	scan((byte*)(&runtime·mheap+1), end - (byte*)(&runtime·mheap+1));
 
 	// mark stacks
 	for(gp=runtime·allg; gp!=nil; gp=gp->alllink) {
@@ -410,18 +669,24 @@ mark(void)
 		case Grunning:
 			if(gp != g)
 				runtime·throw("mark - world not stopped");
-			scanstack(gp);
+			scanstack(scan, gp);
 			break;
 		case Grunnable:
 		case Gsyscall:
 		case Gwaiting:
-			scanstack(gp);
+			scanstack(scan, gp);
 			break;
 		}
 	}
 
 	// mark things pointed at by objects with finalizers
+	if(scan == debug_scanblock)
+		runtime·walkfintab(debug_markfin);
+	else
 		runtime·walkfintab(markfin);
+
+	// in multiproc mode, join in the queued work.
+	scan(nil, 0);
 }
 
 // Sweep frees or calls finalizers for blocks not marked in the mark phase.
@@ -435,8 +700,17 @@ sweep(void)
 	byte *p;
 	MCache *c;
 	Finalizer *f;
+	byte *arena_start;
+
+	arena_start = runtime·mheap.arena_start;
+
+	for(;;) {
+		s = work.spans;
+		if(s == nil)
+			break;
+		if(!runtime·casp(&work.spans, s, s->allnext))
+			continue;
 
-	for(s = runtime·mheap.allspans; s != nil; s = s->allnext) {
 		if(s->state != MSpanInUse)
 			continue;
 
@@ -452,12 +726,14 @@ sweep(void)
 			n = (npages << PageShift) / size;
 		}
 
-		// sweep through n objects of given size starting at p.
+		// Sweep through n objects of given size starting at p.
+		// This thread owns the span now, so it can manipulate
+		// the block bitmap without atomic operations.
 		for(; n > 0; n--, p += size) {
 			uintptr off, *bitp, shift, bits;
 
-			off = (uintptr*)p - (uintptr*)runtime·mheap.arena_start;
-			bitp = (uintptr*)runtime·mheap.arena_start - off/wordsPerBitmapWord - 1;
+			off = (uintptr*)p - (uintptr*)arena_start;
+			bitp = (uintptr*)arena_start - off/wordsPerBitmapWord - 1;
 			shift = off % wordsPerBitmapWord;
 			bits = *bitp>>shift;
 
@@ -465,17 +741,27 @@ sweep(void)
 				continue;
 
 			if((bits & bitMarked) != 0) {
+				if(DebugMark) {
+					if(!(bits & bitSpecial))
+						runtime·printf("found spurious mark on %p\n", p);
+					*bitp &= ~(bitSpecial<<shift);
+				}
 				*bitp &= ~(bitMarked<<shift);
 				continue;
 			}
 
-			if((bits & bitSpecial) != 0) {
+			if(DebugMark || (bits & bitSpecial) != 0) {
 				// Special means it has a finalizer or is being profiled.
+				// In DebugMark mode, the bit has been coopted so
+				// we have to assume all blocks are special.
 				f = runtime·getfinalizer(p, 1);
 				if(f != nil) {
 					f->arg = p;
+					for(;;) {
 						f->next = finq;
-					finq = f;
+						if(runtime·casp(&finq, f->next, f))
+							break;
+					}
 					continue;
 				}
 				runtime·MProf_Free(p, size);
@@ -503,6 +789,23 @@ sweep(void)
 	}
 }
 
+void
+runtime·gchelper(void)
+{
+	// Wait until main proc is ready for mark help.
+	runtime·lock(&work.markgate);
+	runtime·unlock(&work.markgate);
+	scanblock(nil, 0);
+
+	// Wait until main proc is ready for sweep help.
+	runtime·lock(&work.sweepgate);
+	runtime·unlock(&work.sweepgate);
+	sweep();
+
+	if(runtime·xadd(&work.ndone, +1) == work.nproc-1)
+		runtime·notewakeup(&work.alldone);
+}
+
 // Semaphore, not Lock, so that the goroutine
 // reschedules when there is contention rather
 // than spinning.
@@ -562,6 +865,7 @@ runtime·gc(int32 force)
 	uint64 heap0, heap1, obj0, obj1;
 	byte *p;
 	Finalizer *fp;
+	bool extra;
 
 	// The gc is turned off (via enablegc) until
 	// the bootstrap has completed.
@@ -600,6 +904,7 @@ runtime·gc(int32 force)
 	nlookup = 0;
 	nsizelookup = 0;
 	naddrlookup = 0;
+	nhandoff = 0;
 
 	m->gcing = 1;
 	runtime·stoptheworld();
@@ -608,10 +913,30 @@ runtime·gc(int32 force)
 	heap0 = mstats.heap_alloc;
 	obj0 = mstats.nmalloc - mstats.nfree;
 
-	mark();
+	runtime·lock(&work.markgate);
+	runtime·lock(&work.sweepgate);
+
+	work.nproc = 1;
+	if(runtime·gomaxprocs > 1 && runtime·ncpu > 1) {
+		runtime·noteclear(&work.alldone);
+		work.nproc += runtime·helpgc();
+	}
+	work.nwait = 0;
+	work.ndone = 0;
+
+	runtime·unlock(&work.markgate);  // let the helpers in
+	mark(scanblock);
+	if(DebugMark)
+		mark(debug_scanblock);
 	t1 = runtime·nanotime();
+
+	work.spans = runtime·mheap.allspans;
+	runtime·unlock(&work.sweepgate);  // let the helpers in
 	sweep();
+	if(work.nproc > 1)
+		runtime·notesleep(&work.alldone);
 	t2 = runtime·nanotime();
+
 	stealcache();
 	cachestats();
 
@@ -643,15 +968,25 @@ runtime·gc(int32 force)
 		runtime·printf("pause %D\n", t3-t0);
 
 	if(gctrace) {
-		runtime·printf("gc%d: %D+%D+%D ms %D -> %D MB %D -> %D (%D-%D) objects %D pointer lookups (%D size, %D addr)\n",
-			mstats.numgc, (t1-t0)/1000000, (t2-t1)/1000000, (t3-t2)/1000000,
+		runtime·printf("gc%d(%d): %D+%D+%D ms %D -> %D MB %D -> %D (%D-%D) objects %D pointer lookups (%D size, %D addr) %D handoff\n",
+			mstats.numgc, work.nproc, (t1-t0)/1000000, (t2-t1)/1000000, (t3-t2)/1000000,
 			heap0>>20, heap1>>20, obj0, obj1,
 			mstats.nmalloc, mstats.nfree,
-			nlookup, nsizelookup, naddrlookup);
+			nlookup, nsizelookup, naddrlookup, nhandoff);
 	}
 
 	runtime·semrelease(&gcsema);
-	runtime·starttheworld();
+
+	// If we could have used another helper proc, start one now,
+	// in the hope that it will be available next time.
+	// It would have been even better to start it before the collection,
+	// but doing so requires allocating memory, so it's tricky to
+	// 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.
+	extra = work.nproc < runtime·gomaxprocs && work.nproc < MaxGcproc;
+
+	runtime·starttheworld(extra);
 
 	// give the queued finalizers, if any, a chance to run
 	if(fp != nil)
@@ -674,7 +1009,7 @@ runtime·UpdateMemStats(void)
 	cachestats();
 	m->gcing = 0;
 	runtime·semrelease(&gcsema);
-	runtime·starttheworld();
+	runtime·starttheworld(0);
 }
 
 static void
@@ -858,6 +1193,9 @@ runtime·blockspecial(void *v)
 {
 	uintptr *b, off, shift;
 
+	if(DebugMark)
+		return true;
+
 	off = (uintptr*)v - (uintptr*)runtime·mheap.arena_start;
 	b = (uintptr*)runtime·mheap.arena_start - off/wordsPerBitmapWord - 1;
 	shift = off % wordsPerBitmapWord;
@@ -870,6 +1208,9 @@ runtime·setblockspecial(void *v)
 {
 	uintptr *b, off, shift, bits, obits;
 
+	if(DebugMark)
+		return;
+
 	off = (uintptr*)v - (uintptr*)runtime·mheap.arena_start;
 	b = (uintptr*)runtime·mheap.arena_start - off/wordsPerBitmapWord - 1;
 	shift = off % wordsPerBitmapWord;

src/pkg/runtime/print.c

@@ -51,7 +51,7 @@ vprintf(int8 *s, byte *base)
 	uintptr arg, narg;
 	byte *v;
 
-//	lock(&debuglock);
+	//runtime·lock(&debuglock);
 
 	lp = p = s;
 	arg = 0;
@@ -152,7 +152,7 @@ vprintf(int8 *s, byte *base)
 	if(p > lp)
 		runtime·write(2, lp, p-lp);
 
-//	unlock(&debuglock);
+	//runtime·unlock(&debuglock);
 }
 
 #pragma textflag 7

src/pkg/runtime/proc.c

@@ -15,6 +15,7 @@ static void unwindstack(G*, byte*);
 static void schedule(G*);
 static void acquireproc(void);
 static void releaseproc(void);
+static M *startm(void);
 
 typedef struct Sched Sched;
 
@@ -323,6 +324,9 @@ mcommoninit(M *m)
 	m->fastrand = 0x49f6428aUL + m->id;
 	m->stackalloc = runtime·malloc(sizeof(*m->stackalloc));
 	runtime·FixAlloc_Init(m->stackalloc, FixedStack, runtime·SysAlloc, nil, nil);
+
+	if(m->mcache == nil)
+		m->mcache = runtime·allocmcache();
 }
 
 // Try to increment mcpu.  Report whether succeeded.
@@ -422,7 +426,7 @@ mget(G *g)
 	M *m;
 
 	// if g has its own m, use it.
-	if((m = g->lockedm) != nil)
+	if(g && (m = g->lockedm) != nil)
 		return m;
 
 	// otherwise use general m pool.
@@ -507,6 +511,7 @@ nextgandunlock(void)
 	G *gp;
 	uint32 v;
 
+top:
 	if(atomic_mcpu(runtime·sched.atomic) >= maxgomaxprocs)
 		runtime·throw("negative mcpu");
 
@@ -584,12 +589,49 @@ nextgandunlock(void)
 	schedunlock();
 
 	runtime·notesleep(&m->havenextg);
+	if(m->helpgc) {
+		runtime·gchelper();
+		m->helpgc = 0;
+		runtime·lock(&runtime·sched);
+		goto top;
+	}
 	if((gp = m->nextg) == nil)
 		runtime·throw("bad m->nextg in nextgoroutine");
 	m->nextg = nil;
 	return gp;
 }
 
+int32
+runtime·helpgc(void)
+{
+	M *m;
+	int32 n, max;
+
+	// Figure out how many CPUs to use.
+	// Limited by gomaxprocs, number of actual CPUs, and MaxGcproc.
+	max = runtime·gomaxprocs;
+	if(max > runtime·ncpu)
+		max = runtime·ncpu;
+	if(max > MaxGcproc)
+		max = MaxGcproc;
+
+
+	// We're going to use one CPU no matter what.
+	// Figure out the max number of additional CPUs.
+	max--;
+
+	runtime·lock(&runtime·sched);
+	n = 0;
+	while(n < max && (m = mget(nil)) != nil) {
+		n++;
+		m->helpgc = 1;
+		m->waitnextg = 0;
+		runtime·notewakeup(&m->havenextg);
+	}
+	runtime·unlock(&runtime·sched);
+	return n;
+}
+
 void
 runtime·stoptheworld(void)
 {
@@ -626,15 +668,28 @@ runtime·stoptheworld(void)
 	schedunlock();
 }
 
-// TODO(rsc): Remove. This is only temporary,
-// for the mark and sweep collector.
 void
-runtime·starttheworld(void)
+runtime·starttheworld(bool extra)
 {
+	M *m;
+
 	schedlock();
 	runtime·gcwaiting = 0;
 	setmcpumax(runtime·gomaxprocs);
 	matchmg();
+	if(extra && canaddmcpu()) {
+		// Start a new m that will (we hope) be idle
+		// and so available to help when the next
+		// garbage collection happens.
+		// canaddmcpu above did mcpu++
+		// (necessary, because m will be doing various
+		// initialization work so is definitely running),
+		// but m is not running a specific goroutine,
+		// so set the helpgc flag as a signal to m's
+		// first schedule(nil) to mcpu--.
+		m = startm();
+		m->helpgc = 1;
+	}
 	schedunlock();
 }
 
@@ -644,8 +699,6 @@ runtime·mstart(void)
 {
 	if(g != m->g0)
 		runtime·throw("bad runtime·mstart");
-	if(m->mcache == nil)
-		m->mcache = runtime·allocmcache();
 
 	// Record top of stack for use by mcall.
 	// Once we call schedule we're never coming back,
@@ -677,19 +730,29 @@ struct CgoThreadStart
 static void
 matchmg(void)
 {
-	G *g;
+	G *gp;
+	M *mp;
 
 	if(m->mallocing || m->gcing)
 		return;
 
 	while(haveg() && canaddmcpu()) {
-		g = gget();
-		if(g == nil)
+		gp = gget();
+		if(gp == nil)
 			runtime·throw("gget inconsistency");
 
-		// Find the m that will run g.
+		// Find the m that will run gp.
+		if((mp = mget(gp)) == nil)
+			mp = startm();
+		mnextg(mp, gp);
+	}
+}
+
+static M*
+startm(void)
+{
 	M *m;
-		if((m = mget(g)) == nil){
+
 	m = runtime·malloc(sizeof(M));
 	mcommoninit(m);
 
@@ -712,9 +775,8 @@ matchmg(void)
 			m->g0 = runtime·malg(8192);
 		runtime·newosproc(m, m->g0, m->g0->stackbase, runtime·mstart);
 	}
-		}
-		mnextg(m, g);
-	}
+
+	return m;
 }
 
 // One round of scheduler: find a goroutine and run it.
@@ -767,6 +829,12 @@ schedule(G *gp)
 			gp->readyonstop = 0;
 			readylocked(gp);
 		}
+	} else if(m->helpgc) {
+		// atomic { mcpu-- }
+		v = runtime·xadd(&runtime·sched.atomic, -1<<mcpuShift);
+		if(atomic_mcpu(v) > maxgomaxprocs)
+			runtime·throw("negative mcpu in scheduler");
+		m->helpgc = 0;
 	}
 
 	// Find (or wait for) g to run.  Unlocks runtime·sched.

src/pkg/runtime/runtime.h

@@ -238,6 +238,7 @@ struct	M
 	int32	waitnextg;
 	int32	dying;
 	int32	profilehz;
+	int32	helpgc;
 	uint32	fastrand;
 	uint64	ncgocall;
 	Note	havenextg;
@@ -406,6 +407,7 @@ extern	bool	runtime·singleproc;
 extern	uint32	runtime·panicking;
 extern	int32	runtime·gcwaiting;		// gc is waiting to run
 int8*	runtime·goos;
+int32	runtime·ncpu;
 extern	bool	runtime·iscgo;
 extern	void	(*runtime·destroylock)(Lock*);
 
@@ -515,6 +517,7 @@ void	runtime·startpanic(void);
 void	runtime·sigprof(uint8 *pc, uint8 *sp, uint8 *lr, G *gp);
 void	runtime·resetcpuprofiler(int32);
 void	runtime·setcpuprofilerate(void(*)(uintptr*, int32), int32);
+void	runtime·usleep(uint32);
 
 #pragma	varargck	argpos	runtime·printf	1
 #pragma	varargck	type	"d"	int32
@@ -534,7 +537,7 @@ void	runtime·setcpuprofilerate(void(*)(uintptr*, int32), int32);
 // TODO(rsc): Remove. These are only temporary,
 // for the mark and sweep collector.
 void	runtime·stoptheworld(void);
-void	runtime·starttheworld(void);
+void	runtime·starttheworld(bool);
 
 /*
  * mutual exclusion locks.  in the uncontended case,

test/garbage/Makefile

@@ -18,7 +18,7 @@ all: $(addsuffix .out, $(ALL))
 	$(LD) -o $@ $*.$O
 
 %.bench: %.out
-	./$*.out
+	time ./$*.out
 
 bench: $(addsuffix .bench, $(ALL))
 

test/garbage/parser.go

@@ -73,10 +73,6 @@ func parseDir(dirpath string) map[string]*ast.Package {
 }
 
 func main() {
-	runtime.GOMAXPROCS(4)
-	go func() {}()
-	go func() {}()
-	go func() {}()
 	st := &runtime.MemStats
 	packages = append(packages, packages...)
 	packages = append(packages, packages...)
@@ -132,7 +128,6 @@ func main() {
 	}
 }
 
-
 var packages = []string{
 	"archive/tar",
 	"asn1",
@@ -148,7 +143,6 @@ var packages = []string{
 	"container/ring",
 	"container/vector",
 	"crypto/aes",
-	"crypto/block",
 	"crypto/blowfish",
 	"crypto/hmac",
 	"crypto/md4",
@@ -167,7 +161,6 @@ var packages = []string{
 	"debug/macho",
 	"debug/elf",
 	"debug/gosym",
-	"debug/proc",
 	"ebnf",
 	"encoding/ascii85",
 	"encoding/base64",
@@ -177,9 +170,6 @@ var packages = []string{
 	"encoding/pem",
 	"exec",
 	"exp/datafmt",
-	"exp/draw",
-	"exp/eval",
-	"exp/iterable",
 	"expvar",
 	"flag",
 	"fmt",