diff --git a/src/cmd/6a/lex.c b/src/cmd/6a/lex.c
index 5009a7b832f4c..42f4b1d11d396 100644
--- a/src/cmd/6a/lex.c
+++ b/src/cmd/6a/lex.c
@@ -527,6 +527,7 @@ struct
 	"OUTSB",	LTYPE0,	AOUTSB,
 	"OUTSL",	LTYPE0,	AOUTSL,
 	"OUTSW",	LTYPE0,	AOUTSW,
+	"PAUSE",	LTYPEN,	APAUSE,
 	"POPAL",	LTYPE0,	APOPAL,
 	"POPAW",	LTYPE0,	APOPAW,
 	"POPFL",	LTYPE0,	APOPFL,
diff --git a/src/cmd/6l/6.out.h b/src/cmd/6l/6.out.h
index 24fede53c2c74..262da02abd6de 100644
--- a/src/cmd/6l/6.out.h
+++ b/src/cmd/6l/6.out.h
@@ -190,6 +190,7 @@ enum	as
 	AOUTSB,
 	AOUTSL,
 	AOUTSW,
+	APAUSE,
 	APOPAL,
 	APOPAW,
 	APOPFL,
diff --git a/src/cmd/6l/optab.c b/src/cmd/6l/optab.c
index 928ad5d9ae73d..36806ec4b625d 100644
--- a/src/cmd/6l/optab.c
+++ b/src/cmd/6l/optab.c
@@ -919,6 +919,7 @@ Optab optab[] =
 	{ APADDW,	ymm,	Py, 0xfd,Pe,0xfd },
 	{ APAND,	ymm,	Py, 0xdb,Pe,0xdb },
 	{ APANDN,	ymm,	Py, 0xdf,Pe,0xdf },
+	{ APAUSE,	ynone,	Px, 0xf3,0x90 },
 	{ APAVGB,	ymm,	Py, 0xe0,Pe,0xe0 },
 	{ APAVGW,	ymm,	Py, 0xe3,Pe,0xe3 },
 	{ APCMPEQB,	ymm,	Py, 0x74,Pe,0x74 },
diff --git a/src/cmd/8a/lex.c b/src/cmd/8a/lex.c
index ab4de417a506c..e56460e4bd2d7 100644
--- a/src/cmd/8a/lex.c
+++ b/src/cmd/8a/lex.c
@@ -421,6 +421,7 @@ struct
 	"OUTSB",	LTYPE0,	AOUTSB,
 	"OUTSL",	LTYPE0,	AOUTSL,
 	"OUTSW",	LTYPE0,	AOUTSW,
+	"PAUSE",	LTYPEN,	APAUSE,
 	"POPAL",	LTYPE0,	APOPAL,
 	"POPAW",	LTYPE0,	APOPAW,
 	"POPFL",	LTYPE0,	APOPFL,
diff --git a/src/cmd/8l/8.out.h b/src/cmd/8l/8.out.h
index 03db0016b591d..9a8483aaf4fbc 100644
--- a/src/cmd/8l/8.out.h
+++ b/src/cmd/8l/8.out.h
@@ -180,6 +180,7 @@ enum	as
 	AOUTSB,
 	AOUTSL,
 	AOUTSW,
+	APAUSE,
 	APOPAL,
 	APOPAW,
 	APOPFL,
diff --git a/src/cmd/8l/optab.c b/src/cmd/8l/optab.c
index 1e89a21053842..f5c195d75f308 100644
--- a/src/cmd/8l/optab.c
+++ b/src/cmd/8l/optab.c
@@ -495,6 +495,7 @@ Optab optab[] =
 	{ AOUTSB,	ynone,	Pb, 0x6e },
 	{ AOUTSL,	ynone,	Px, 0x6f },
 	{ AOUTSW,	ynone,	Pe, 0x6f },
+	{ APAUSE,	ynone,	Px, 0xf3,0x90 },
 	{ APOPAL,	ynone,	Px, 0x61 },
 	{ APOPAW,	ynone,	Pe, 0x61 },
 	{ APOPFL,	ynone,	Px, 0x9d },
diff --git a/src/pkg/runtime/386/asm.s b/src/pkg/runtime/386/asm.s
index 24e64a11e5a64..2505e4df6a9f4 100644
--- a/src/pkg/runtime/386/asm.s
+++ b/src/pkg/runtime/386/asm.s
@@ -334,6 +334,20 @@ TEXT runtime·xadd(SB), 7, $0
 	ADDL	CX, AX
 	RET
 
+TEXT runtime·xchg(SB), 7, $0
+	MOVL	4(SP), BX
+	MOVL	8(SP), AX
+	XCHGL	AX, 0(BX)
+	RET
+
+TEXT runtime·procyield(SB),7,$0
+	MOVL	4(SP), AX
+again:
+	PAUSE
+	SUBL	$1, AX
+	JNZ	again
+	RET
+
 TEXT runtime·atomicstorep(SB), 7, $0
 	MOVL	4(SP), BX
 	MOVL	8(SP), AX
diff --git a/src/pkg/runtime/amd64/asm.s b/src/pkg/runtime/amd64/asm.s
index 6ac84c408d601..4723018a7aa9f 100644
--- a/src/pkg/runtime/amd64/asm.s
+++ b/src/pkg/runtime/amd64/asm.s
@@ -378,6 +378,20 @@ TEXT runtime·xadd(SB), 7, $0
 	ADDL	CX, AX
 	RET
 
+TEXT runtime·xchg(SB), 7, $0
+	MOVQ	8(SP), BX
+	MOVL	16(SP), AX
+	XCHGL	AX, 0(BX)
+	RET
+
+TEXT runtime·procyield(SB),7,$0
+	MOVL	8(SP), AX
+again:
+	PAUSE
+	SUBL	$1, AX
+	JNZ	again
+	RET
+
 TEXT runtime·atomicstorep(SB), 7, $0
 	MOVQ	8(SP), BX
 	MOVQ	16(SP), AX
diff --git a/src/pkg/runtime/arm/atomic.c b/src/pkg/runtime/arm/atomic.c
index d229e8c347584..3199afe622787 100644
--- a/src/pkg/runtime/arm/atomic.c
+++ b/src/pkg/runtime/arm/atomic.c
@@ -19,6 +19,29 @@ runtime·xadd(uint32 volatile *val, int32 delta)
 	}
 }
 
+#pragma textflag 7
+uint32
+runtime·xchg(uint32 volatile* addr, uint32 v)
+{
+	uint32 old;
+
+	for(;;) {
+		old = *addr;
+		if(runtime·cas(addr, old, v))
+			return old;
+	}
+}
+
+#pragma textflag 7
+void
+runtime·procyield(uint32 cnt)
+{
+	uint32 volatile i;
+
+	for(i = 0; i < cnt; i++) {
+	}
+}
+
 #pragma textflag 7
 uint32
 runtime·atomicload(uint32 volatile* addr)
diff --git a/src/pkg/runtime/linux/386/defs.h b/src/pkg/runtime/linux/386/defs.h
index 6ae1c4e139acd..73fe23ef98b28 100644
--- a/src/pkg/runtime/linux/386/defs.h
+++ b/src/pkg/runtime/linux/386/defs.h
@@ -61,6 +61,8 @@ enum {
 	ITIMER_REAL = 0,
 	ITIMER_VIRTUAL = 0x1,
 	ITIMER_PROF = 0x2,
+	O_RDONLY = 0,
+	O_CLOEXEC = 02000000,
 };
 
 // Types
diff --git a/src/pkg/runtime/linux/386/sys.s b/src/pkg/runtime/linux/386/sys.s
index e8b4233242862..0b4a34986c187 100644
--- a/src/pkg/runtime/linux/386/sys.s
+++ b/src/pkg/runtime/linux/386/sys.s
@@ -22,9 +22,31 @@ TEXT runtime·exit1(SB),7,$0
 	INT $3	// not reached
 	RET
 
+TEXT runtime·open(SB),7,$0
+	MOVL	$5, AX		// syscall - open
+	MOVL	4(SP), BX
+	MOVL	8(SP), CX
+	MOVL	12(SP), DX
+	INT	$0x80
+	RET
+
+TEXT runtime·close(SB),7,$0
+	MOVL	$6, AX		// syscall - close
+	MOVL	4(SP), BX
+	INT	$0x80
+	RET
+
 TEXT runtime·write(SB),7,$0
 	MOVL	$4, AX		// syscall - write
-	MOVL	4(SP),  BX
+	MOVL	4(SP), BX
+	MOVL	8(SP), CX
+	MOVL	12(SP), DX
+	INT	$0x80
+	RET
+
+TEXT runtime·read(SB),7,$0
+	MOVL	$3, AX		// syscall - read
+	MOVL	4(SP), BX
 	MOVL	8(SP), CX
 	MOVL	12(SP), DX
 	INT	$0x80
@@ -315,3 +337,8 @@ TEXT runtime·setldt(SB),7,$32
 	MOVW	AX, GS
 
 	RET
+
+TEXT runtime·osyield(SB),7,$0
+	MOVL	$158, AX
+	INT	$0x80
+	RET
diff --git a/src/pkg/runtime/linux/amd64/defs.h b/src/pkg/runtime/linux/amd64/defs.h
index 70d63145c694e..8053dd16fe954 100644
--- a/src/pkg/runtime/linux/amd64/defs.h
+++ b/src/pkg/runtime/linux/amd64/defs.h
@@ -61,6 +61,8 @@ enum {
 	ITIMER_REAL = 0,
 	ITIMER_VIRTUAL = 0x1,
 	ITIMER_PROF = 0x2,
+	O_RDONLY = 0,
+	O_CLOEXEC = 02000000,
 };
 
 // Types
diff --git a/src/pkg/runtime/linux/amd64/sys.s b/src/pkg/runtime/linux/amd64/sys.s
index 66fdab2083db2..8b4dcd921e8fd 100644
--- a/src/pkg/runtime/linux/amd64/sys.s
+++ b/src/pkg/runtime/linux/amd64/sys.s
@@ -28,6 +28,12 @@ TEXT runtime·open(SB),7,$0-16
 	SYSCALL
 	RET
 
+TEXT runtime·close(SB),7,$0-16
+	MOVL	8(SP), DI
+	MOVL	$3, AX			// syscall entry
+	SYSCALL
+	RET
+
 TEXT runtime·write(SB),7,$0-24
 	MOVL	8(SP), DI
 	MOVQ	16(SP), SI
@@ -36,6 +42,14 @@ TEXT runtime·write(SB),7,$0-24
 	SYSCALL
 	RET
 
+TEXT runtime·read(SB),7,$0-24
+	MOVL	8(SP), DI
+	MOVQ	16(SP), SI
+	MOVL	24(SP), DX
+	MOVL	$0, AX			// syscall entry
+	SYSCALL
+	RET
+
 TEXT runtime·raisesigpipe(SB),7,$12
 	MOVL	$186, AX	// syscall - gettid
 	SYSCALL
@@ -232,3 +246,7 @@ TEXT runtime·settls(SB),7,$32
 	CALL	runtime·notok(SB)
 	RET
 
+TEXT runtime·osyield(SB),7,$0
+	MOVL	$24, AX
+	SYSCALL
+	RET
diff --git a/src/pkg/runtime/linux/arm/defs.h b/src/pkg/runtime/linux/arm/defs.h
index 6b2f22c66ada7..09b558ed0f882 100644
--- a/src/pkg/runtime/linux/arm/defs.h
+++ b/src/pkg/runtime/linux/arm/defs.h
@@ -61,6 +61,8 @@ enum {
 	ITIMER_REAL = 0,
 	ITIMER_PROF = 0x2,
 	ITIMER_VIRTUAL = 0x1,
+	O_RDONLY = 0,
+	O_CLOEXEC = 02000000,
 };
 
 // Types
diff --git a/src/pkg/runtime/linux/arm/sys.s b/src/pkg/runtime/linux/arm/sys.s
index ab53498222ff6..8619f0945caa9 100644
--- a/src/pkg/runtime/linux/arm/sys.s
+++ b/src/pkg/runtime/linux/arm/sys.s
@@ -15,7 +15,10 @@
 #define SYS_BASE 0x0
 
 #define SYS_exit (SYS_BASE + 1)
+#define SYS_read (SYS_BASE + 3)
 #define SYS_write (SYS_BASE + 4)
+#define SYS_open (SYS_BASE + 5)
+#define SYS_close (SYS_BASE + 6)
 #define SYS_gettimeofday (SYS_BASE + 78)
 #define SYS_clone (SYS_BASE + 120)
 #define SYS_rt_sigreturn (SYS_BASE + 173)
@@ -29,10 +32,25 @@
 #define SYS_mincore (SYS_BASE + 219)
 #define SYS_gettid (SYS_BASE + 224)
 #define SYS_tkill (SYS_BASE + 238)
+#define SYS_sched_yield (SYS_BASE + 158)
 
 #define ARM_BASE (SYS_BASE + 0x0f0000)
 #define SYS_ARM_cacheflush (ARM_BASE + 2)
 
+TEXT runtime·open(SB),7,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_open, R7
+	SWI	$0
+	RET
+
+TEXT runtime·close(SB),7,$0
+	MOVW	0(FP), R0
+	MOVW	$SYS_close, R7
+	SWI	$0
+	RET
+
 TEXT runtime·write(SB),7,$0
 	MOVW	0(FP), R0
 	MOVW	4(FP), R1
@@ -41,6 +59,14 @@ TEXT runtime·write(SB),7,$0
 	SWI	$0
 	RET
 
+TEXT runtime·read(SB),7,$0
+	MOVW	0(FP), R0
+	MOVW	4(FP), R1
+	MOVW	8(FP), R2
+	MOVW	$SYS_read, R7
+	SWI	$0
+	RET
+
 TEXT runtime·exit(SB),7,$-4
 	MOVW	0(FP), R0
 	MOVW	$SYS_exit_group, R7
@@ -287,3 +313,7 @@ cascheck:
 TEXT runtime·casp(SB),7,$0
 	B	runtime·cas(SB)
 
+TEXT runtime·osyield(SB),7,$0
+	MOVW	$SYS_sched_yield, R7
+	SWI	$0
+	RET
diff --git a/src/pkg/runtime/linux/thread.c b/src/pkg/runtime/linux/thread.c
index 7c7ca7b4e10ca..8efba2b98b216 100644
--- a/src/pkg/runtime/linux/thread.c
+++ b/src/pkg/runtime/linux/thread.c
@@ -8,6 +8,11 @@
 #include "stack.h"
 
 extern SigTab runtime·sigtab[];
+static int32 proccount;
+
+int32 runtime·open(uint8*, int32, int32);
+int32 runtime·close(int32);
+int32 runtime·read(int32, void*, int32);
 
 // Linux futex.
 //
@@ -15,11 +20,19 @@ extern SigTab runtime·sigtab[];
 //	futexwakeup(uint32 *addr)
 //
 // Futexsleep atomically checks if *addr == val and if so, sleeps on addr.
-// Futexwakeup wakes up one thread sleeping on addr.
+// Futexwakeup wakes up threads sleeping on addr.
 // Futexsleep is allowed to wake up spuriously.
 
 enum
 {
+	MUTEX_UNLOCKED = 0,
+	MUTEX_LOCKED = 1,
+	MUTEX_SLEEPING = 2,
+
+	ACTIVE_SPIN = 4,
+	ACTIVE_SPIN_CNT = 30,
+	PASSIVE_SPIN = 1,
+
 	FUTEX_WAIT = 0,
 	FUTEX_WAKE = 1,
 
@@ -52,13 +65,13 @@ futexsleep(uint32 *addr, uint32 val)
 	runtime·futex(addr, FUTEX_WAIT, val, &longtime, nil, 0);
 }
 
-// If any procs are sleeping on addr, wake up at least one.
+// If any procs are sleeping on addr, wake up at most cnt.
 static void
-futexwakeup(uint32 *addr)
+futexwakeup(uint32 *addr, uint32 cnt)
 {
 	int64 ret;
 
-	ret = runtime·futex(addr, FUTEX_WAKE, 1, nil, nil, 0);
+	ret = runtime·futex(addr, FUTEX_WAKE, cnt, nil, nil, 0);
 
 	if(ret >= 0)
 		return;
@@ -66,70 +79,96 @@ futexwakeup(uint32 *addr)
 	// I don't know that futex wakeup can return
 	// EAGAIN or EINTR, but if it does, it would be
 	// safe to loop and call futex again.
-
-	runtime·prints("futexwakeup addr=");
-	runtime·printpointer(addr);
-	runtime·prints(" returned ");
-	runtime·printint(ret);
-	runtime·prints("\n");
+	runtime·printf("futexwakeup addr=%p returned %D\n", addr, ret);
 	*(int32*)0x1006 = 0x1006;
 }
 
+static int32
+getproccount(void)
+{
+	int32 fd, rd, cnt, cpustrlen;
+	byte *cpustr, *pos, *bufpos;
+	byte buf[256];
+
+	fd = runtime·open((byte*)"/proc/stat", O_RDONLY|O_CLOEXEC, 0);
+	if(fd == -1)
+		return 1;
+	cnt = 0;
+	bufpos = buf;
+	cpustr = (byte*)"\ncpu";
+	cpustrlen = runtime·findnull(cpustr);
+	for(;;) {
+		rd = runtime·read(fd, bufpos, sizeof(buf)-cpustrlen);
+		if(rd == -1)
+			break;
+		bufpos[rd] = 0;
+		for(pos=buf; pos=runtime·strstr(pos, cpustr); cnt++, pos++) {
+		}
+		if(rd < cpustrlen)
+			break;
+		runtime·memmove(buf, bufpos+rd-cpustrlen+1, cpustrlen-1);
+		bufpos = buf+cpustrlen-1;
+	}
+	runtime·close(fd);
+	return cnt ? cnt : 1;
+}
 
-// Lock and unlock.
-//
-// The lock state is a single 32-bit word that holds
-// a 31-bit count of threads waiting for the lock
-// and a single bit (the low bit) saying whether the lock is held.
-// The uncontended case runs entirely in user space.
-// When contention is detected, we defer to the kernel (futex).
-//
-// A reminder: compare-and-swap runtime·cas(addr, old, new) does
-//	if(*addr == old) { *addr = new; return 1; }
-//	else return 0;
-// but atomically.
-
+// Possible lock states are MUTEX_UNLOCKED, MUTEX_LOCKED and MUTEX_SLEEPING.
+// MUTEX_SLEEPING means that there is presumably at least one sleeping thread.
+// Note that there can be spinning threads during all states - they do not
+// affect mutex's state.
 static void
 futexlock(Lock *l)
 {
-	uint32 v;
+	uint32 i, v, wait, spin;
 
-again:
-	v = l->key;
-	if((v&1) == 0){
-		if(runtime·cas(&l->key, v, v|1)){
-			// Lock wasn't held; we grabbed it.
-			return;
+	// Speculative grab for lock.
+	v = runtime·xchg(&l->key, MUTEX_LOCKED);
+	if(v == MUTEX_UNLOCKED)
+		return;
+
+	// wait is either MUTEX_LOCKED or MUTEX_SLEEPING
+	// depending on whether there is a thread sleeping
+	// on this mutex.  If we ever change l->key from
+	// MUTEX_SLEEPING to some other value, we must be
+	// careful to change it back to MUTEX_SLEEPING before
+	// returning, to ensure that the sleeping thread gets
+	// 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)
+		spin = ACTIVE_SPIN;
+
+	for(;;) {
+		// Try for lock, spinning.
+		for(i = 0; i < spin; i++) {
+			while(l->key == MUTEX_UNLOCKED)
+				if(runtime·cas(&l->key, MUTEX_UNLOCKED, wait))
+						return;
+			runtime·procyield(ACTIVE_SPIN_CNT);
 		}
-		goto again;
-	}
 
-	// Lock was held; try to add ourselves to the waiter count.
-	if(!runtime·cas(&l->key, v, v+2))
-		goto again;
-
-	// We're accounted for, now sleep in the kernel.
-	//
-	// We avoid the obvious lock/unlock race because
-	// the kernel won't put us to sleep if l->key has
-	// changed underfoot and is no longer v+2.
-	//
-	// We only really care that (v&1) == 1 (the lock is held),
-	// and in fact there is a futex variant that could
-	// accommodate that check, but let's not get carried away.)
-	futexsleep(&l->key, v+2);
-
-	// We're awake: remove ourselves from the count.
-	for(;;){
-		v = l->key;
-		if(v < 2)
-			runtime·throw("bad lock key");
-		if(runtime·cas(&l->key, v, v-2))
-			break;
-	}
+		// Try for lock, rescheduling.
+		for(i=0; i < PASSIVE_SPIN; i++) {
+			while(l->key == MUTEX_UNLOCKED)
+				if(runtime·cas(&l->key, MUTEX_UNLOCKED, wait))
+					return;
+			runtime·osyield();
+		}
 
-	// Try for the lock again.
-	goto again;
+		// Sleep.
+		v = runtime·xchg(&l->key, MUTEX_SLEEPING);
+		if(v == MUTEX_UNLOCKED)
+			return;
+		wait = MUTEX_SLEEPING;
+		futexsleep(&l->key, MUTEX_SLEEPING);
+	}
 }
 
 static void
@@ -137,34 +176,26 @@ futexunlock(Lock *l)
 {
 	uint32 v;
 
-	// Atomically get value and clear lock bit.
-again:
-	v = l->key;
-	if((v&1) == 0)
+	v = runtime·xchg(&l->key, MUTEX_UNLOCKED);
+	if(v == MUTEX_UNLOCKED)
 		runtime·throw("unlock of unlocked lock");
-	if(!runtime·cas(&l->key, v, v&~1))
-		goto again;
-
-	// If there were waiters, wake one.
-	if(v & ~1)
-		futexwakeup(&l->key);
+	if(v == MUTEX_SLEEPING)
+		futexwakeup(&l->key, 1);
 }
 
 void
 runtime·lock(Lock *l)
 {
-	if(m->locks < 0)
-		runtime·throw("lock count");
-	m->locks++;
+	if(m->locks++ < 0)
+		runtime·throw("runtime·lock: lock count");
 	futexlock(l);
 }
 
 void
 runtime·unlock(Lock *l)
 {
-	m->locks--;
-	if(m->locks < 0)
-		runtime·throw("lock count");
+	if(--m->locks < 0)
+		runtime·throw("runtime·unlock: lock count");
 	futexunlock(l);
 }
 
@@ -175,35 +206,24 @@ runtime·destroylock(Lock*)
 
 
 // One-time notifications.
-//
-// Since the lock/unlock implementation already
-// takes care of sleeping in the kernel, we just reuse it.
-// (But it's a weird use, so it gets its own interface.)
-//
-// We use a lock to represent the event:
-// unlocked == event has happened.
-// Thus the lock starts out locked, and to wait for the
-// event you try to lock the lock.  To signal the event,
-// you unlock the lock.
-
 void
 runtime·noteclear(Note *n)
 {
-	n->lock.key = 0;	// memset(n, 0, sizeof *n)
-	futexlock(&n->lock);
+	n->state = 0;
 }
 
 void
 runtime·notewakeup(Note *n)
 {
-	futexunlock(&n->lock);
+	runtime·xchg(&n->state, 1);
+	futexwakeup(&n->state, 1<<30);
 }
 
 void
 runtime·notesleep(Note *n)
 {
-	futexlock(&n->lock);
-	futexunlock(&n->lock);	// Let other sleepers find out too.
+	while(runtime·atomicload(&n->state) == 0)
+		futexsleep(&n->state, 0);
 }
 
 
diff --git a/src/pkg/runtime/runtime.h b/src/pkg/runtime/runtime.h
index 15b1e8eb9d8c1..eee346844bf0b 100644
--- a/src/pkg/runtime/runtime.h
+++ b/src/pkg/runtime/runtime.h
@@ -131,7 +131,10 @@ struct	Usema
 union	Note
 {
 	struct {	// Linux
-		Lock	lock;
+		uint32	state;
+	};
+	struct {	// Windows
+		Lock lock;
 	};
 	struct {	// OS X
 		int32	wakeup;
@@ -382,6 +385,7 @@ extern	bool	runtime·iscgo;
  * common functions and data
  */
 int32	runtime·strcmp(byte*, byte*);
+byte*	runtime·strstr(byte*, byte*);
 int32	runtime·findnull(byte*);
 int32	runtime·findnullw(uint16*);
 void	runtime·dump(byte*, int32);
@@ -427,6 +431,7 @@ bool	runtime·casp(void**, void*, void*);
 // Don't confuse with XADD x86 instruction,
 // this one is actually 'addx', that is, add-and-fetch.
 uint32	runtime·xadd(uint32 volatile*, int32);
+uint32	runtime·xchg(uint32 volatile*, uint32);
 uint32	runtime·atomicload(uint32 volatile*);
 void*	runtime·atomicloadp(void* volatile*);
 void	runtime·atomicstorep(void* volatile*, void*);
@@ -596,6 +601,8 @@ void	runtime·semacquire(uint32*);
 void	runtime·semrelease(uint32*);
 String	runtime·signame(int32 sig);
 int32	runtime·gomaxprocsfunc(int32 n);
+void	runtime·procyield(uint32);
+void	runtime·osyield(void);
 
 void	runtime·mapassign(Hmap*, byte*, byte*);
 void	runtime·mapaccess(Hmap*, byte*, byte*, bool*);
diff --git a/src/pkg/runtime/string.goc b/src/pkg/runtime/string.goc
index e0daac49ab6b0..48bf3183b5257 100644
--- a/src/pkg/runtime/string.goc
+++ b/src/pkg/runtime/string.goc
@@ -203,6 +203,28 @@ runtime·strcmp(byte *s1, byte *s2)
 	}
 }
 
+byte*
+runtime·strstr(byte *s1, byte *s2)
+{
+	byte *sp1, *sp2;
+
+	if(*s2 == 0)
+		return s1;
+	for(; *s1; s1++) {
+		if(*s1 != *s2)
+			continue;
+		sp1 = s1;
+		sp2 = s2;
+		for(;;) {
+			if(*sp2 == 0)
+				return s1;
+			if(*sp1++ != *sp2++)
+				break;
+		}
+	}
+	return nil;
+}
+
 func slicestring(si String, lindex int32, hindex int32) (so String) {
 	int32 l;