major semaphore improvements (performance and correctness)

1. make sem_[timed]wait interruptible by signals, per POSIX
2. keep a waiter count in order to avoid unnecessary futex wake syscalls

src/thread/__timedwait.c

@@ -18,6 +18,6 @@ int __timedwait(volatile int *addr, int val, clockid_t clk, const struct timespe
 	}
 	if (priv) priv = 128; priv=0;
 	r = -__syscall(__NR_futex, (long)addr, FUTEX_WAIT | priv, val, at ? (long)&to : 0);
-	if (r == ETIMEDOUT) return r;
+	if (r == ETIMEDOUT || r == EINTR) return r;
 	return 0;
 }

src/thread/pthread_cond_timedwait.c

@@ -16,7 +16,8 @@ int pthread_cond_timedwait(pthread_cond_t *c, pthread_mutex_t *m, const struct t
 	if ((r=pthread_mutex_unlock(m))) return r;
 
 	CANCELPT_BEGIN;
-	e = __timedwait(&c->_c_block, 1, c->_c_clock, ts, 0);
+	do e = __timedwait(&c->_c_block, 1, c->_c_clock, ts, 0);
+	while (e == EINTR);
 	CANCELPT_END;
 
 	pthread_cleanup_pop(0);

src/thread/sem_post.c

@@ -3,7 +3,8 @@
 
 int sem_post(sem_t *sem)
 {
-	if (!a_fetch_add(sem->__val, 1))
+	a_inc(sem->__val);
+	if (sem->__val[1])
 		__wake(sem->__val, 1, 0);
 	return 0;
 }

src/thread/sem_timedwait.c

@@ -1,26 +1,41 @@
 #include <semaphore.h>
 #include "pthread_impl.h"
 
+static void cleanup(void *p)
+{
+	a_dec(p);
+}
+
 int sem_timedwait(sem_t *sem, const struct timespec *at)
 {
-	int val;
+	int r;
 
+	if (a_fetch_add(sem->__val, -1) > 0) return 0;
+	a_inc(sem->__val);
+
+	if (at && at->tv_nsec >= 1000000000UL) {
+		errno = EINVAL;
+		return -1;
+	}
+
+	a_inc(sem->__val+1);
+	pthread_cleanup_push(cleanup, sem->__val+1)
+
+	CANCELPT_BEGIN;
 	for (;;) {
-		if (a_fetch_add(sem->__val, -1) > 0) return 0;
-		val = a_fetch_add(sem->__val, 1)+1;
-		if (val==1) __wake(sem->__val, 1, 0);
-		if (at && at->tv_nsec >= 1000000000UL) {
-			errno = EINVAL;
-			return -1;
-		}
-		CANCELPT_BEGIN;
-		if (val <= 0 && __timedwait(sem->__val, val, CLOCK_REALTIME, at, 0) == ETIMEDOUT) {
-			errno = ETIMEDOUT;
-			CANCELPT_TRY;
-			CANCELPT_END;
-			return -1;
+		r = 0;
+		if (!sem_trywait(sem)) break;
+		r = __timedwait(sem->__val, 0, CLOCK_REALTIME, at, 0);
+		if (r) {
+			errno = r;
+			r = -1;
+			break;
 		}
 		CANCELPT_TRY;
-		CANCELPT_END;
 	}
+	CANCELPT_END;
+
+	pthread_cleanup_pop(1);
+
+	return r;
 }

src/thread/sem_trywait.c

@@ -3,9 +3,8 @@
 
 int sem_trywait(sem_t *sem)
 {
-	int val = a_fetch_add(sem->__val, -1);
-	if (val > 0) return 0;
-	if (!a_fetch_add(sem->__val, 1))
+	if (a_fetch_add(sem->__val, -1) > 0) return 0;
+	if (!a_fetch_add(sem->__val, 1) && sem->__val[1])
 		__wake(sem->__val, 1, 0);
 	errno = EAGAIN;
 	return -1;