new internal locking primitive; drop spinlocks

we use priority inheritance futexes if possible so that the library
cannot hit internal priority inversion deadlocks in the presence of
realtime priority scheduling (full support to be added later).

src/internal/libc.h

@@ -45,10 +45,11 @@ extern struct __libc *__libc_loc(void) __attribute__((const));
 
 /* Designed to avoid any overhead in non-threaded processes */
 void __lock(volatile int *);
+void __unlock(volatile int *);
 int __lockfile(FILE *);
 void __unlockfile(FILE *);
 #define LOCK(x) (libc.threads_minus_1 ? (__lock(x),1) : ((void)(x),1))
-#define UNLOCK(x) (*(volatile int *)(x)=0)
+#define UNLOCK(x) (libc.threads_minus_1 ? (__unlock(x),1) : ((void)(x),1))
 
 void __synccall(void (*)(void *), void *);
 void __synccall_wait(void);

src/thread/__lock.c

@@ -1,11 +1,32 @@
 #include "pthread_impl.h"
 
+void __lock_2(volatile int *l)
+{
+	if (!__syscall(SYS_futex, l, FUTEX_LOCK_PI, 0, 0))
+		return;
+	int old, tid = __pthread_self()->tid;
+	while ((old = a_cas(l, 0, tid))) {
+		a_cas(l, old, old|INT_MIN);
+		__syscall(SYS_futex, l, FUTEX_WAIT, old|INT_MIN, 0);
+	}
+}
+
 void __lock(volatile int *l)
 {
-	int spins=10000;
-	/* Do not use futexes because we insist that unlocking is a simple
-	 * assignment to optimize non-pathological code with no contention. */
-	while (a_swap(l, 1))
-		if (spins) spins--, a_spin();
-		else __syscall(SYS_sched_yield);
+	if (a_cas(l, 0, __pthread_self()->tid)) __lock_2(l);
+}
+
+void __unlock_2(volatile int *l)
+{
+	if (__syscall(SYS_futex, l, FUTEX_UNLOCK_PI)) {
+		*l = 0;
+		__syscall(SYS_futex, l, FUTEX_WAKE, 1);
+	}
+}
+
+void __unlock(volatile int *l)
+{
+	int old = *l;
+	if (!(old & INT_MIN) && a_cas(l, old, 0)==old) return;
+	__unlock_2(l);
 }