process-shared barrier support, based on discussion with bdonlan

this implementation is rather heavy-weight, but it's the first
solution i've found that's actually correct. all waiters actually wait
twice at the barrier so that they can synchronize exit, and they hold
a "vm lock" that prevents changes to virtual memory mappings (and
blocks pthread_barrier_destroy) until all waiters are finished
inspecting the barrier.

thus, it is safe for any thread to destroy and/or unmap the barrier's
memory as soon as pthread_barrier_wait returns, without further
synchronization.

src/internal/pthread_impl.h

@@ -73,10 +73,12 @@ struct __timer {
 #define _c_waiters2 __u.__i[7]
 #define _rw_lock __u.__i[0]
 #define _rw_waiters __u.__i[1]
-#define _b_inst __u.__p[0]
+#define _b_lock __u.__i[0]
+#define _b_waiters __u.__i[1]
 #define _b_limit __u.__i[2]
-#define _b_lock __u.__i[3]
-#define _b_waiters __u.__i[4]
+#define _b_count __u.__i[3]
+#define _b_seq __u.__i[4]
+#define _b_inst __u.__p[4]
 
 #include "pthread_arch.h"
 

src/mman/mmap.c

@@ -5,16 +5,25 @@
 #include "syscall.h"
 #include "libc.h"
 
+static void dummy1(int x) { }
+static void dummy0(void) { }
+weak_alias(dummy1, __vm_lock);
+weak_alias(dummy0, __vm_unlock);
+
 void *__mmap(void *start, size_t len, int prot, int flags, int fd, off_t off)
 {
+	void *ret;
 	if (sizeof(off_t) > sizeof(long))
 		if (((long)off & 0xfff) | ((long)((unsigned long long)off>>(12 + 8*(sizeof(off_t)-sizeof(long))))))
 			start = (void *)-1;
+	if (flags & MAP_FIXED) __vm_lock(-1);
 #ifdef SYS_mmap2
-	return (void *)syscall(SYS_mmap2, start, len, prot, flags, fd, off>>12);
+	ret = (void *)syscall(SYS_mmap2, start, len, prot, flags, fd, off>>12);
 #else
-	return (void *)syscall(SYS_mmap, start, len, prot, flags, fd, off);
+	ret = (void *)syscall(SYS_mmap, start, len, prot, flags, fd, off);
 #endif
+	if (flags & MAP_FIXED) __vm_unlock();
+	return ret;
 }
 
 weak_alias(__mmap, mmap);

src/mman/munmap.c

@@ -3,9 +3,18 @@
 #include "syscall.h"
 #include "libc.h"
 
+static void dummy1(int x) { }
+static void dummy0(void) { }
+weak_alias(dummy1, __vm_lock);
+weak_alias(dummy0, __vm_unlock);
+
 int __munmap(void *start, size_t len)
 {
-	return syscall(SYS_munmap, start, len);
+	int ret;
+	__vm_lock(-1);
+	ret = syscall(SYS_munmap, start, len);
+	__vm_unlock();
+	return ret;
 }
 
 weak_alias(__munmap, munmap);

src/thread/pthread_barrier_destroy.c

 #include "pthread_impl.h"
 
+void __vm_lock(int), __vm_unlock(void);
+
 int pthread_barrier_destroy(pthread_barrier_t *b)
 {
+	if (b->_b_limit < 0) {
+		__vm_lock(-1);
+		__vm_unlock();
+	}
 	return 0;
 }

src/thread/pthread_barrier_init.c

@@ -2,7 +2,7 @@
 
 int pthread_barrier_init(pthread_barrier_t *b, const pthread_barrierattr_t *a, unsigned count)
 {
-	if (!count) return EINVAL;
-	*b = (pthread_barrier_t){ ._b_limit = count-1 };
+	if (count-1 > INT_MAX-1) return EINVAL;
+	*b = (pthread_barrier_t){ ._b_limit = count-1 | (a?*a:0) };
 	return 0;
 }

src/thread/pthread_barrier_wait.c

 #include "pthread_impl.h"
 
+static int vmlock[2];
+
+void __vm_lock(int inc)
+{
+	for (;;) {
+		int v = vmlock[0];
+		if (inc*v < 0) __wait(vmlock, vmlock+1, v, 1);
+		else if (a_cas(vmlock, v, v+inc)==v) break;
+	}
+}
+
+void __vm_unlock(void)
+{
+	if (vmlock[0]>0) a_dec(vmlock);
+	else a_inc(vmlock);
+	if (vmlock[1]) __wake(vmlock, 1, 1);
+}
+
+static int pshared_barrier_wait(pthread_barrier_t *b)
+{
+	int limit = (b->_b_limit & INT_MAX) + 1;
+	int seq;
+	int ret = 0;
+
+	if (!limit) return PTHREAD_BARRIER_SERIAL_THREAD;
+
+	while (a_swap(&b->_b_lock, 1))
+		__wait(&b->_b_lock, &b->_b_waiters, 1, 0);
+
+	seq = b->_b_seq;
+
+	if (++b->_b_count == limit) {
+		ret = PTHREAD_BARRIER_SERIAL_THREAD;
+		b->_b_seq++;
+		__wake(&b->_b_seq, -1, 0);
+	} else {
+		a_store(&b->_b_lock, 0);
+		if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+		__wait(&b->_b_seq, 0, seq, 0);
+	}
+
+	__vm_lock(+1);
+
+	if (a_fetch_add(&b->_b_count, -1)==1) {
+		b->_b_seq++;
+		__wake(&b->_b_seq, -1, 0);
+		a_store(&b->_b_lock, 0);
+		if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+	} else {
+		__wait(&b->_b_seq, 0, seq+1, 0);
+	}
+	
+	__vm_unlock();
+
+	return 0;
+}
+
 struct instance
 {
 	int count;
@@ -16,9 +73,12 @@ int pthread_barrier_wait(pthread_barrier_t *b)
 	/* Trivial case: count was set at 1 */
 	if (!limit) return PTHREAD_BARRIER_SERIAL_THREAD;
 
+	/* Process-shared barriers require a separate, inefficient wait */
+	if (limit < 0) return pshared_barrier_wait(b);
+
 	/* Otherwise we need a lock on the barrier object */
 	while (a_swap(&b->_b_lock, 1))
-		__wait(&b->_b_lock, &b->_b_waiters, 1, 0);
+		__wait(&b->_b_lock, &b->_b_waiters, 1, 1);
 	inst = b->_b_inst;
 
 	/* First thread to enter the barrier becomes the "instance owner" */
@@ -27,7 +87,7 @@ int pthread_barrier_wait(pthread_barrier_t *b)
 		int spins = 10000;
 		b->_b_inst = inst = &new_inst;
 		a_store(&b->_b_lock, 0);
-		if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+		if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
 		while (spins-- && !inst->finished)
 			a_spin();
 		a_inc(&inst->finished);
@@ -40,19 +100,19 @@ int pthread_barrier_wait(pthread_barrier_t *b)
 	if (++inst->count == limit) {
 		b->_b_inst = 0;
 		a_store(&b->_b_lock, 0);
-		if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
+		if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
 		a_store(&inst->last, 1);
 		if (inst->waiters)
-			__wake(&inst->last, -1, 0);
+			__wake(&inst->last, -1, 1);
 	} else {
 		a_store(&b->_b_lock, 0);
-		if (b->_b_waiters) __wake(&b->_b_lock, 1, 0);
-		__wait(&inst->last, &inst->waiters, 0, 0);
+		if (b->_b_waiters) __wake(&b->_b_lock, 1, 1);
+		__wait(&inst->last, &inst->waiters, 0, 1);
 	}
 
 	/* Last thread to exit the barrier wakes the instance owner */
 	if (a_fetch_add(&inst->count,-1)==1 && a_fetch_add(&inst->finished,1))
-		__wake(&inst->finished, 1, 0);
+		__wake(&inst->finished, 1, 1);
 
 	return 0;
 }

src/thread/pthread_barrierattr_setpshared.c

@@ -2,6 +2,6 @@
 
 int pthread_barrierattr_setpshared(pthread_barrierattr_t *a, int pshared)
 {
-	*a = !!pshared;
+	*a = pshared ? INT_MIN : 0;
 	return 0;
 }