improve pshared barriers

eliminate the sequence number field and instead use the counter as the
futex because of the way the lock is held, sequence numbers are
completely useless, and this frees up a field in the barrier structure
to be used as a waiter count for the count futex, which lets us avoid
some syscalls in the best case.

as of now, self-synchronized destruction and unmapping should be fully
safe. before any thread can return from the barrier, all threads in
the barrier have obtained the vm lock, and each holds a shared lock on
the barrier. the barrier memory is not inspected after the shared lock
count reaches 0, nor after the vm lock is released.

diff --git a/src/internal/pthread_impl.h b/src/internal/pthread_impl.h
index 82d62426f618a23f5f628bdaf665f574feae2f4a..dd3863fcc58cbf2244048fdf778caba0f3bd91a1 100644 (file)
--- a/src/internal/pthread_impl.h
+++ b/src/internal/pthread_impl.h
@@ -77,7 +77,7 @@ struct __timer {
 #define _b_waiters __u.__i[1]
 #define _b_limit __u.__i[2]
 #define _b_count __u.__i[3]
-#define _b_seq __u.__i[4]
+#define _b_waiters2 __u.__i[4]
 #define _b_inst __u.__p[4]
 
 #include "pthread_arch.h"

diff --git a/src/thread/pthread_barrier_wait.c b/src/thread/pthread_barrier_wait.c
index 9b1edbf94ad19a46a93fc9d68e497db662148ff4..71f7b5feebac79867b2fbbcd4cd1feac178194f7 100644 (file)
--- a/src/thread/pthread_barrier_wait.c
+++ b/src/thread/pthread_barrier_wait.c
@@ -21,7 +21,6 @@ void __vm_unlock(void)
 static int pshared_barrier_wait(pthread_barrier_t *b)
 {
        int limit = (b->_b_limit & INT_MAX) + 1;
-       int seq;
        int ret = 0;
        int v, w;
 
@@ -30,33 +29,36 @@ static int pshared_barrier_wait(pthread_barrier_t *b)
        while ((v=a_cas(&b->_b_lock, 0, limit)))
                __wait(&b->_b_lock, &b->_b_waiters, v, 0);
 
-       seq = b->_b_seq;
-
+       /* Wait for <limit> threads to get to the barrier */
        if (++b->_b_count == limit) {
+               a_store(&b->_b_count, 0);
                ret = PTHREAD_BARRIER_SERIAL_THREAD;
-               b->_b_seq++;
-               __wake(&b->_b_seq, -1, 0);
+               if (b->_b_waiters2) __wake(&b->_b_count, -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);
+               while ((v=b->_b_count)>0)
+                       __wait(&b->_b_count, &b->_b_waiters2, v, 0);
        }
 
        __vm_lock(+1);
 
-       if (a_fetch_add(&b->_b_count, -1)==1) {
-               b->_b_seq++;
-               __wake(&b->_b_seq, -1, 0);
+       /* Ensure all threads have a vm lock before proceeding */
+       if (a_fetch_add(&b->_b_count, -1)==1-limit) {
+               a_store(&b->_b_count, 0);
+               if (b->_b_waiters2) __wake(&b->_b_count, -1, 0);
        } else {
-               __wait(&b->_b_seq, 0, seq+1, 0);
+               while ((v=b->_b_count))
+                       __wait(&b->_b_count, &b->_b_waiters2, v, 0);
        }
        
        /* Perform a recursive unlock suitable for self-sync'd destruction */
        do {
                v = b->_b_lock;
                w = b->_b_waiters;
-       } while (a_cas(&b->_b_lock, v, v-1 & INT_MAX) != v);
+       } while (a_cas(&b->_b_lock, v, v==INT_MIN+1 ? 0 : v-1) != v);
 
+       /* Wake a thread waiting to reuse or destroy the barrier */
        if (v==INT_MIN+1 || (v==1 && w))
                __wake(&b->_b_lock, 1, 0);