pingpong, etc. programs

[Saurabh Desai <sdesai austin ibm com>]

I found out why the pingpong and other programs like test_str03 (an NGPT
test 
program), etc. performed unstable and poor on a 4-way system using NPTL. 
The pthread_cond_broadcast wakes up 200+ threads and they all contend
for 
an internal condvar lock. And again they wait in kernel using futex. 
To verify this I tried adding this patch in the library and that made a
hugh difference.

Without this patch:
./pp -v -n 128 -i 1 -S 32768
256 threads initialised in 0s.10121ms
128 games completed in 53s.104041ms

With this patch:
./pp -v -n 128 -i 1 -S 32768
256 threads initialised in 0s.9894ms
128 games completed in 0s.5049ms

Again, I did this to confirm the problem area. I guess the right fix
also involves futexes related kernel changes.

- Saurabh

==================
--- pthread_cond_broadcast.c.orig       2003-04-30 16:56:22.764088584
-0500
+++ pthread_cond_broadcast.c    2003-04-30 16:56:32.206653096 -0500
@@ -68,7 +68,7 @@
 #endif

       /* Wake everybody.  */
-      lll_futex_wake (futex, INT_MAX);
+      lll_futex_wake (futex, 1);

       /* That's all.  */
       return 0;

--- pthread_cond_wait.c.orig    2003-04-30 17:15:58.114792440 -0500
+++ pthread_cond_wait.c 2003-04-30 17:16:05.540663536 -0500
@@ -101,7 +101,7 @@
 {
   struct _pthread_cleanup_buffer buffer;
   struct _condvar_cleanup_buffer cbuffer;
-  int err;
+  int err, futex_wakeup=0;

   /* Make sure we are along.  */
   lll_mutex_lock (cond->__data.__lock);
@@ -167,8 +167,12 @@
   /* Another thread woken up.  */
   ++cond->__data.__woken_seq;

+  if (val > cond->__data.__woken_seq)
+       futex_wakeup = 1;
   /* We are done with the condvar.  */
   lll_mutex_unlock (cond->__data.__lock);
+  if (futex_wakeup)
+       lll_futex_wake (futex, 1);

   /* The cancellation handling is back to normal, remove the handler. 
*/
   __pthread_cleanup_pop (&buffer, 0);