While reviewing futex.c re Esben's rt_mutex_next_owner() serialization
posting, I had a question regarding the use of
futex_atomic_cmpxchg_inatomic()
In all but two places in futex.c we wrap that call with
(inc|dec)_preempt_count() calls. Nothing stood out to me that would
account for the difference. Can some comment on why we don't have to
wrap the call in those two places (see patch for locations). Just in
case this happens to be a bug, I've included a patch to fix it.
Index: 2.6.17-rt8/kernel/futex.c
===================================================================
--- 2.6.17-rt8.orig/kernel/futex.c 2006-08-02 20:29:08.000000000 -0700
+++ 2.6.17-rt8/kernel/futex.c 2006-08-02 20:29:42.000000000 -0700
@@ -1303,8 +1303,10 @@
ret = get_user(uval, uaddr);
while (!ret) {
newval = (uval & FUTEX_OWNER_DIED) | newtid;
+ inc_preempt_count();
curval = futex_atomic_cmpxchg_inatomic(uaddr,
uval, newval);
+ dec_preempt_count();
if (curval == -EFAULT)
ret = -EFAULT;
if (curval == uval)
@@ -1684,8 +1686,11 @@
* thread-death.) The rest of the cleanup is done in
* userspace.
*/
+ inc_preempt_count();
nval = futex_atomic_cmpxchg_inatomic(uaddr, uval,
uval | FUTEX_OWNER_DIED);
+ dec_preempt_count();
+
if (nval == -EFAULT)
return -1;
--
Thanks,
Darren Hart
IBM Linux Technology Center
Realtime Linux Team