Re: [RFC][PATCH v2 5/5] mutex: Give spinners a chance to spin_on_owner if need_resched() triggered while queued

[Jason Low <jason.low2@hp.com>]

On Fri, 2014-01-31 at 21:08 +0100, Peter Zijlstra wrote:
> On Fri, Jan 31, 2014 at 12:01:37PM -0800, Jason Low wrote:
> > Currently still getting soft lockups with the updated version.
> 
> Bugger.. ok clearly I need to think harder still. I'm fairly sure this
> cancelation can work though, just seems tricky to get right :-)

Ok, I believe I have found a race condition between m_spin_lock() and
m_spin_unlock().

In m_spin_unlock(), we do "next = ACCESS_ONCE(node->next)". Then, if
next is not NULL, we proceed to set next->locked to 1.

A thread in m_spin_lock() in the unqueue path could execute
"next = cmpxchg(&prev->next, node, NULL)" after the thread in
m_spin_unlock() accesses its node->next and finds that it is not NULL.
Then, the thread in m_spin_lock() could check !node->locked before
the thread in m_spin_unlock() sets next->locked to 1.

The following addition change was able to solve the initial lockups that were
occurring when running fserver on a 2 socket box.

---
diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
index 9eb4dbe..e71a84a 100644
--- a/kernel/locking/mutex.c
+++ b/kernel/locking/mutex.c
@@ -513,8 +513,13 @@ static void m_spin_unlock(struct m_spinlock **lock)
 			return;
 
 		next = ACCESS_ONCE(node->next);
-		if (unlikely(next))
-			break;
+
+		if (unlikely(next)) {
+			next = cmpxchg(&node->next, next, NULL);
+
+			if (next)
+				break;
+		}
 
 		arch_mutex_cpu_relax();
 	}