Workaround for rq->lock deadlock

The following patch converts double_lock_balance to a full DP alogorithm to work around a deadlock in the scheduler when running on an 8-way SMP system.

I think the original algorithm in this function is technically correct. So really this patch is plastering over another lurking issue. However, it does fix the observed deadlock on our systems here so I thought I would at least share the discovery.

The actual problem is probably related to a code path which takes task_rq_locks without using the balancer code. It might also be a race between an rq_lock and something else. TBD

---

 kernel/sched.c |   19 ++++++++++++-------
 1 files changed, 12 insertions(+), 7 deletions(-)

index 6f2cf6a..e946e3f 100644
@@ -2507,14 +2507,19 @@ static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
 		BUG_ON(1);
 	}
 	if (unlikely(!spin_trylock(&busiest->lock))) {
-		if (busiest < this_rq) {
-			spin_unlock(&this_rq->lock);
-			spin_lock(&busiest->lock);
-			spin_lock(&this_rq->lock);
+		struct rq *rq_l =  busiest < this_rq ? busiest : this_rq;
+		struct rq *rq_h =  busiest > this_rq ? busiest : this_rq;
 
-			return 1;
-		} else
-			spin_lock(&busiest->lock);
+		spin_unlock(&this_rq->lock);
+
+		while (1) {
+			if (spin_trylock(&rq_l->lock)) {
+				if (spin_trylock(&rq_h->lock))
+					return 1;
+				else
+					spin_unlock(&rq_l->lock);
+			}
+		}
 	}
 	return 0;
 }

To LKML: This is an RT specific patch. I forgot to change the subject line. Sorry for the confusion.

-Greg