[YARN-1398] Deadlock in capacity scheduler leaf queue and parent queue for getQueueInfo and completedContainer call

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 2.2.0
• Fix Version/s: 2.4.0
• Component/s: resourcemanager
• Target Version/s: 2.4.0
• Hadoop Flags: Reviewed
• Assignee: Vinod Kumar Vavilapalli
• Reporter: Sunil G
• Created: 11/Nov/13 14:48
• Updated: 10/Apr/14 13:11
• Resolved: 21/Feb/14 02:22

Description

getQueueInfo in parentQueue will call child.getQueueInfo().
This will try acquire the leaf queue lock over parent queue lock.

Now at same time if a completedContainer call comes and acquired LeafQueue lock and it will wait for ParentQueue's completedConatiner call.

This lock usage is not in synchronous and can lead to deadlock.

With JCarder, this is showing as a potential deadlock scenario.

Issue Links

• duplicates: YARN-1422 RM CapacityScheduler can deadlock when getQueueUserAclInfo() is called and a container is completing
• duplicates: YARN-1737 Deadlock in org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity
• is broken by: YARN-569 CapacityScheduler: support for preemption (using a capacity monitor)