[LU-1715] flock deadlock detection does not work properly

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.5.0
• Labels: patch
• Severity: 3
• Assignee: Keith Mannthey (Inactive)
• Reporter: Andriy Skulysh

Description

flock deadlocks are checked on the first attempt to grant the flock only. if we try again to grant it after its blocking lock is cancelled, we do not check for deadlocks which also may exist.

the reproducer is attached, its output is the following:
4730: sleeping 1
4731: taking lock0 [0, 100]
4732: taking lock1 [100, 200]
4731: done
4731 sleeping 1
4732: done
4732 sleeping 2
4730: taking lock2 [200, 300]
4730: done
4730: taking lock0 [0, 100]
4731: taking lock3 [100, 300]
4732: putting lock1 [100, 200]
4732: done
4732 Exit
--> it is waiting forever here.

Issue Links

• is related to: LU-3783 MDS crash, LustreError: 34483:0:(ldlm_flock.c:208:ldlm_flock_deadlock()) ASSERTION( req != lock ) failed