-
Bug
-
Resolution: Unresolved
-
Critical
-
Lustre 2.15.8
-
None
-
16-MDT DNE filesystem, ~535 TiB, Lustre 2.15, EL9 kernel (5.14.x). Triggered during failover/recovery of a single MDT.
-
2
-
9223372036854775807
On a DNE filesystem, an MDS can permanently deadlock its entire LDLM
blocking-AST (ldlm_bl) thread pool when a cross-MDT close is issued from lock
cancellation while the MDC import to the target MDT is not LUSTRE_IMP_FULL. All
lock cancellation on the affected MDS stops, the MDS begins evicting its own
internal client and then its regular clients, peer MDTs evict the affected MDS,
and the condition does not self-heal – only a reboot clears it.
Observed impact: on a 16-MDT production filesystem, two peer MDS entered this
state during failover of a third MDT. The filesystem was unusable for
approximately two hours and was recovered only by rebooting all MDS.