-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
Lustre 2.15.6
-
None
-
3
-
9223372036854775807
When a discovery-on client reboots and remounts with fewer networks and with peer discovery disabled, the server's LNet peer table keeps the client's old secondary NIDs — the client sends no push, and the server already has the peer flagged NIDS-UPTODATE so discovery never re-pings it. Server-originated messages such as blocking ASTs then get sent to a NID that no longer exists. The entry stays wrong until the server is restarted.
The stale NID gets selected rather than passed over because lnet_select_peer_ni round-robins across the peer's NIDs once they are equally healthy, and a NID stays at full health until a send over it fails. We see the blocking AST go to the dead NIDs and the client evicted on ldlm_bl_timeout. That escalation is specific to AWS EFA LND, whose recovery path is weaker than tcp/o2iblnd; with those, health-based resend should turn this into added latency rather than an eviction. The stale entry itself is not LND-specific — nothing ages a secondary NID out of a peer.
Proposed fix: on an initial connect, have the server force a re-discovery of the client so discovery re-pings it and the peer entry gets corrected.