-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
Summary:
When a multi-rail peer's secondary NID is renumbered while
its primary NID remains reachable, LNetAddPeer() builds a temporary
peer for the new NID and relies on discovery to merge it into the
existing peer. In lnet_peer_add_nid() the merge collides with a safety
check when the existing peer has LNET_PEER_LOCK_PRIMARY set
with more than one NI. Both peers are marked LNET_PEER_BAD_CONFIG and
the client hangs on all I/O to that server.
Scenario:
Server peer state on the client transitions:
[TCP_NID, OLD_NID] --> [TCP_NID, NEW_NID]
--> means the server rebooted directly. No lnetctl peer del and no
client-side unmount happens in between.
On reconnection the client learns the new NID list and
calls LNetAddPeer([TCP_NID, NEW_NID]). TCP_NID is the primary and remains
reachable; OLD_NID was a secondary on the same peer and has been replaced
by NEW_NID.
LNetAddPeer() builds a fresh peer lp with lp_primary_nid = NEW_NID
and lp_merge_primary_nid = TCP_NID, and kicks discovery on NEW_NID.
Discovery's ping reply lists [TCP_NID, NEW_NID]. lnet_peer_merge_data(lp, pbuf)
computes addnis = [TCP_NID] and calls lnet_peer_add_nid(lp, TCP_NID, ...).
This will fail because TCP_NID already has it's peer [TCP_NID, OLD_NID].
Would like to know if this is a valid concern, and if so if the proposed fix makes sense.