-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
A push from a peer with discovery off can delete a peer that the user configured with lnetctl peer add. A ping reply with the same content does not delete that peer.
Problem
lnet_discovery_event_reply() keeps a peer that has LNET_PEER_CONFIGURED set. LU-14661 added this rule. The matching check in lnet_peer_push_event() is older than LU-14661, and it does not test LNET_PEER_CONFIGURED:
if (!(lp->lp_state & (LNET_PEER_NO_DISCOVERY |
LNET_PEER_LOCK_PRIMARY |
LNET_PEER_DISCOVERING)) &&
lp->lp_state & LNET_PEER_DISCOVERED) {
...
lp->lp_state |= LNET_PEER_MARK_DELETION;
}
The sequence:
- The user adds a peer with lnetctl peer add. Discovery on the local node then discovers the peer.
- The remote node turns discovery off. lnet_set_peer_discovery() calls lnet_push_update_to_peers(1), and the remote node pushes to its Multi-Rail peers.
- lnet_peer_push_event() on the local node sets LNET_PEER_MARK_DELETION.
- The discovery thread calls lnet_peer_deletion(), which calls lnet_peer_del_locked(). That function deletes the configured peer and all its peer NIs.
The local node thus loses the user configuration. Traffic later creates a new peer that has no LNET_PEER_CONFIGURED flag and none of the configured NIDs.
Proposed fix
Add LNET_PEER_CONFIGURED to the push check, so that both paths keep a configured peer. Add a test that configures a peer on a second node, discovers it, turns discovery off on that node, and checks that the peer stays.
Source
The Gerrit AI review found this case on https://review.whamcloud.com/70549 (LU-21145), which changes the same check. The finding is from the source only. Nobody has reproduced it at runtime yet.
- is related to
-
LU-21145 'network' target mount option does not work with LNet dynamic peer discovery
-
- Open
-