-
Bug
-
Resolution: Unresolved
-
Major
-
None
-
Lustre 2.14.0, Lustre 2.18.0
-
None
-
3
-
9223372036854775807
If the server has to send to a client it has no connection to, it dials the client. When that SYN is silently dropped (e.g. a firewall that only allows client->server port 988), the connect hangs for the TCP SYN retry period (127 s by default) with ksnr_connecting set. During that time ksocknal_create_conn() rejects every connect from the lower-NID client as "connection race resolution". Without BULK_IN the client can't send ZC-ACKs (nonblk_zcack), so the server closes everything on "stale ZC_REQs ... resid: 0", and the cycle repeats until the client is evicted.
Reproducer: sanity-lnet test_239 (in the patch). It uses two NIs on one node, drops SYNs from the higher NID to the lower one, and has the lower NID ping the higher one while the higher NID's connect is stuck.
Fix: a real race settles within min_reconnectms. Make ksocknal_connecting() report a race only while our own connect is younger than that, and accept the peer's connection after that.
Lab result: 580 race rejects and an eviction without the fix; 0 rejects and all connections back in under 35 s with it.
- is related to
-
LU-21219 ksocklnd: positive EALREADY after retry cap crashes node
-
- Open
-