[LU-13461] LNet routing: wrong gw ni may be selected to reach undiscovered peer

Tue, 2 Feb 2021 18:17:43 +0000

Test scenario is as follows:

PeerA: two networks, one interface per network (A1@tcp, A2@tcp1)

PeerB: one network, two interfaces (B1@tcp2, B2@tcp2)

GW1: two interfaces on peerA's first net, two facing peerB (R1@tcp, R2@tcp, R3@tcp2, R4@tcp2)

GW2: two interfaces on peerA's second net, two facing peerB (R1@tcp1, R2@tcp1, R3@tcp2, R4@tcp2)

Routes on peer A: reach tcp2 via GW1, reach tcp2 via GW2

Routes on peer B: reach tcp via GW1, reach tcp1 via GW2

Do not run discovery on peer A from peer B or vice versa.

Once everything is configured, the following ping may fail from peer A to peer B:

lnetctl ping B1@tcp2

It looks like wrong gateway NI may be selected by peer A:

(lib-move.c:1921:lnet_handle_send()) TRACE: 192.168.122.103@tcp(192.168.122.103@tcp:<?>) -> 192.168.122.110@tcp2(192.168.122.110@tcp2:192.168.122.150@tcp1) <?> : GET try# 0

Whamcloud Community JIRA

[LU-13461] LNet routing: wrong gw ni may be selected to reach undiscovered peer