-
Bug
-
Resolution: Unresolved
-
Medium
-
None
-
None
-
None
-
3
-
9223372036854775807
Problem
sanity-lnet test_291 and test_292 fail on a node whose NIC reports a link speed of 4 Gb/s or more. Both tests expect socklnd to use conns_per_peer=1. socklnd sets conns_per_peer from the link speed, so the tests fail on fast links.
sanity-lnet test_291: @@@@@@ FAIL: failed to set conns-per-peer to 1 sanity-lnet test_292: @@@@@@ FAIL: expected 2 tcp connections on port 7988
Environment
- master at 7c3c9f8814 (LU-20076)
- rocky9 aarch64 VM, tcp NETTYPE, virtio-net NIC that reports 100000 Mb/s
Reproduction
- Use a node with a NIC that reports 100 Gb/s in ethtool.
- Run ONLY="291 292" bash sanity-lnet.sh.
Analysis
ksocklnd_speed2cpp() returns ilog2(speed / 1000) / 2 + 1. A 100 Gb/s link gives conns_per_peer=4. A NIC with no speed gives 1. We think that the CI VMs report no link speed. That would explain why CI does not see the failure.
Test 291 checks for conns_per_peer: 1 and fails at once. Test 292 expects 1 control connection and 1 bulk connection on the acceptor port. But socklnd makes conns_per_peer bulk connections. At 100 Gb/s, test 292 sees 5 connections and times out.
Proposed fix
Read conns_per_peer from lnetctl net show -v. Expect conns_per_peer bulk connections in test_291, and 1 + conns_per_peer connections in test_292.