Repository navigation
UCT/GDAKI: Restore "Check for HCA reachability like rc_mlx5 (#11781)" - #11945
Conversation
|
🤖 Starting review — findings will be posted here when done. |
| .ep_invalidate = (uct_ep_invalidate_func_t)ucs_empty_function_return_unsupported, | ||
| .ep_connect_to_ep_v2 = uct_rc_gdaki_ep_connect_to_ep_v2, | ||
| .iface_is_reachable_v2 = (uct_iface_is_reachable_v2_func_t)ucs_empty_function_return_one_int, | ||
| .iface_is_reachable_v2 = uct_ib_iface_is_reachable_v2, |
There was a problem hiding this comment.
gda's iface address only carries the RC address-type constant, so a peer with a different NUM_CHANNELS is still reported reachable and uct_rc_gdaki_ep_connect_to_ep_v2 then reads iface->num_channels QPNs out of a shorter remote ep address. can we compare num_channels here too, the way rc_mlx5 compares the TM address type before delegating?
| @@ -98,3 +98,4 @@ UCS_TEST_P(test_uct_iface, is_reachable) | |||
| } | |||
There was a problem hiding this comment.
The new rc_gda case only exercises self-address reachability plus random-address rejection; it does not cover a real cross-node gda pair or a RoCE-vs-IB link-layer mismatch. That matches what the existing IB transports get from this fixture, so no extra test is warranted.
| @@ -98,3 +98,4 @@ UCS_TEST_P(test_uct_iface, is_reachable) | |||
| } | |||
There was a problem hiding this comment.
rc_gda/test_uct_iface.is_reachable silently registers zero tests on machines without GDA support (no active CUDA primary context, no peermem/DMABUF), so the effective validation depends on the GDA CI job in contrib/test_jenkins.sh actually running gtest.
This reverts commit dcf19c1.