Add broker tests for a NIC split between shared and dedicated ports

[kthare10]

Context

A ConnectX-6 can now run one port in SR-IOV/shared mode and the sibling port dedicated. The dedicated half is advertised as a separate component with model ConnectX-6-1P.

1. No broker change was needed for mixed-mode cards

Verified by reading, now pinned down by tests:

• __check_component_labels_and_capacities matches on model string equality.
• __update_smart_nic_labels_and_capacities matches ports by local_name, so it is port-count agnostic.
• __exclude_allocated_component removes by component name, so the two halves of the card are already independent.
• get_matching_nodes_with_components counts by (Type, Model) and AggregatedBQMPlugin.occupied_node_capacity buckets by (resource_type, resource_model), so the new model gets its own capacity/allocated figures for free.

2. Bug fix: Shared NIC delegated capacity never decremented

Found while writing those tests, and fixed here.

__exclude_allocated_pci_device_from_shared_nic did:

delegated_capacity -= allocated.get_capacity_allocations()

Capacities defines __sub__ but not __isub__, so this builds a new object and rebinds the local — the delegation keeps its original unit count, while the BDF list beside it is mutated in place.

Consequence: a Shared NIC with every BDF allocated was never excluded from the candidate node, because the capacity check still reported the original units. The next allocation then reached usable_bdfs[0] on an empty list and raised IndexError instead of a clean INSUFFICIENT_RESOURCES. The comment at the capacity check — "Delegated capacity would have been decremented already to exclude allocated shared NICs" — shows the decrement was meant to persist.

The remaining BDFs are the source of truth, so the fix derives the unit count from them and mutates the delegated object in place, exactly the way the BDF list is already updated. Also guards a missing or empty label delegation, which previously raised AttributeError.

This bug predates the mixed-mode work and is independent of it.

Tests

New fabric_cf/actor/test/unit/test_mixed_mode_nic.py, 9 cases, no infrastructure required.

Mixed-mode:

• the dedicated half allocates its single port, MAC and local_name
• a two-port ConnectX-6 request does not match the single-port ConnectX-6-1P card
• allocating either half leaves the other untouched
• exhausting every VF of PF0 removes only the shared half; PF1 keeps its BDF and unit

Capacity tracking:

• delegated capacity follows the remaining BDFs at every step
• the last BDF reports exhausted, earlier ones do not
• an unknown BDF leaves capacity alone
• an exhausted card raises INSUFFICIENT_RESOURCES rather than IndexError

4 of these fail against the old code and pass against this branch. The full fabric_cf/actor/test/unit/ directory is green (39 passed).

Version

Bumped to 2.0.1b1 for dev deployment. 2.0.1 is not yet on PyPI, so 2.0.1b1 sorts ahead of the released 2.0.0 and behind the eventual 2.0.1. Local image tags in the three docker-compose.yml files follow the version.

⚠️ Two commits to revert before merge

• DEV ONLY: pin fabric-fim to 2.0.1b1
• DEV ONLY: pin fabric-am-handlers to 2.0.2b1 in Dockerfile-auth

Both exist so the branches can be deployed together on dev. main should not depend on or build against pre-releases — drop both before merging.

Depends on

fabric-testbed/InformationModel#223

🤖 Generated with Claude Code

https://claude.ai/code/session_016wB1CV7hRtwmWVLPY8GBxh