[selekkala]

Pre-submission checklist

• git diff --check
• pre-commit run

Summary

Removing an FBOSS load balancer can leave its SAI hash object attached to the switch on ASICs that do not support SAI_HASH_FIELDS_CLEAR_BEFORE_SET. The subsequent SaiHash release then removes an object that SAI still considers referenced, which can fail with OBJECT_IN_USE and crash the SW agent during load-balancer removal.

This PR detaches ECMP and LAG hash objects unconditionally before releasing their owning references and adds a regression test for the unsupported-feature case.

Root Cause

SaiSwitchManager::removeLoadBalancer() used resetLoadBalancer() to clear the switch hash attributes. resetLoadBalancer() is intentionally gated by HwAsic::Feature::SAI_HASH_FIELDS_CLEAR_BEFORE_SET, because it was introduced as a workaround for replacing hash fields on selected SAI implementations.

On GR2, that feature is unsupported. The switch therefore continued to reference the old hash object while ecmpV[46]Hash_ or lagV[46]Hash_ was reset. SAI object lifetime ordering was violated: the hash object was released before the switch reference was cleared.

Fix

• Add an unconditional detachLoadBalancer<HashAttrT>() helper that writes SAI_NULL_OBJECT_ID to the switch hash attribute.
• Use it for ECMP and LAG IPv4/IPv6 hash attributes in removeLoadBalancer() before resetting the corresponding shared_ptr.
• Keep resetLoadBalancer() feature-gated for its original hash-field replacement workaround; this change does not alter add/update programming behavior.

Test Plan

• Added SaiHashDetachTest.removeLoadBalancerDetachesHashesBeforeRelease.
  • Uses a fake ASIC with SAI_HASH_FIELDS_CLEAR_BEFORE_SET disabled.
  • Programs and removes both ECMP and LAG hashes.
  • Verifies the switch attributes are cleared and the fake SAI hash objects are released.
• git diff --check passed.
• Existing g200 FBOSS image build passed in the associated buildimage workspace.
• Live R1 reproduction using the affected load-balancer configuration completed with both agents active, zero restarts, no coredump, and no crash/fatal/segmentation-fault messages after the SW-agent restart.

The focused Buck2 test could not be executed in this checkout because its generated external-cell configuration is not present; the test is included for mainline CI.