Re: BUG #19382: Server crash at __nss_database_lookup

Hi Andrey,

Thank you for the detailed review and test cases. I've addressed all three
points in v6:

Point 1 (domain at root): collect_composite_type_versions() now resolves
domain types at the root level, not just for nested attributes. This
ensures the type tree walk works correctly when the record variable itself
is a domain over composite.

Point 2 (missing pfree): Added pfree() calls for compTypeOids and
compTypeVersions before NULLing them in the record-clear path.

Point 3 (Cases 5 and 6 i.e dot assignment and SELECT INTO fields): These
bypassed assign_record_var() because they modify the ExpandedRecord in
place via expanded_record_set_field(). The fix was to restructure
check_record_type_not_altered() into a two-level check:
i) Outermost type: Always checked using erh->er_tupdesc_id, which is set
when the ExpandedRecord is created. This works for all code paths (whole
assignment, field assignment, SELECT INTO) without needing a snapshot.
ii) Nested types: Checked against the snapshot when available (taken at
assign_record_var() and instantiate_empty_record_variable()).

I also added a fast-path optimization in snapshot_record_composite_types()
to avoid repeated type tree walks when a record is assigned in a loop — it
skips the snapshot if the outermost type's tupDesc_identifier hasn't
changed since the last snapshot.

Added both new test cases (Cases 5 and 6) to the regression. All 248 core
regression tests and all 13 PL/pgSQL tests pass.

Regards,
Surya Poondla