On 25/10/2026 17:40, Baji Shaik wrote:
I noticed two places that can segfault if a schema is dropped while
they run: pg_identify_object(), and schema_to_xml() with the related
SQL/XML functions. Both look up a schema name by OID without holding a
lock on the schema, and dereference the NULL that get_namespace_name()
returns after a concurrent DROP SCHEMA. It's a narrow race and rare in
practice, but any user can hit it, and it takes the cluster through
crash recovery.
0001 makes pg_identify_object() return NULL for the schema, as it
already does for objects that have gone away, so catalog-wide scans
keep working. 0002 makes the single-schema XML functions report
that the schema does not exist, as they do for a nonexistent schema,
and makes database_to_xml* skip it. Both go back to 14.
Committed down to v14, thanks!
- Heikki
