Summary: PostgreSQL crashes with SIGSEGV when a cursor is open over a composite type and the type is modified via ALTER TYPE during the same transaction, followed by a second FETCH.
Reproduction steps (minimal): CREATE TYPE foo AS (a INT, b INT); BEGIN; DECLARE c CURSOR FOR SELECT (i, power(2, 30))::foo FROM generate_series(1,10) i; FETCH c; ALTER TYPE foo ALTER ATTRIBUTE b TYPE TEXT; FETCH c; COMMIT;
Expected: Error message (type modified during active cursor) Actual: Server process terminated with signal 11 (Segmentation fault)
Server log: client backend (PID 85) was terminated by signal 11: Segmentation fault Failed process was running: [above SQL]
I confirmed the crash on master and traced the root cause. EEOP_ROW was the only rowtype-aware expression step that cached its TupleDesc at init time without an ExprEvalRowtypeCache guard. When ALTER TYPE changes an attribute's storage properties (e.g. int to text), the stale descriptor leads to SIGSEGV.
I looked through nearby rowtype-producing paths and found a few more cases with the same general shape, so I split the proposed fix into two patches:
0001: Detect row type changes in EEOP_ROW expressions
EEOP_ROW was caching the target composite type's TupleDesc at executor startup and never re-checking it. The patch stores the TypeCacheEntry pointer and tupDesc_identifier for named composite row results, then compares the identifier in ExecEvalRow() before forming the composite Datum. If the type changed, the cursor now reports:
ERROR: row type <type-name> has changed
instead of producing a malformed Datum that can later crash the backend.
0002: Detect row type changes in additional composite result paths
While checking for similar stale TupleDesc reuse, I found the same kind of guard is also needed for:
- whole-row Vars returning a named composite type - SQL-language functions returning a whole named composite result - targetlist SRFs returning named composite results
The second patch adds the same typcache TupleDesc identity check in those paths and adds regression coverage for cursor/ALTER TYPE/FETCH cases that previously crashed or could reuse stale layout information.
With these patches, the original repro and the additional repro cases fail cleanly with "row type ... has changed" errors, and the backend remains running.
