From: SupperSu (@SupperSu) <noreply+SupperSu@github.com>
To: postgresql-interfaces/psqlodbc <noreply+postgresql-interfaces-psqlodbc@github.com>
Subject: [postgresql-interfaces/psqlodbc] issue #223: `CC_from_PGresult()` in `connection.c` can dereference a NULL `ConnectionClass *conn` and segfault.
Date: Fri, 23 Oct 2026 18:20:56 +0000
Message-ID: <gh-postgresql-interfaces-psqlodbc-223@github.com> (raw)
## Summary
`CC_from_PGresult()` in `connection.c` can dereference a NULL `ConnectionClass *conn` and segfault. On the cursor "fetch next batch" path, `QR_next_tuple()` calls into `CC_send_query_append()`, which invokes `CC_from_PGresult(res, stmt, NULL, cursor, &pgres)` with a hard-coded `NULL` connection. If `QR_from_PGresult()` then returns `FALSE`, the error branch runs `CC_get_errornumber(conn)` with `conn == NULL`, crashing the process with a read through a near-NULL address. The crash is in the error-handling path, so a recoverable fetch failure is turned into a hard process crash.
## Environment
- psqlODBC: observed with **12.01.0000**; the same code path is **still present on `main`** (checked HEAD, Oct 2026).
- Unicode driver (`psqlodbcw.so`), Linux x86_64, with unixODBC.
- `UseDeclareFetch` enabled (cursor-based fetch), reading from a PostgreSQL source via `SQLFetchScroll`/`SQLExtendedFetch`.
## Affected code (links to `main`)
- Crash — unguarded deref in the failure branch: https://github.com/postgresql-interfaces/psqlodbc/blob/main/connection.c#L1702-L1716 (the `if (0 >= CC_get_errornumber(conn))` at L1711, and `switch (QR_get_rstatus(res))` at L1713).
- The call site that passes `NULL` for `conn` (comment: *"called from QR_next_tuple and must return immediately"*): https://github.com/postgresql-interfaces/psqlodbc/blob/main/connection.c#L2258-L2261
- The cursor fetch-continuation that reaches it (`qi.result_in = self`): https://github.com/postgresql-interfaces/psqlodbc/blob/main/qresult.c#L1174-L1176
- `QR_from_PGresult()` deliberately tolerates a NULL `conn` (guards it before use, and its failure returns never touch `conn`): https://github.com/postgresql-interfaces/psqlodbc/blob/main/qresult.c#L598 (also the `if (stmt && conn)` and `if (NULL != conn)` guards further down). The caller `CC_from_PGresult` is missing the same guard.
## Root cause
`QR_from_PGresult()` treats `conn` as optional (it is called with `NULL` on the `QR_next_tuple` continuation and guards every use). But `CC_from_PGresult()`'s failure branch does not apply the same guard: it calls `CC_get_errornumber(conn)` (and `QR_get_rstatus(res)` / `QR_get_message(res)`) unconditionally. When `QR_from_PGresult()` fails on the NULL-`conn` path, `CC_get_errornumber(NULL)` dereferences NULL.
The faulting access is a read at the small fixed offset of the `__error_number` field within `ConnectionClass` (i.e. `NULL + offsetof(...)`), consistent with `CC_get_errornumber(conn)` on `conn == NULL`.
## Captured backtrace (production core dump, symbolized — driver frames)
Symbolized backtrace from a real crash with this signature. Frames above the ODBC entry point are application-side and omitted:
```
#0 CC_from_PGresult psqlodbcw.so <- SIGSEGV: CC_get_errornumber(conn) with conn == NULL
#1 CC_send_query_append psqlodbcw.so <- CC_from_PGresult(res, stmt, NULL, cursor, &pgres)
#2 QR_next_tuple psqlodbcw.so <- "FETCH n IN cursor", reuse result (qi.result_in = self)
#3 SC_fetch psqlodbcw.so
#4 PGAPI_ExtendedFetch psqlodbcw.so
#5 SQLFetchScroll psqlodbcw.so
#6 SQLFetchScroll SQLFetchScroll.c (unixODBC driver manager)
... application fetch loop calling SQLFetchScroll (omitted) ...
```
The crash signature is a `SIGSEGV` with a read fault at the low fixed address noted above, and it is deterministic — every occurrence lands on the same instruction in `CC_from_PGresult`. This backtrace matches the source analysis exactly: crash at `#0 CC_from_PGresult`, reached from the `QR_next_tuple` cursor fetch-continuation (`#2 → #1`).
## How to reproduce
With `UseDeclareFetch` enabled, run a `SELECT` that opens a cursor and fetch rows so the driver issues repeated `FETCH … IN "cursor"` via `QR_next_tuple()`. If `QR_from_PGresult()` returns `FALSE` on one of the continuation fetches (e.g. the backend returns an error/FATAL for that `FETCH`, the connection is interrupted, or result construction otherwise fails), the driver enters the `CC_from_PGresult` failure branch with `conn == NULL` and crashes.
I observed this deterministically in production under concurrent bulk reads of a table with large text columns — once the failed-fetch condition occurs, it crashes at the same instruction every time. I have not reduced it to a standalone minimal reproducer, but the `NULL` originates at a fixed call site in the driver itself (connection.c L2261), so the path is unambiguous from the source.
I'm happy to open a PR with the fix if that's welcome.