pgsql: libpq: Poison the v2 part of a v1 Bearer request - Mailing list pgsql-committers

From Jacob Champion
Subject pgsql: libpq: Poison the v2 part of a v1 Bearer request
Date
Msg-id E1w7e9X-002Lf3-2K@gemulon.postgresql.org
Whole thread Raw
List pgsql-committers
libpq: Poison the v2 part of a v1 Bearer request

The new PGoauthBearerRequestV2 API (which has similarities to the
"subclass" pointer architecture in use by the backend, for Nodes)
carries the risk of a developer ignoring the type of hook in use and
just casting directly to the V2 struct. This will appear to work fine in
19, but crash (or worse) when speaking to libpq 18.

However, we're in a unique position to catch this problem, because we
have tight control over the struct. Add poisoning code to the v1 path
which does the following:

- masks the v2 request->issuer pointer, to hopefully point at nonsense
  memory
- abort()s if the v2 request->error is assigned by the hook
- attempts to cover both with VALGRIND_MAKE_MEM_NOACCESS for the
  duration of the callback (a potential AddressSanitizer implementation
  is left for future work)

The struct is unpoisoned after the call, so we can switch back to the v2
internal implementation when necessary.

Reviewed-by: Zsolt Parragi <zsolt.parragi@percona.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/CAOYmi%2BnCg5upBVOo_UCSjMfO%3DYMkZXcSEsgaADKXqerr5wahZQ%40mail.gmail.com

Branch
------
master

Details
-------
https://git.postgresql.org/pg/commitdiff/0af4d402cb900364f275cc6f9c28dca4a5bec36b

Modified Files
--------------
src/interfaces/libpq/fe-auth-oauth.c | 142 ++++++++++++++++++++++++++++++++---
src/interfaces/libpq/fe-auth-oauth.h |   1 +
2 files changed, 134 insertions(+), 9 deletions(-)


pgsql-committers by date:

Previous
From: Nathan Bossart
Date:
Subject: pgsql: Avoid including vacuum.h in tableam.h and heapam.h.
Next
From: Melanie Plageman
Date:
Subject: pgsql: Fix test_aio read_buffers() to work without cassert