Re: Check the fd number against FD_SETSIZE in PQsocketPoll

From: Abraham Pérez <abraham1798(at)gmail(dot)com>
To: pgsql-hackers(at)lists(dot)postgresql(dot)org
Subject: Re: Check the fd number against FD_SETSIZE in PQsocketPoll
Date: 2026-11-21 00:31:29
Message-ID: CAF6dGh3m1OQOf6JgUtqeUsn4V=n=DMzRop-omPTP11zixApcCA@mail.gmail.com
Views: Whole Thread | Raw Message | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Hi Roman,

Thanks for the patch. I took a look at v2.

What I checked
--------------
- Applies cleanly on current master (git am, no fuzz/offset) at
207b5293b8d.
- Builds with --enable-debug --enable-cassert --enable-tap-tests; no new
compiler warnings.
- "make check" in src/test/regress is green (All 243 tests passed). As
expected this exercises nothing of the change (see below), but confirms
the tree is otherwise unaffected.

The concern: as written, the new code is never compiled
--------------------------------------------------------
The guard is added inside the "#else /* !HAVE_POLL */" arm of
PQsocketPoll() -- i.e. the select(2) fallback -- and is itself wrapped in
"#ifndef WIN32":

#ifndef WIN32
if (sock >= FD_SETSIZE)
{
SOCK_ERRNO_SET(EINVAL);
return -1;
}
#endif

But src/include/port.h unconditionally does, for every non-Windows build:

#ifndef WIN32
...
#define HAVE_POLL 1
...
#endif

(that line is a plain #define, not guarded by #ifndef HAVE_POLL). So:

* On any non-Windows platform HAVE_POLL is always defined, hence the
"#else /* !HAVE_POLL */" select() arm -- where the new check lives --
is never compiled. I confirmed this by preprocessing fe-misc.c: the
FD_SETSIZE check does not appear in the output, and even forcing
-UHAVE_POLL on the command line does not help, because port.h redefines
HAVE_POLL back to 1.

* On Windows the select() arm *is* compiled (HAVE_POLL is not defined
there), but the new block is excluded by the surrounding "#ifndef
WIN32" -- which is also the behavior you intended per your v2 note that
the check isn't needed on Windows.

So the net effect of v2 on every supported platform is a no-op: there is
no build configuration in which the new "sock >= FD_SETSIZE" test is
reached. The block itself is fine in isolation (I compiled it standalone
with -Werror; FD_SETSIZE, SOCK_ERRNO_SET and EINVAL are all in scope, and
the shape matches the existing precedent in fe-secure-gssapi.c), but it
can't protect anything as placed.

Because PostgreSQL assumes poll(2) on all Unix (that is exactly what the
port.h block encodes), the select() fallback in PQsocketPoll() is
effectively Windows-only today, and Windows is the one platform you
established does not need the guard. That seems to leave the original
scenario (a Unix without poll(2) handing out an fd >= FD_SETSIZE) without
a live code path to hit in the first place -- which may be why this reads
as "mostly theoretical".

Questions / suggestions
-----------------------
- Is there a real, currently-supported platform where PQsocketPoll() takes
the select() path *and* lacks poll(2)? If not, it's worth asking on-list
whether the hardening is still wanted, or whether the right fix is
elsewhere (e.g. the pgbench/client-fd thread you referenced in [0]).
- If it is still wanted, the check needs to sit where it actually compiles
and runs. As-is it will not.

Given the above, my tentative verdict is that v2 does not yet do what it
intends; I'd suggest clarifying the target platform before another spin.
Happy to re-test once that's settled.

Best regards,
Abraham

Browse pgsql-hackers by date

  From Date Subject
Next Message Prashant Chinnam 2026-11-21 00:48:48 Re: Fix exclusion of the init fork in table_block_relation_size()
Previous Message Chao Li 2026-11-21 00:27:37 Re: Fix resource leaks in pg_receivewal on LZ4 close failures