Re: [PATCH] - Re: libpq: decouple the .pgpass lookup port from the connection port

Hello,

It has been a couple of weeks with no comments, so let me add a small,
self-contained reproducer that makes both the problem and the fix easy
to see. I am also cc'ing a few people who have
worked on the relevant code, in case it is of interest: Tom (the
host/hostaddr split for the .pgpass lookup that this mirrors, and the
2018 .pgpass fix I referenced), Michael (the recent passwordFromFile()
work), and Heikki (the multi-host port parsing that the patch reuses).
Apologies for the noise if this is not the right moment.

The reproducer below uses a real SSH tunnel to itself, so the port that
libpq connects to genuinely differs from the real server port. Only the
patch changes the outcome; everything else (server, database, .pgpass)
is identical. It needs a libpq built with the v1 patch posted upstream
in this thread.

daf(at)t:postgres$ DEMO=/home/daf/scripts/postgres/passfileport-demo
daf(at)t:postgres$
daf(at)t:postgres$ BIN_CON="$DEMO/con-el-patch/usr/local/bin"
daf(at)t:postgres$
daf(at)t:postgres$ LIB_CON="$DEMO/con-el-patch/usr/local/lib/x86_64-linux-gnu"
daf(at)t:postgres$
daf(at)t:postgres$ PSQL_SIN="$DEMO/sin-patch/usr/local/bin/psql"
daf(at)t:postgres$
daf(at)t:postgres$ LIB_SIN="$DEMO/sin-patch/usr/local/lib/x86_64-linux-gnu"
daf(at)t:postgres$
daf(at)t:postgres$ PSQL_CON="$DEMO/con-el-patch/usr/local/bin/psql"
daf(at)t:postgres$
daf(at)t:postgres$ unset PGPASSWORD PGPASSFILEPORT
daf(at)t:postgres$
daf(at)t:postgres$ export LD_LIBRARY_PATH="$LIB_CON"
daf(at)t:postgres$
daf(at)t:postgres$ export PGDATA=/tmp/pgd-demo
daf(at)t:postgres$
daf(at)t:postgres$ export PGPASSFILE=/tmp/pgpass-demo
daf(at)t:postgres$
daf(at)t:postgres$ rm -rf "$PGDATA"
daf(at)t:postgres$
daf(at)t:postgres$ "$BIN_CON/initdb" -D "$PGDATA" -U postgres -A trust
--no-sync
The files belonging to this database system will be owned by user "daf".
This user must also own the server process.

The database cluster will be initialized with locale "C.UTF-8".
The default database encoding has accordingly been set to "UTF8".
The default text search configuration will be set to "english".

Data page checksums are enabled.

creating directory /tmp/pgd-demo ... ok
creating subdirectories ... ok
selecting dynamic shared memory implementation ... posix
selecting default "max_connections" ... 100
selecting default "shared_buffers" ... 128MB
selecting default time zone ... America/Argentina/Buenos_Aires
creating configuration files ... ok
running bootstrap script ... ok
performing post-bootstrap initialization ... ok

Sync to disk skipped.
The data directory might become corrupt if the operating system crashes.

Success. You can now start the database server using:

/home/daf/scripts/postgres/passfileport-demo/con-el-patch/usr/local/bin/pg_ctl
-D /tmp/pgd-demo -l logfile start

daf(at)t:postgres$ printf 'local all all trust\nhost all all 127.0.0.1/32
scram-sha-256\n' > "$PGDATA/pg_hba.conf"
daf(at)t:postgres$
daf(at)t:postgres$ "$BIN_CON/pg_ctl" -D "$PGDATA" -o "-p 5440 -c
listen_addresses=127.0.0.1" -w start
waiting for server to start....2026-07-08 13:39:19.186 -03 [666970]
LOG:  starting PostgreSQL 19beta1 on x86_64-linux, compiled by
gcc-12.2.0, 64-bit
2026-07-08 13:39:19.186 -03 [666970] LOG:  listening on IPv4 address
"127.0.0.1", port 5440
2026-07-08 13:39:19.189 -03 [666970] LOG:  listening on Unix socket
"/tmp/.s.PGSQL.5440"
2026-07-08 13:39:19.196 -03 [666975] LOG:  database system was shut down
at 2026-07-08 13:38:56 -03
2026-07-08 13:39:19.200 -03 [666970] LOG:  database system is ready to
accept connections
 done
server started
daf(at)t:postgres$
daf(at)t:postgres$ "$BIN_CON/psql" -X -p 5440 -U postgres -h /tmp -d
postgres -c "create role tuser login password 'sekret';"
CREATE ROLE
daf(at)t:postgres$
daf(at)t:postgres$ echo "127.0.0.1:5441:postgres:tuser:sekret" > "$PGPASSFILE"
daf(at)t:postgres$
daf(at)t:postgres$ chmod 600 "$PGPASSFILE"
daf(at)t:postgres$
daf(at)t:postgres$ LD_LIBRARY_PATH="$LIB_SIN" "$PSQL_SIN" -X -w
"host=127.0.0.1 port=5440 user=tuser dbname=postgres" -tAc "select
'AUTENTICO-OK'"
psql: error: connection to server at "127.0.0.1", port 5440 failed:
fe_sendauth: no password supplied
daf(at)t:postgres$
daf(at)t:postgres$ LD_LIBRARY_PATH="$LIB_SIN" "$PSQL_SIN" -X -w
"host=127.0.0.1 port=5440 user=tuser dbname=postgres passfileport=5441"
-tAc "select 'AUTENTICO-OK'"
psql: error: invalid connection option "passfileport"
daf(at)t:postgres$
daf(at)t:postgres$ LD_LIBRARY_PATH="$LIB_CON" "$PSQL_CON" -X -w
"host=127.0.0.1 port=5440 user=tuser dbname=postgres passfileport=5441"
-tAc "select 'AUTENTICO-OK'"
AUTENTICO-OK
daf(at)t:postgres$
daf(at)t:postgres$ cat $PGPASSFILE
127.0.0.1:5441:postgres:tuser:sekret
daf(at)t:postgres$
daf(at)t:postgres$ "$BIN_CON/pg_ctl" -D "$PGDATA" -m immediate stop
waiting for server to shut down....2026-07-08 13:44:30.677 -03 [666970]
LOG:  received immediate shutdown request
2026-07-08 13:44:30.688 -03 [666970] LOG:  database system is shut down
 done
server stopped
daf(at)t:postgres$

The connection port stays 5441; only the password-file lookup key uses
5440. Without passfileport the default is unchanged, so existing setups
behave exactly as before.

As a sanity check that the difference is solely this patch, I built two
libpq trees from the same commit, identical except for the change, and
ran the same command line against both through the same tunnel: with the
unpatched libpq passfileport is simply "invalid connection option", and
with the patched one the connection succeeds. Same server, same .pgpass,
same command; the patch is the only variable.

The design is unchanged from the original proposal, and the v1 patch is
the one already posted in this thread (adds the passfileport conninfo
option and PGPASSFILEPORT env var, splits per host like port, uses it as
the .pgpass lookup key, with docs and a TAP test). I would still welcome
opinions on two points in particular:

- the parameter name (passfileport is only tentative), and
- the security angle: this matches the real server port explicitly,
which is stricter than the port wildcard people use today, so I
believe it improves on the "password sent to the wrong server" risk
rather than adding to it.

Thanks for taking a look.

Regards,
Diego

On 2026-06-29 14:14, Diego wrote:
>
> Hi,
>
> Following up on my proposal from last week. Since the idea didn't draw
> objections, I went ahead and wrote the patch (code, docs and a TAP test)
> so the discussion can be more concrete. v1 attached.
>
> It compiles cleanly, the new TAP test 008_passfileport passes (10
> subtests), and the full authentication suite stays green. I also
> verified it end-to-end against a real server reached through an SSH
> tunnel: the .pgpass entry matches with passfileport set and fails
> without it.
>
> I'd really value your read on whether the
> host/hostaddr-style decoupling is the right shape here, and on the
> parameter name (still tentative).
>
> I'm not a professional C developer and I need help.
> Ofc, I used Cursor and Claude to review the code before this mail. ;P
>
> Thanks,
> Diego
>
> On 2026-06-23 12:57, Diego wrote:
>>
>> Hello hackers,
>>
>> I would like to float an idea before writing a patch, to find out whether
>> it is wanted and to get the design right.
>>
>> Problem
>> -------
>>
>> libpq looks up a password in .pgpass using the connection's host and port
>> as part of the key (host:port:database:user:password). When a client
>> connects through an SSH tunnel, or through a connection pooler that
>> listens on a different local port, the port that libpq actually connects
>> to is not the port of the real server. As a result, the .pgpass lookup is
>> done against the local/tunnel port and fails to match entries written for
>> the real server port.
>>
>> Concretely, suppose the real server is db.example.com:5432 and a user
>> opens an SSH tunnel so that 127.0.0.1:54321 forwards to it. The natural
>> .pgpass entry is:
>>
>>     db.example.com:5432:appdb:alice:secret
>>
>> The client then connects with host=db.example.com (kept for .pgpass and
>> TLS), hostaddr=127.0.0.1 and port=54321 (the tunnel). libpq looks up
>>
>>     db.example.com:54321:appdb:alice
>>
>> which does not match the 5432 entry, so no password is found and the user
>> is prompted (or the connection fails under -w).
>>
>> The host side of this exact problem was already solved
>> -------------------------------------------------------
>>
>> libpq already decouples the *host* used for the .pgpass lookup from the
>> real network endpoint: hostaddr gives the address actually connected to,
>> while host remains the logical name used for the .pgpass lookup and for
>> TLS verification. This is the pwhost logic in fe-connect.c, which goes
>> back to the 2018 thread "Bizarre behavior in libpq's searching of
>> ~/.pgpass":
>>
>> https://www.postgresql.org/message-id/30805.1532749137%40sss.pgh.pa.us
>>
>> The port has no equivalent. passwordFromFile() is called with
>> conn->connhost[i].port, i.e. the real connection port, with no way to say
>> "connect to this port, but look up .pgpass under that port". The host has
>> host/hostaddr; the port only has port. This proposal is to close that
>> asymmetry.
>>
>> Why the port wildcard is not enough
>> -----------------------------------
>>
>> One can write the entry with a wildcard port:
>>
>>     db.example.com:*:appdb:alice:secret
>>
>> and it does match the tunnel. But the wildcard over-matches: a single
>> local forwarding port (say 54321, or even a fixed local port reused for
>> several tunnels at different times) ends up matching every server reached
>> through that port, so the same password line can be applied to different
>> servers. That is precisely the kind of "password sent to the wrong
>> server" situation the 2018 host fix was trying to avoid. The wildcard
>> trades safety for convenience; it is not a substitute for matching the
>> real server port.
>>
>> Proposal
>> --------
>>
>> Add a libpq connection parameter that specifies the port to be used for
>> the .pgpass lookup, independently of the port libpq connects to. The
>> connection still uses port (and hostaddr); only the password-file lookup
>> key uses the new value. When the new parameter is not set, behavior is
>> unchanged: the lookup uses port exactly as today.
>>
>> I do not have a strong opinion on the name and would rather not bikeshed
>> it before the idea itself is judged. Candidates that came to mind:
>>
>>     - pgpassport / passfileport (it only affects the password file)
>>     - portaddr (mirrors hostaddr: "port stays logical, portaddr is the
>>       real endpoint"), though that would invert today's meaning of port,
>>       which is probably too invasive
>>
>> A dedicated parameter that affects only the .pgpass lookup (the first
>> option) seems the least surprising and the smallest change. It is also
>> easy to reason about for security: it is an explicit, opt-in assertion by
>> the user, exactly like hostaddr/host.
>>
>> This is not hypothetical. I ran into it myself while adding SSH tunnel
>> support to pgcli (a widely used Postgres CLI): with the tunnel active, an
>> explicit-port .pgpass entry never matches, because the lookup happens
>> against the random local forwarding port. The user is prompted for a
>> password even though the matching entry is right there, and only a
>> wildcard port papers over it. Other tools hit the same wall:
>>
>>     - pgcli: SSH tunnel rewrites the port before the .pgpass lookup
>> https://github.com/dbcli/pgcli/pull/1546
>>     - DBeaver: .pgpass looked up by 127.0.0.1 through an SSH tunnel
>> https://github.com/dbeaver/dbeaver/issues/16499
>>     - pgAdmin 4: control the SSH tunnel local port for .pgpass matching
>> https://github.com/pgadmin-org/pgadmin4/issues/6903
>>
>> Questions for the list
>> ----------------------
>>
>>   1. Is decoupling the .pgpass-lookup port from the connection port
>>      something libpq wants, given that host/hostaddr already does the
>>      equivalent for the host?
>>   2. Is a dedicated lookup-only parameter the right shape, or would you
>>      prefer a different model?
>>   3. Naming preferences?
>>
>> If there is interest, I am happy to write the patch (code, docs and
>> tests).
>>
>> Thanks for reading,
>> Diego
>>