Horton

Re: [PoC] Federated Authn/z with OAUTHBEARER

Jacob Champion <jacob.champion@enterprisedb.com>

View thread
From: Jacob Champion <jacob.champion@enterprisedb.com>
To: PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Cc: Daniel Gustafsson <daniel@yesql.se>, Peter Eisentraut <peter@eisentraut.org>, Antonin Houska <ah@cybertec.at>, Kashif Zeeshan <kashi.zeeshan@gmail.com>
Date: 2025-01-14T01:00:00Z
Lists: pgsql-hackers

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. meson: Fix install-quiet after clean

  2. oauth: Run Autoconf tests with correct compiler flags

  3. Link libpq with libdl if the platform needs that.

  4. Doc: correct spelling of meson switch.

  5. oauth: Correct SSL dependency for libpq-oauth.a

  6. oauth: Fix Autoconf build on macOS

  7. oauth: Move the builtin flow into a separate module

  8. Remove a stray "pgrminclude" annotation

  9. oauth: Simplify copy of PGoauthBearerRequest

  10. oauth: Improve validator docs on interruptibility

  11. oauth: Disallow synchronous DNS in libcurl

  12. oauth: Fix postcondition for set_timer on macOS

  13. oauth: Use IPv4-only issuer in oauth_validator tests

  14. Work around OAuth/EVFILT_TIMER quirk on NetBSD.

  15. oauth: Fix incorrect const markers in struct

  16. Add missing entry to oauth_validator test .gitignore

  17. cirrus: Temporarily fix libcurl link error

  18. Add support for OAUTHBEARER SASL mechanism

  19. libpq: Handle asynchronous actions during SASL

  20. require_auth: prepare for multiple SASL mechanisms

  21. Move PG_MAX_AUTH_TOKEN_LENGTH to libpq/auth.h

  22. Make SASL max message length configurable

  23. jsonapi: fully initialize dummy lexer

  24. common/jsonapi: support libpq as a client

  25. Remove fe_memutils from libpgcommon_shlib

  26. Revert ECPG's use of pnstrdup()

  27. Explicitly require password for SCRAM exchange

  28. Refactor SASL exchange to return tri-state status

On Mon, Jan 13, 2025 at 3:21 PM Jacob Champion
<jacob.champion@enterprisedb.com> wrote:
> Next email will discuss the architectural bug that Kashif found.

Okay, here goes. A standard OAuth connection attempt looks like this
(oh, I hope Gmail doesn't mangle it):

    Issuer    User    libpq   Backend
      |        |        |
      |        x -----> x -----> o      [1] Startup Packet
      |        |        |        |
      |        |        x <----- x      [2] OAUTHBEARER Request
      |        |        |        |
      |        |        x -----> x      [3] Parameter Discovery
      |        |        |        |
      |        |        x <----- o      [4] Parameters Stored
      |        |        |
      |        |        |
      |        |        |
      |        |        x -----> o      [5] New Startup Packet
      |        |        |        |
      |        |        x <----- x      [6] OAUTHBEARER Request
      |        |        |        |
      x <----- x <----> x        |
      x <----- x <----> x        |      [7] OAuth Handshake
      x <----- x <----> x        |
      |        |        |        |
      o        |        x -----> x      [8] Send Token
               |        |        |
               | <----- x <----- x      [9] Connection Established
               |        |        |
               x <----> x <----> x
               x <----> x <----> x      [10] Use the DB
               .        .        .
               .        .        .
               .        .        .

When the server first asks for a token via OAUTHBEARER (step 2), the
client doesn't necessarily know what the server's requirements are for
a given user. It uses the rest of the doomed OAUTHBEARER exchange to
store the issuer and scope information in the PGconn (step 3-4), then
disconnects and sets need_new_connection in PQconnectPoll() so that a
second connection is immediately opened (step 5). When the OAUTHBEARER
mechanism takes control the second time, it has everything it needs to
conduct the login flow with the issuer (step 7). It then sends the
obtained token to establish a connection (steps 8 onward).

The problem is that step 7 is consuming the authentication_timeout for
the backend. I'm very good at completing these flows quickly, but if
you can't complete the browser prompts in time, you will simply not be
able to log into the server. Which is harsh to say the least. (Imagine
the pain if the standard psql password prompt timed out.) DBAs can get
around it by increasing the timeout, obviously, but that doesn't feel
very good as a solution.

Last week I looked into a fix where libpq would simply try again with
the stored token if the backend hangs up on it during the handshake,
but I think that will end up making the UX worse. The token validation
on the server side isn't going to be instantaneous, so if the client
is able to complete the token exchange in 59 seconds and send it to
the backend, there's an excellent chance that the connection is still
going to be torn down in a way that's indistinguishable from a crash.
We don't want the two sides to fight for time.

So I think what I'm going to need to do is modify v41-0003 to allow
the mechanism to politely hang up the connection while the flow is in
progress. This further decouples the lifetimes of the mechanism and
the async auth -- the async state now has to live outside of the SASL
exchange -- but I think it's probably more architecturally sound. Yell
at me if that sounds unmaintainable or if there's a more obvious fix
I'm missing.

Huge thanks to Kashif for pointing this out!

--Jacob