Re: [PoC] Federated Authn/z with OAUTHBEARER

On Fri, Feb 28, 2025 at 4:37 PM Thomas Munro <thomas(dot)munro(at)gmail(dot)com> wrote:
> In case it's relevant, it was green for me, but I also ran it in
> combination with my 3x-go-faster patch on that other thread. . o O {
> Timing/race stuff? Normally the build farm shakes that stuff out a
> bit more reliably than CI, but I doubt libcurl is set up on many
> animals... }

That does help, thanks. Luckily, I can still sometimes reproduce with
that patch, which should speed things up nicely.

Commenting out the failing test causes the next test to fail with
basically the same error, so there's something stateful going on.
There are some suspicious messages that occasionally show up right
before the failure:

# [libcurl] * IPv6: ::1
# [libcurl] * IPv4: 127.0.0.1
# [libcurl] * Trying [::1]:65269...
# [libcurl] * getsockname() failed with errno 22: Invalid argument
# [libcurl] * connect to ::1 port 65269 from ::1 port 65270
failed: Connection refused
# [libcurl] * Trying 127.0.0.1:65269...
# [libcurl] * Connected to localhost (127.0.0.1) port 65269

Later, Curl reconnects via IPv6 -- this time succeeding -- but then
the response gets mangled in some way. I assume headers are being
truncated, based on Curl's complaint about "HTTP/0.9".

The NetBSD man pages say that EINVAL is returned when the socket is
already shut down, suggesting some sort of bad interaction between
Curl and the test authorization server (and/or the OS?). I wonder if
my test server doesn't handle dual-stack setups correctly. I'll see if
I can get ktruss working on either side.

Thanks,
--Jacob