Distributed Erlang over TLS does not work correctly using OTP 26

[lukebakken]

Describe the bug

When using OTP 26 (26.0.2 to be precise) and -proto_dist inet_tls -ssl_dist_optfile ... two Erlang nodes can't connect via distributed Erlang.

Using the latest version of OTP 25 works correctly.

To Reproduce

Reproduction steps that show SUCCESS

Note: clone my repository and change to the main branch to show success:

https://github.com/lukebakken/erlang-otp-7497#reproduction-steps

Reproduction steps that show FAILURE

Note: clone my repository and change to the otp-26 branch to show failure:

https://github.com/lukebakken/erlang-otp-7497/tree/otp-26#reproduction-steps

The main difference between the two branches is the Erlang version specified in the .tool-versions file.

Expected behavior

This code succeeds:

net_kernel:connect('a@HOSTNAME').

Affected versions

OTP 26.0.2 is the version with which I am testing that demonstrates failure. OTP 25.3.2.3 succeeds.

[u3s]

• are you using the same OTP version on both nodes?
• what are you configuring in ssl_dist_optfile?
• is there some error message, stack trace or some other symptoms of failure?

[lukebakken]

@u3s I provided comprehensive reproduction steps in this git repository, in two branches:

This succeeds: https://github.com/lukebakken/erlang-otp-7497#reproduction-steps

This fails: https://github.com/lukebakken/erlang-otp-7497/tree/otp-26#reproduction-steps

If you have time, I'd appreciate someone else confirming what I have found. This issue was originally discovered when Team RabbitMQ and Docker tried to upgrade to OTP 26 -

docker-library/rabbitmq#652

There is a test that uses TLS-enabled distributed Erlang that used to succeed with OTP 25 and then started to fail with 26. I traced it down to this issue.

are you using the same OTP version on both nodes?

Yes, they are the same. Please check out the reproduction steps.

what are you configuring in ssl_dist_optfile?

This is the template that is turned into the optfile. It is identical for the working version and the one that fails. The x509 certs are identical as well.

https://github.com/lukebakken/erlang-otp-7497/blob/main/inet-dist-tls.config.in

is there some error message, stack trace or some other symptoms of failure?

No, there is not. It would be great to have some sort of debug logging available for distributed Erlang but I am not aware of it.

[RaimoNiskanen]

Try net_kernel:verbose(1). in both nodes before connecting. 2 is also possible, but start with 1.

[lukebakken]

@RaimoNiskanen on it!

[lukebakken]

Here is the transcript of a successful run using OTP 25.3.2.3 and net_kernel:verbose(2):

Node a

lbakken@shostakovich ~/development/lukebakken/erlang-otp-7497 (main *=)
$ ./run-node.sh a
elixir          1.15.2-otp-26   /home/lbakken/.tool-versions
erlang          25.3.2.3        /home/lbakken/development/lukebakken/erlang-otp-7497/.tool-versions
python          3.11.4          /home/lbakken/.tool-versions
ruby            3.2.2           /home/lbakken/.tool-versions
Erlang/OTP 25 [erts-13.2.2.2] [source] [64-bit] [smp:8:8] [ds:8:8:10] [async-threads:1] [jit:ns]

Eshell V13.2.2.2  (abort with ^G)
(a@shostakovich)1> =INFO REPORT==== 14-Jul-2023::05:43:07.749556 ===
{net_kernel,{'EXIT',<0.109.0>,shutdown}}
=INFO REPORT==== 14-Jul-2023::05:43:07.752467 ===
{net_kernel,{net_kernel,1473,nodedown,b@shostakovich}}
=INFO REPORT==== 14-Jul-2023::05:43:07.752525 ===
{net_kernel,{'EXIT',<0.112.0>,shutdown}}

(a@shostakovich)1> init:stop().
ok

Node b

lbakken@shostakovich ~/development/lukebakken/erlang-otp-7497 (main *=)
$ ./run-node.sh b
elixir          1.15.2-otp-26   /home/lbakken/.tool-versions
erlang          25.3.2.3        /home/lbakken/development/lukebakken/erlang-otp-7497/.tool-versions
python          3.11.4          /home/lbakken/.tool-versions
ruby            3.2.2           /home/lbakken/.tool-versions
Erlang/OTP 25 [erts-13.2.2.2] [source] [64-bit] [smp:8:8] [ds:8:8:10] [async-threads:1] [jit:ns]

=INFO REPORT==== 14-Jul-2023::05:42:55.575656 ===
{net_kernel,{connect,normal,a@shostakovich}}
Eshell V13.2.2.2  (abort with ^G)
(b@shostakovich)1> ['a@shostakovich']

(b@shostakovich)1> init:stop().
ok

[lukebakken]

AHA @RaimoNiskanen using net_kernel:verbose(2) reveals the error:

=INFO REPORT==== 14-Jul-2023::07:55:51.621871 ===
{net_kernel,
    {conn_own_exit,<0.103.0>,
        {ssl_connect_failed,
            {192,168,1,5},
            36843,
            {error,{option,server_only,fail_if_no_peer_cert}}},
        a@shostakovich}}

I did notice this while working on the original issue (docker-library/official-images@7335ae3) but it didn't make sense because there was no useful error logged or returned, and that setting worked just fine with OTP 25.

Before I close this issue I'm going to re-check the OTP 26 READMEs and release announcements to see where this change was (or was not) announced.

[lukebakken]

@RaimoNiskanen the change wasn't announced in my opinion. I see a brief mention of "Improved error checking and handling of ssl options" here -

https://www.erlang.org/news/164#ssl

...and one mention of fail_if_no_peer_cert here -

https://erlang.org/download/otp_src_26.0.readme

However, isn't fail_if_no_peer_cert a valid client option? Or was it seen as redundant when verify is taken into account?

[IngelaAndin]

No it is not a valid client option, has never been. Is documented as server option but was ignored by client previously!
https://www.erlang.org/doc/man/ssl.html

[lukebakken]

OK! I'll make sure all of the RabbitMQ docs reflect this. Thanks!