[Pulp-dev] Moving Content Guard Authorization to Webserver and out of pulp-content
bmbouter at redhat.com
Mon Mar 16 19:45:32 UTC 2020
tl;dr I'm going back to the original plan to have Content Guard
authorization in Pulp's processes and not the webserver. Also now we know
a) to use urlencoded certificates instead of stripping newlines and b) how
to make it work on both apache and nginx having tested it already.
After lots of testing and digging, I believe the best thing to do is to go
back to the original idea of having nginx/apache forward a certificate used
in a TLS connection with the client to pulp-container for authorization.
Here's the summary of reasons why:
* Trying to use a WSGIAccessScript to check in Python in the webserver
isn't an option when you're reverse proxying with either Apache or Nginx.
For example, to even do it in Nginx you would have to write the script in
Lua. Yikes. So moving the checking to apache/nginx isn't even possible in a
* Having the additional requirement to install all of Pulp on nginx or
apache is undesirable at best. It's a requirement likely not easily
achieved in deployments like k8s for example meaning maybe Pulp couldn't
run there if we did this.
* Forward the certificate is not as bad as I originally thought. Yes you
are forwarding the clients cert but you aren't forwarding or re-presenting
the client's key so you can't impersonate the client like a MITM would.
Also you are forwarding it to a trusted service so that is ok.
* stripping newlines from rhsm certs causes them to be unreadable by rhsm
library so we need to use urlencoding instead of newline strippping to make
rhsm certs transport safe for headers
* urlencoding a tls cert for forwarding is so common nginx even has a
pre-defined variable for exactly this. For apache I have a simple
mod_rewrite rule handling it. Both were working for me.
Feedback is welcome! Please send any thoughts, questions, or concerns you
have. Next steps wise I will close the Epic I opened
https://pulp.plan.io/issues/6323 and it's subtasks as WONTFIX. I will
continue the pulp-certguard work on https://pulp.plan.io/issues/4664
On Thu, Mar 12, 2020 at 9:44 AM Brian Bouterse <bmbouter at redhat.com> wrote:
> On Wed, Mar 11, 2020 at 9:18 PM Justin Sherrill <jsherril at redhat.com>
>> Thanks and it does appears that we use certs that could be too large for
>> default header sizes (by several multiples).
>> Could you elaborate a bit about the design a bit more? I'm curious what
>> the requirements of web service layer (will it need to talk to the pulp3
>> api? the db?) Will it just add some header after reading the cert (and
>> validating the path) and then pass it on to the reverse proxy with apache?
> This was the idea, but in looking through the nginx and apache
> documentation it doesn't seem possible. I had planned to use
> WSGIAccessScript, but the pulp-content app isn't wsgi it's ProxyPass so I
> don't see an access script option for that which means my plan isn't
> achievable as is. At best nginx has the auth request model
> https://nginx.org/en/docs/http/ngx_http_auth_request_module.html but that
> makes a subrequest and apache doesn't support it. We'd still have the
> "certs are too big for headers" problem to deal with. I'm configuring nginx
> and http and doing some testing. Articles are saying the header options
> would work for both nginx and httpd, so I'm going to do some testing with
> their configs.
> I'll post my findings and bump this list also for more input.
>> On 3/11/20 2:58 PM, Brian Bouterse wrote:
>> On Wed, Mar 11, 2020 at 2:34 PM Justin Sherrill <jsherril at redhat.com>
>>> We had discussed base64 encoding the cert in the webserver on the way in
>>> and then letting cert guard decode it. While that's not ideal I think it
>>> has some advantages over moving the full auth into the webserver. What was
>>> your motivation for going with that approach over the base64 encoding
>> Thank you for this question! I ended up with a few different concerns
>> about the base64 encode-and-forward idea. Architecturally the concern with
>> it is that it's frowned upon to forward the client's TLS cert beyond the
>> TLS termination point because that is what MITM software does. Also, there
>> are some practical concerns: one, I don't think nginx can provide a similar
>> runtime base64 encoding feature. Also I was concerned with header length
>> truncation and what happens when the certificates get longer.
>> Overall having auth that is based on TLS certificates brought me to the
>> conclusion that we need to auth where the TLS is terminated. What do you
>> More thoughts and questions are welcome. This is a good discussion.
>> On 3/11/20 2:11 PM, Brian Bouterse wrote:
>>> tl;dr: What we have today cannot work with rhsm certificates which
>>> Katello uses. To resolve, we need to have content guard checking moved to
>>> the webserver configs for apache and nginx and not done in pulp-content as
>>> it is today. https://pulp.plan.io/issues/6323
>>> We need to bring the auth to where TLS is terminated because we can't
>>> being the client certs to pulp-content due to invalid header characters. As
>>> is, pulp-certguard cannot work with Katello's cert types (rhsm certs) so
>>> that is driving my changes.
>>> If anyone has major concerns or other ideas please let me know. In the
>>> meantime I'm proceeding moving the authorization to the webserver and then
>>> updating pulp-certguard to work with that. This will make pulp-certguard's
>>> GA tied to pulpcore 3.3.0. Feedback is welcome.
>>> : https://pulp.plan.io/issues/6323
>>> Pulp-dev mailing listPulp-dev at redhat.comhttps://www.redhat.com/mailman/listinfo/pulp-dev
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Pulp-dev