[wildfly-dev] HTTP/2 out of the box in Wildfly 10.1

Jason T. Greene jason.greene at redhat.com
Wed Jun 8 08:28:43 EDT 2016



> On Jun 7, 2016, at 8:51 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
> 
> I have created a PR for this here: https://github.com/wildfly/wildfly-core/pull/1596 (it will also require some upstream changes).
> 
> Basically this just creates a new schema version, and add the 'generate-self-signed-certificate-host' attribute to the keystore.
> 
> I have not added a script to enable HTTPS over management as Jason suggested, I am not 100% sure if that really belongs in core or as part of the full distribution.
> 

It probably belongs in core so that other layered projects/products can get it as well. However for 10.1 I think it's fine if we stick it in full for now.



> Stuart
> 
>> On Wed, Jun 8, 2016 at 6:55 AM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>> 
>> 
>>> On Wed, Jun 8, 2016 at 6:51 AM, Jason Greene <jason.greene at redhat.com> wrote:
>>> So after reviewing this thread and discussing with a few folks, I’d like to propose, for 10.1:
>>> 
>>> #1b - Same as the previous #1, we don’t enable TLS for management by default for now, but we additionally include an extra cli script to enable TLS.
>> 
>> We would leave the cert generation bit in the security realm, but just don't enable the HTTPS interface. That way all that is required is for the user to add the https="managements-https" attribute.
>> 
>> Stuart
>>  
>>> 
>>> For 11 I think we should move to TLS by default, perhaps with a configurable URL policy on redirects, and address the incongruence with upgrade over app.
>>> 
>>> I think its likely reasonable to redirect by default for 11, but we can hash that out further. One nice thing I had forgotten about is that the JDK will prompt for you to add unknown certs, and this all works with the CLI[1]. So it’s really only non-interactive clients we have to worry about, and that sounds like a reasonable burden for upgrade.
>>> 
>>> [1]
>>> 
>>> [disconnected /] connect
>>> Unable to connect due to unrecognised server certificate
>>> Subject    - CN=foo,OU=foo,L=Madison,ST=WI,C=US
>>> Issuer     - CN=myServer, OU=test, L=Madison, ST=WI, C=US
>>> Valid From - Tue Jun 07 15:22:06 CDT 2016
>>> Valid To   - Thu Jun 07 15:22:06 CDT 2018
>>> MD5 : cd:68:be:0b:e0:c0:1c:63:d5:2a:85:c8:d1:9d:e7:7d
>>> SHA1 : ae:f8:35:fd:09:c9:b3:08:05:59:a6:40:5e:ac:6e:e8:ce:85:72:4b
>>> 
>>> Accept certificate? [N]o, [T]emporarily, [P]ermenantly : t
>>> 
>>> 
>>> 
>>>> On Jun 7, 2016, at 6:24 AM, Jason T. Greene <jason.greene at redhat.com> wrote:
>>>> 
>>>> Long term I think we want management using TLS, but that can of course come in phases. Assuming 2) is one of those phases to come (either now or later), a following step is that the CLI, and really any remoting client, should prefer TLS with a defaulted trust store location that points to the keystore. 
>>>> 
>>>> With 2) if we have the default of the attribute that forces redirect be true, and our default config be false, then someone that carries over their old config would not have a potential security weakness. If they have a CLI script that adds the https port, it will fail, hopefully sending a signal to look. Although, the user might just assume that oh it's there, I don't have to do anything. 
>>>> 
>>>> Another interesting thing about 2 is that IIRC we have conflicting behavior between the app port which doesn't force upgrade and the management port which does.
>>>> 
>>>> So my preference is 2, because at some point we have to do it anyway, and if we have TLS out of the box might as well use it. 
>>>> 
>>>>> On Jun 6, 2016, at 10:48 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>>>>> 
>>>>> So while implementing this I have noticed a potential problem that it would be good to get some feedback on.
>>>>> 
>>>>> If the management interface has SSL by default then the HTTP interface will always redirect to the HTTPS interface. This effectively breaks the management API, as clients such as the CLI, Arquillian etc will be redirected to HTTPS, and then reject the self signed certificate (as they should).
>>>>> 
>>>>> I am not sure what to do about this, these are the options as I see them:
>>>>> 
>>>>> 1) Don't enable SSL for the management interface (just for the Undertow subsystem). The management interface can still use this auto-generation capability, it just won't be enable by default (we could even leave the cert in the security domain, but just not enable the https interface).
>>>>> 
>>>>> 2) Disable automatic redirects for HTTP upgrade requests (potentially controlled by an attribute). This will allow the CLI etc to work, but at the price of potentially reducing security, as some connections that would have previously been redirected to use HTTPS will no longer do this. 
>>>>> 
>>>>> 3) Enable it by default and leave it broken. We can setup some kind of automatic trust store thing so the local CLI works, and can get our test suite to work with Arquillian in a similar manner. Personally I think this is a terrible idea, but I am including it for completeness.
>>>>> 
>>>>> Personally I think we should go for 1). Given that this is supposed to be about developer usability I don't think having management also use SSL as being that important.
>>>>> 
>>>>> Stuart
>>>>> 
>>>>>> On Mon, Jun 6, 2016 at 10:24 PM, Jason T. Greene <jason.greene at redhat.com> wrote:
>>>>>> Awesome! Another idea I had on how we could get away with it being in server boot, is to have a pre-boot first time setup task, either launched from the shell/batch scripts or as a special pre-step before the AS module loads. We could then report boot time as the time AFTER first time installation tasks have completed, which I think is fair because the server hasn't yet been started.
>>>>>> 
>>>>>>> On Jun 5, 2016, at 11:53 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>>>>>>> 
>>>>>>> I have some initial work on this at: https://github.com/stuartwdouglas/wildfly-core/tree/WFCORE-1576
>>>>>>> 
>>>>>>> If you go to https://localhost:9993 it will generate the certificate (although all that will be served is a 404 page as the console is not installed).
>>>>>>> 
>>>>>>> Stuart
>>>>>>> 
>>>>>>>> On Mon, Jun 6, 2016 at 12:46 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>>>>>>>> I think that would actually end up being more complex.
>>>>>>>> 
>>>>>>>> Stuart
>>>>>>>> 
>>>>>>>>> On Mon, Jun 6, 2016 at 12:45 PM, Jason T. Greene <jason.greene at redhat.com> wrote:
>>>>>>>>> Another option could be a post boot task. So it's still eager but don't block completed start. We'd still need to block Tls ports though. So maybe this does not help
>>>>>>>>> 
>>>>>>>>>> On Jun 5, 2016, at 9:31 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>>>>>>>>>> 
>>>>>>>>>> 2048 bits adds close to a second to first boot on my machine (obviously subsequent boots are unaffected). 
>>>>>>>>>> 
>>>>>>>>>> This is probably a bit much, I will work on getting a POC for the lazy loading approach implemented.
>>>>>>>>>> 
>>>>>>>>>> Stuart
>>>>>>>>>> 
>>>>>>>>>>> On Mon, Jun 6, 2016 at 12:27 PM, Jason T. Greene <jason.greene at redhat.com> wrote:
>>>>>>>>>>> We should really be generating 2048 bit keys. 
>>>>>>>>>>> 
>>>>>>>>>>> I don't like adding to our boot time, we have already seen it grow and this would be yet another case.
>>>>>>>>>>> 
>>>>>>>>>>>> On Jun 5, 2016, at 8:57 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>>>>>>>>>>>> 
>>>>>>>>>>>> So I just did up a very quick prototype that generates self signed certificates on startup and it looks like the difference in startup time is negligible (at least when generating 1024 bit RSA keys). Even if the difference is measurable it only affects the very first startup.
>>>>>>>>>>>> 
>>>>>>>>>>>> I think that in order to simplify the implementation of this it may be better to simply generate the key of first startup, instead of attempting to do it lazily. 
>>>>>>>>>>>> 
>>>>>>>>>>>> Stuart
>>>>>>>>>>>> 
>>>>>>>>>>>>> On Sat, Jun 4, 2016 at 12:09 AM, Jason T. Greene <jason.greene at redhat.com> wrote:
>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> What will be default keysize? It has to be probably choosen to work also without "Java Cryptography Extension (JCE) Unlimited Strength Jurisdiction Policy"
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> Probably the largest that is supported without JCE. It does not matter that much, self signed certs are inherently insecure, this is a developer usability feature, not something that can be used in production.
>>>>>>>>>>>>> 
>>>>>>>>>>>>> IIRC there is actually no limit on RSA key size, it's only symmetric algs that are limited, so we could use a standard 2048 bit key without issue.
>>>>>>>>>>>>> 
>>>>>>>>>>>>> 
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> Stuart
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> On Thu, Jun 2, 2016 at 10:01 PM, Stuart Douglas <stuart.w.douglas at gmail.com> wrote:
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> So I guess we should talk about how this should actually work. 
>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> In terms of auto generating the key I was thinking we would need to add a new attribute to the 'keystore' element under the security realm, something like 'auto-generate-cert-host="localhost"'. I am not sure what other options we would need, or how configurable we should make it, but as this is for testing/development purposes I don't think we need to expose full control over the certificate generation process.
>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> In terms of the implementation we could just implement an SSLContext wrapper, that can do the generation and then create a 'real' SSLContext the first time it is asked to create and SSLEngine.
>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> Stuart
>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>>> On Fri, Jun 3, 2016 at 3:19 AM, Jason Greene <jason.greene at redhat.com> wrote:
>>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>>> > On Jun 2, 2016, at 11:29 AM, Harold Campbell <hcamp at muerte.net> wrote:
>>>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>>>> > On Thu, 2016-06-02 at 09:22 +1000, Stuart Douglas wrote:
>>>>>>>>>>>>>>>>> >> Hi All,
>>>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>>>> >> I would like to propose that we add support for HTTP/2 out of the box
>>>>>>>>>>>>>>>>> >> in Wildfly 10.1.
>>>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>>>> > This lowly user desperately wants a release containing the fix to WFLY-
>>>>>>>>>>>>>>>>> > 6283 sooner rather than later. I'm sure other people have other pet
>>>>>>>>>>>>>>>>> > bugs awaiting release.
>>>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>>>> > I have no opinion on HTTP/2 being added other than to ask that pent up
>>>>>>>>>>>>>>>>> > bug fixes be kept in mind.
>>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>>> Hi Harold,
>>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>>> That fix is already in master, so it will be included in 10.1.
>>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>>> --
>>>>>>>>>>>>>>>>> Jason T. Greene
>>>>>>>>>>>>>>>>> WildFly Lead / JBoss EAP Platform Architect
>>>>>>>>>>>>>>>>> JBoss, a division of Red Hat
>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>>>>> wildfly-dev mailing list
>>>>>>>>>>>>>>>> wildfly-dev at lists.jboss.org
>>>>>>>>>>>>>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>>> wildfly-dev mailing list
>>>>>>>>>>>>>> wildfly-dev at lists.jboss.org
>>>>>>>>>>>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>>> _______________________________________________
>>>> wildfly-dev mailing list
>>>> wildfly-dev at lists.jboss.org
>>>> https://lists.jboss.org/mailman/listinfo/wildfly-dev
>>> 
>>> --
>>> Jason T. Greene
>>> WildFly Lead / JBoss EAP Platform Architect
>>> JBoss, a division of Red Hat
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/wildfly-dev/attachments/20160608/702560d1/attachment-0001.html 


More information about the wildfly-dev mailing list