On Mon, May 31, 2021 at 10:44 AM Richard Opalka <ropalka@redhat.com> wrote:
Helo,

Comments inlined:

On Thu, May 27, 2021 at 9:44 PM Brian Stansberry <brian.stansberry@redhat.com> wrote:
For the last month we've been focusing quite a bit of energy toward seeing what it will take for WildFly to run well on the upcoming JDK 17. This post is one of two I plan. This one is an attempt to start a discussion around a couple topics; the other will be more of a status update. 

Status summary is things are progressing well, no show-stoppers so far, but plenty more to do. More on that in the other post....

WF 23 runs well on SE 13. We want to get to 17. The key barriers are:

1) SE 14 dropped the java.security.acl package.
2) SE 15 introduced hidden classes (JEP 371)
3) SE 16 strongly encapsulated JDK internals by default (JEP 396)
SE 17 will eliminate global --illegal-access command line option (JEP 403) so explicit enumeration
of --add-opens and --add-exports will be the only possibility to open JDK internal packages.

The discussion points relate to #3. WildFly does quite a lot reflection stuff, plus we have some use cases where end users may want to use internal JDK classes. SE 16 locks this down. For a good primer on the basic things SE allows us when we need things to be made available, see [1]. Richard Opalka did a lot of good analysis of what JPMS-related VM launch settings we need for WF to work properly; see [2] and [3]. It's not a huge set, which is nice.

But, it's not complete, because it doesn't account for user applications. If application code requires additional deep reflection, then additional VM launch settings will be needed. I think that's ok in general; we provide hooks for users to add things to the JAVA_OPTS flags that are passed to java. But the less users need to do that the better. Hence the discussion topics:
AFAIR there are two places where modular jdk params must be specified.
One place is shell scripts (common.sh(.bat), etc) and second place is org.jboss.as.host.controller.jvm.JvmType .
That is a minor problem. We should configure these params either in shell scripts or in config files but not in Java code.

The problematic part is also how to propagate (configure) modular jdk parameters?
1) Should we have just one global place where modular jdk params will be specified (e.g. common.sh(.bat)) and propagated ( e.g. domain -> host controller(s) -> server(s) )?
2) Or should users be able to specify modular jdk params for each entity like domain, server, host controller?
Maybe that is already supported via standard configuration?

The key requirement is that each HC (DC or otherwise) is independently configurable, and each server is as well. A domain mode server doesn't have to run on the same VM as its HC, so we cannot force HC-level settings (e.g. stuff from common.sh) onto the servers.

The various 'jvm' config settings in domain.xml/host.xml let users customize these things (or any other JVM launch setting) on a per-server basis. JvmType provides a kind of simple ease-of-use thing. Otherwise we'd have to have standard config blocks per JVM type in the standard config files we ship, and then users (and our own testsuite) would have to be taught to use them.

It is clunky that there is no way to turn off applying those defaults though.


1) The ClassReflectionIndex[4] constructor iterates over all fields and methods in a class and marks them as accessible. For any class that is used as an EE component type, *as well its superclasses*, a ClassReflectionIndex is created. This means if an application uses some JDK class as a superclass (ignore Object, which gets some special handling that makes it not a problem in this discussion), then that superclass's package is going to need to be opened. We have no way to know what superclasses our users' component might have, so we can't open them up in for them as part of our standard launch args.

My general understanding is we do this in order to allow things like injection of values into fields or wrapping calls to non-public methods with interceptors.

Is there anything we can do about this? Any intelligence we can apply to avoid doing unnecessary opening? (See [5] for a very specific example of such a thing.) 

Or is this maybe not a big problem? We already need to open the java.util package for other reasons, so EE component based on classes in that package won't have a problem.
There are many utility classes in various JDK packages, e.g.
 * Readable or Runnable in java.lang. 
 * Closeable or Flushable in java.io.
We do open all these packages already so maybe it will be sufficient for our users too without need to open other packages.

Thanks; that's a good point.


2) There are cases where our configuration allows users to specify a class to use as the impl of interface, as an instruction for the server to instantiate an instance and use it. Examples include NamingContext and java.security.Policy impls.  In some cases well known examples of those interfaces are internal JDK classes.

Should we identify likely cases of these things and proactively include those packages in our server launch --add-opens set? My general instinct is no, but there may be cases where my instincts are wrong.
I would also say no. Few of the answers will be obvious after a successful 100% TCK run.
_______________________________________________
wildfly-dev mailing list -- wildfly-dev@lists.jboss.org
To unsubscribe send an email to wildfly-dev-leave@lists.jboss.org
%(web_page_url)slistinfo%(cgiext)s/%(_internal_name)s


--
Brian Stansberry
Principal Architect, Red Hat JBoss EAP
He/Him/His