Hello Ioi,
On 22/10/21 12:29 pm, Ioi Lam wrote:
On 10/21/21 9:09 PM, Jaikiran Pai wrote:
Hello Ioi,
This is my initial analysis of the problem.
====>>> Can anyone think of similar problems that may happen elsewhere?
The static constructors of enum classes are executed at both CDS
dump time and run time. E.g.,
public enum Modifier {
OPEN
}
The <clinit> method essentially does this:
public static final Modifier OPEN = new Modifier("OPEN");
If a reference of Modifier.OPEN is stored inside the CDS archived
heap during dump time, it will be different than the value of
Modifier.OPEN that is re-created at runtime by the execution of
Modifier.<clinit>
I have almost next to nothing knowledge about CDS internals. My only
understanding of it is based on some documentation that I have read.
One of them being this one
https://docs.oracle.com/en/java/javase/17/vm/class-data-sharing.html#GUID-7EAA3411-8CF0-4D19-BD05-DF5E1780AA91.
Based on that documentation (and other similar ones), it was my
understanding that CDS was meant to store/share class "metadata" like
it states in that doc:
"When the JVM starts, the shared archive is memory-mapped to allow
sharing of read-only JVM metadata for these classes among multiple
JVM processes."
But from what you explain in that enum example, it looks like it also
stores class instance data that is computed at build time on the host
where the JDK image was generated? Did I understand it correctly? Is
this only for enums or does it also store the static initialization
data of "class" types too? If it does store the static init data of
class types too, then wouldn't such data be host/build time specific
and as such the classes that need to be enrolled into the default CDS
archive of the JDK should be very selective (by reviewing what they
do in their static init)? Like I said, I haven't looked into this in
detail so perhaps it already is selective in the JDK build?
Hi Jaikiran,
Thank you very much for the detailed response.
CDS also has the ability to archive Java heap object. Since
https://bugs.openjdk.java.net/browse/JDK-8244778 , we have archived
the entire module graph to improve start-up time. At run time, the
module graph (as well as other archived heap objects) are loaded from
the CDS archive and put into the Java heap (either through memory
mapping or copying).
That is interesting and something that I hadn't known.
You can see the related code in
jdk.internal.module.ModuleBootstrap::boot()
I just had a look at it and it's quite elaborate and it'll take a me
while to fully grasp it (if at all) given its understandable complexity.
When the module system has started up, the module graph will reference
a copy of the OPEN enum object that was created as part of the
archive. However, the Modifier.<clinit> will also be executed at VM
start-up, causing a second copy of the OPEN enum object to be stored
into the static field Modified::OPEN.
Thank you for that detail. That helps me understand this a bit more (and
opens a few questions). To be clear - the VM startup code which creates
that other copy, ends up creating that copy because that piece of
initialization happens even before the module system has fully started
up and created those references from the archived state? Otherwise, the
classloaders I believe would be smart enough to not run that static init
again, had the module system with that graph from the archived state
been fully "up"?
So would this mean that this not just impacts enums but essentially
every class referenced within the module system (of just boot layer?)
that has state which is initialized during static init? To be more
precise, consider the very common example of loggers which are typically
static initialized and stored in a static (final) field:
private static final java.util.logger.Logger logger =
Logger.getLogger(SomeClass.class);
If the boot layer module graph has any classes which has state like
this, it would then mean that if such classes do get initialized very
early on during VM startup, then they too are impacted and the module
graph holding instances of such classes will end up using a different
instance for such fields as compared to the rest of the application code?
In essence, such classes which get accessed early (before module system
with the archived graph is "up") during VM startup can end up
_virtually_ having their static initialization run twice (I understand
it won't be run twice, but that's the end result, isn't it)?
-Jaikiran
