Java 21 clinit deadlock

[David Holmes]

On 15/09/2023 5:00 pm, Simone Bordet wrote:
> David,
> 
> On Thu, Sep 14, 2023 at 4:33 AM David Holmes <david.holmes at oracle.com> wrote:
>> I've created a draft PR here:
>>
>> https://github.com/openjdk/jdk/pull/15732
>>
>> can you apply that patch, build and re-test? Hopefully that will show
>> which thread is doing the missing initialization.
> 
> We have attached the logs to the pull request above.
> 
> Seems like a classic deadlock:

Indeed it does. The reason this was not visible in the thread dump is 
that we have not reached the point of <clinit> execution for 
MutableHttpFields, but are blocked trying to get the class 
initialization lock for a superclass/interface to ensure it is 
initialized first.
> '[8.061s][debug][class,init] Thread ForkJoinPool-1-worker-1 is
> initializing org.eclipse.jetty.http.HttpFields'.
> ...
> '[8.113s][debug][class,init] Thread ForkJoinPool-1-worker-2 is
> initializing org.eclipse.jetty.http.MutableHttpFields'.
> '[8.113s][debug][class,init] Thread ForkJoinPool-1-worker-2 waiting
> for initialization of org.eclipse.jetty.http.HttpFields by thread
> ForkJoinPool-1-worker-1'.
> ...
> '[8.117s][debug][class,init] Thread ForkJoinPool-1-worker-1
> recursively initializing org.eclipse.jetty.http.HttpFields'.
> '[8.117s][debug][class,init] Thread ForkJoinPool-1-worker-1 waiting
> for initialization of org.eclipse.jetty.http.MutableHttpFields by
> thread ForkJoinPool-1-worker-2'.
> 
> It is peculiar that we already get this deadlock for the same classes.
> Perhaps we are using a pattern in the code that makes it more likely to happen?

Perhaps. I don't know the jetty code at all so you would need to go back 
to them with these findings. If two classes have a dependency on each 
other this way then you would normally expect there to be a preferred 
"entry point" so that initialization will always occur in the same 
order. But it is a fragile design where a superclass/interface uses its 
own subclass in its static initialization.

Glad this helped expose the problem.

Cheers,
David
-----

> Thanks!
>