RFR: Strong roots scan should include thread stack oops

[Roman Kennke]

Am 06.07.2017 um 11:42 schrieb Aleksey Shipilev:
> We call into RP::process_strong_roots during init mark here:
>
>     if (heap->concurrentMark()->unload_classes()) {
>       _rp->process_strong_roots(&mark_cl, _process_refs ? NULL : &mark_cl,
>                                 &cldCl, &blobsCl, worker_id);
>
> ...with the rationale:
>
>  a. With unload_classes = true, we only want to scan the actual strong roots
>     from the code cache. This will allow us to identify the dead classes,
>     unload them, *and* invalidate the relevant code cache blobs. This could
>     be only done together with class unloading.
>
> However, inside process_strong_roots definition, we ignore CodeBlobClosure!
> Instead, we should pass it down to process_java_roots, which will pass it to
> stack code oop scans, which are by definition, strong. Makes sense?
>
> Fix:
>   http://cr.openjdk.java.net/~shade/shenandoah/rs-strong-coderoots/webrev.01/
>
> Testing: hotspot_gc_shenandoah
>
> Thanks,
> -Aleksey
>
Eh. WTF?

Yes, please push.

I'd really like to know how that happened and how could this possibly
work at all? How did we not unload most of the code cache?!

Roman