Request for review 8000662: NPG: nashorn ant clean test262 out-of-memory with Java heap
david.holmes at oracle.com
Sun Nov 11 20:32:57 PST 2012
Can you explain the old and new relationships between these various
objects please. As I understand it JSR-292 introduced this special kind
of class - the anonymous class - and they differ from regular classes in
important ways (mainly that they can be collected before their
classloader), and pre-NPG this all "just worked". The NPG changes added
the ClassLoaderData metadata objects and changed some of the
associations that anonymous classes actually relied upon and as a result
they stopped getting GC'd in some cases, and were prematurely GC'd in
others (I hope I got that right). This changeset addresses these problems.
I've been trying to follow this from the beginning and still don't have
a clear understanding on what the relationships between ClassLoader,
Class, "anonymous class" and ClassLoaderData objects are. So a clear
picture of what relationships exist (particularly this new 1:N
association) would be appreciated, and help make the code changes more
understandable to me.
On 12/11/2012 1:51 PM, Coleen Phillimore wrote:
> This change creates a ClassLoaderData object for each JSR 292 anonymous
> class, so that the metadata and mirror object for the anonymous class
> can be reclaimed when the anonymous class is no longer referenced. The
> java_lang_ClassLoader object for the anonymous class is the same as its
> host_class. Most of this change is to break the assumption that there's
> a 1-1 relationship between java_lang_ClassLoader Java objects and
> ClassLoaderData metadata objects in the VM. Also, nmethods and other
> things that were strong roots for java_lang_ClassLoader objects have to
> also be strong roots for java_lang_Class objects for anonymous classes.
> There were also changes to move the dependencies out of the
> java_lang_ClassLoader object to the ClassLoaderData type. This type is
> preallocated so that CMS will have card marks to track additions to the
> dependencies. Please check this, Stefan!
> Also, in this change is the addition of mirrors to the interpreter frame
> and a test case that shows why this is necessary. An interpreter frame
> can be running while it's class loader is unloaded in some special
> circumstances. It is easier to do this with JSR292 static MethodHandle
> code. Some people are looking for a platform independent way to do this,
> by changing code in GC. While this target-dependent interpreter code is
> unfortunate, the concept is simple. If the latter effort succeeds, we
> can remove the mirror from the interpreter frame later. A note to
> openjdk developers - I added this mirror to zero but not to shark. More
> testing is necessary.
> Please review the following change:
> Summary: Add ClassLoaderData object for each anonymous class with
> metaspaces to allocate in. Add mirror interpreter frame.
> Tested with Nashorn tests, NSK full testlist, dacapo with CMS and G1.
More information about the hotspot-dev