On 12/3/19 7:48 AM, erik.osterl...@oracle.com wrote:
This looks great. Thanks for sorting this out!
On 12/2/19 3:43 PM, coleen.phillim...@oracle.com wrote:
On 11/26/19 7:03 PM, David Holmes wrote:
(adding runtime as well)
On 27/11/2019 12:22 am, coleen.phillim...@oracle.com wrote:
Summary: Add local deferred event list to thread to post events
This patch builds on the patch for JDK-8173361. With this patch, I
made the JvmtiDeferredEventQueue an instance class (not AllStatic)
and have one per thread. The CodeBlob event that used to drop the
CodeCache_lock and raced with the sweeper thread, adds the events
it wants to post to its thread local list, and processes it outside
the lock. The list is walked in GC and by the sweeper to keep the
nmethods from being unloaded and zombied, respectively.
Sorry I don't understand why we would want/need a deferred event
queue for every JavaThread? Isn't this only relevant for
non-JavaThreads that need to have the ServiceThread process the
I thought I'd written this in the bug but I had only discussed this
with Erik. I've added a comment to the bug to explain why I added
the per-JavaThread queue. In order to process these events after the
CodeCache_lock is dropped, I have to queue them somewhere safe. The
ServiceThread queue is safe, *but* the ServiceThread can't keep up
with the events, especially from this test case. So the test case
gets a native OOM.
So I've added the safe queue as a field to each JavaThread because
multiple JavaThreads could be posting these events at the same time,
and there didn't seem to be a better safe place to cache them,
without adding another layer of queuing code.
I did write comments to this effect here:
Also, the jmethod_id field in nmethod was only used as a boolean so
don't create a jmethod_id until needed for
Ran hs tier1-8 on linux-x64-debug and the stress test that crashed
in the original bug report.
open webrev at
bug link https://bugs.openjdk.java.net/browse/JDK-8212160