On 12/3/19 8:31 AM, David Holmes wrote:
On 3/12/2019 11:08 pm, coleen.phillim...@oracle.com wrote:
On 12/2/19 11:52 PM, David Holmes wrote:
Hi Coleen,
On 3/12/2019 12:43 am, coleen.phillim...@oracle.com wrote:
On 11/26/19 7:03 PM, David Holmes wrote:
(adding runtime as well)
Hi Coleen,
On 27/11/2019 12:22 am, coleen.phillim...@oracle.com wrote:
Summary: Add local deferred event list to thread to post events
outside CodeCache_lock.
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
deferred event?
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 think I'm getting the picture now. At the time the events are
generated we can't post them directly because the current thread is
inside compiler code. Hence the events must be deferred. Using the
ServiceThread to handle the deferred events is one way to deal with
this - but it can't keep up in this scenario. So instead we store
the events in the current thread and when the current thread returns
to code where it is safe to post the events, it does so itself. Is
that generally correct?
Yes.
I admit I'm not keen on adding this additional field per-thread just
for a temporary usage. Some kind of stack allocated helper would be
preferable, but would need to be passed through the call chain so
that the events could be added to it.
Right, and the GC and nmethods_do has to find it somehow. It wasn't
my first choice of where to put it also because there is too many
things in JavaThread. Might be time for a future cleanup of Thread.
I see.
Also I'm not clear why we aggressively delete the _jvmti_event_queue
after posting the events. I'd be worried about the overhead we are
introducing for creating and deleting this queue. When the
JvmtiDeferredEventQueue data structure was intended only for use by
the ServiceThread its dynamic node allocation may have made more
sense. But now that seems like a liability to me - if
JvmtiDeferredEvents could be linked directly we wouldn't need
dynamic nodes, nor dynamic per-thread queues (just a per-thread
pointer).
I'm not following. The queue is for multiple events that might be
posted while in the CodeCache_lock, so they need to be in order and
linked together. While we post them and take them off, if the
callback safepoints (maybe calls back into the JVM), we don't want to
have GC or nmethods_do walk the one that's been posted already. So a
queue seems to make sense.
Yes but you can make a queue just by having each event have a _next
pointer, rather than dynamically creating nodes to hold the event.
Each event is its own queue node implicitly.
One thing that I experimented with was to have the ServiceThread take
ownership of the queue in it's local thread queue and post them all,
which could be a future enhancement. It didn't help my OOM situation.
Your OOM situation seems to be a basic case of overwhelming the
ServiceThread. A single serviceThread will always have a limit on how
many events it can handle. Maybe this test is being too unrealistic in
its expectations of the current design?
I think the JVMTI API where you can generate an COMPILED_METHOD_LOAD for
all the events in the queue is going to be overwhelming unless it waits
for the events to be posted.
Deleting the queue after all the events are posted allows
JavaThread::oops_do and nmethods_do only a null check to deal with
this jvmti wart.
If the nodes are not dynamically allocated you don't need to delete
you just set the queue-head pointer to NULL - actually it will already
be NULL once the last event has been processed.
I could revisit the data structure as a future RFE. The goal was to
reuse code that's already there, and I don't think there's a significant
difference in performance. I did some measurement of the stress case
and the times were equivalent, actually better in the new code.
Thanks,
Coleen
David
-----
Thanks,
Coleen
Just some thoughts.
Thanks,
David
I did write comments to this effect here:
http://cr.openjdk.java.net/~coleenp/2019/8212160.01/webrev/src/hotspot/share/prims/jvmtiCodeBlobEvents.cpp.udiff.html
Thanks,
Coleen
David
Also, the jmethod_id field in nmethod was only used as a boolean
so don't create a jmethod_id until needed for
post_compiled_method_unload.
Ran hs tier1-8 on linux-x64-debug and the stress test that
crashed in the original bug report.
open webrev at
http://cr.openjdk.java.net/~coleenp/2019/8212160.01/webrev
bug link https://bugs.openjdk.java.net/browse/JDK-8212160
Thanks,
Coleen
