Hi David,
I changed to normal for:
http://rehn-ws.se.oracle.com/cr_mirror/8223306/v2/webrev/src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/CommandProcessor.java.sdiff.html
Full:
http://rehn-ws.se.oracle.com/cr_mirror/8223306/v2/webrev/
Inc:
http://rehn-ws.se.oracle.com/cr_mirror/8223306/v2/inc/webrev/
Passes t1-2
Thanks, Robbin
On 2019-05-07 09:47, David Holmes wrote:
Hi Robbin,
On 7/05/2019 4:50 pm, Robbin Ehn wrote:
Hi David,
On 5/7/19 12:40 AM, David Holmes wrote:
Hi Robbin,
I have a few concerns here.
First I can't see how you are actually integrating the SA with the ThreadSMR.
You've exposed the _java_thread_list for it to iterate but IIRC that list can
be updated when threads are added/removed and I'm not seeing how the SA is
left iterating a valid list - we'd normally using a ThreadsListHandle for
that ?? (I may need a refresher on how this list is actually maintained.)
The processes must be paused. If the processes would be running the linked
list is also broken since if we unlink and delete a JavaThread and then later
SA follows that _next pointer.
Ah good point. Thanks for clarifying.
The conversion from external iteration of the list (the for loop) to internal
iteration (passing a lambda to JavaThreadsDo) is also problematic. First I'd
be very wary about introducing lambda expressions into the SA code - lambda
drags in a lot of supporting code that could have an impact on the way SA
functions. There are places where we have to avoid lambdas due to
bootstrapping/initialization issues and I think the SA may be an area where
we also want to keep the code extremely simple.
There are already several usages of lambdas in SA code, e.g.
LinuxDebuggerLocal.java. SA is not a core module, it's an application, there
is not a bootstrap issue or so.
Hmm okay. Lambda carries a lot of baggage. But if its already being used ...
Second by using lambda's with internal iteration you've lost the ability to
terminate the iteration loop! In the existing code if we have a return in the
for-loop then we not only terminate the loop but the enclosing method. With
the lambda the "return" just ends the current iteration and JavaThreadsDo
will then continue with the next thread - so we don't even terminate the
iteration let alone the method performing the iteration. So for places were
we want to process one thread, for example, we will continue to iterate all
remaining threads and just do nothing with them. That's very inefficient.
That's why I only used the internal iteration where we didn't have early
returns.
src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/CommandProcessor.java -
original code:
1556 new Command("where", "where { -a | id }", false) {
1557 public void doit(Tokens t) {
...
1564 for (JavaThread thread = threads.first(); thread !=
null; thread = thread.next()) {
1565 ByteArrayOutputStream bos = new
ByteArrayOutputStream();
1566 thread.printThreadIDOn(new PrintStream(bos));
1567 if (all || bos.toString().equals(name)) {
1568 out.println("Thread " + bos.toString() + "
Address: " + thread.getAddress());
...
1577 }
1578 if (!all) return;
That looks like an early return to me.
Cheers,
David
-----
Thanks, Robbin
Thanks,
David
On 6/05/2019 5:31 pm, Robbin Ehn wrote:
Hi, please review.
Old threads linked list remove and updated SA to use ThreadsList array instead.
Issue:
https://bugs.openjdk.java.net/browse/JDK-8223306
Webrev:
http://cr.openjdk.java.net/~rehn/8223306/webrev/
Passes t1-3 (which includes SA tests).
Thanks, Robbin
