RFR (S) JDK-8008962: NPG: Memory regression: One extra Monitor per ConstantPool

[David Holmes]

Hi Ioi,

On 16/03/2013 5:07 PM, Ioi Lam wrote:
> Please review:
> http://cr.openjdk.java.net/~iklam/8008962/constpool_lock_001/
> <http://cr.openjdk.java.net/%7Eiklam/8008962/constpool_lock_001/>
>
> Bug: 8008962: NPG: Memory regression: One extra Monitor per ConstantPool
> http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=8008962
> https://jbs.oracle.com/bugs/browse/JDK-8008962 (Oracle Internal)
>
> Summary of fix:
>
> The fix is implemented according to Coleen's suggestion in the bug's
> comments.
>
> Instead of using a Monitor for each ConstantPool (~150 bytes per
> class), I re-use InstanceKlass::init_lock() for locking the
> ConstantPool.
>
> init_lock used to be cleared after the class had been initialized.
> There was complicated code that needed to deal with the clearing
> (using volatile, OrderAccess::storestore(), etc). This code is no
> longer necessary and has been removed.

But the whole point of removing it was that we did not want the extra 
memory usage to exist for the lifetime of the class! This will increase 
dynamic footprint in a potentially unacceptable way.

>      The overhead is reduced to one int[0] per class.
>
> Because the lock is used at two different levels (class
> initialization, and ConstantPool access), I thought about the
> possibility of deadlocks.

Top understand the deadlock potential I need to understand what locks 
the constant pool and why/when. Is there a possibility of a 
self-deadlock if during class initialization we have to lock the 
constant-pool ourselves?

Thanks,
David
------


> However,
>      it seems the only case of a deadlock would be some sort of circular
> dependency
>      in running <clinit> (as in the following example). However, in this
> case, a
>      deadlock would have already happened, just from grabbing the
> initialization
>      locks, before we would even have a chance to access the
> ConstantPools. Hence,
>      reusing the lock for ConstantPool access doesn't add any new paths
> for deadlocks.
>
>      -------------------------------
>      public class ConcurClinit {
>          static void sleep(int ms) {
>              try {Thread.sleep(ms);} catch (Throwable t) {;}
>          }
>          static class A {
>              static {
>                  sleep(1000);
>                  B.xxx();
>                  sleep(1000);
>              }
>              static void xxx() {}
>          }
>          static class B {
>              static {
>                  sleep(1000);
>                  A.xxx();
>                  sleep(1000);
>              }
>              static void xxx() {}
>          }
>          public static void main(String args[]) {
>              (new Thread() {
>                      public void run() {
>                          A a = new A();
>                      }
>                  }).start();
>              (new Thread() {
>                      public void run() {
>                          B b = new B();
>                      }
>                  }).start();
>          }
>      }
>      -------------------------------
>
> Tests executed:
>
>      * JPRT
>      * JCK (vm, lang, api/signaturetest)
>      * UTE (vm.runtime.testlist, vm.quick.testlist,
> vm.parallel_class_loading.testlist
> )
>
> Thanks,
> Ioi