Hi Peter, > On Apr 5, 2016, at 7:41 AM, Peter Levart <peter.lev...@gmail.com> wrote: > > The bottom line is that we need a mechanism that: > > - triggers reference discovery when native memory limit is approached or > reached > - retires native memory reservation at appropriate time slots until > succeeding or until all pending references have been processed and Cleanables > executed at which time native memory reservation can fail with OOME. > - if possible, doesn't execute cleanup functions by the allocating thread but > just waits for system threads to do the job. > - when triggered, does not make native memory allocation a bottleneck. > > I think that what I did in my latest webrevs with ReferenceHandler thread is > an improvement in minimizing contended synchronization and interference of > allocating thread(s) with Reference enqueue-ing. But interaction of > allocating thread(s) with Cleaner background thread could be improved and I > have a couple of ideas to explore.
This is about timely native memory deallocation. Since direct byte buffer is the only one using jdk.internal.ref.Cleaner, I am inclined to suggest keep jdk.internal.ref.Cleaner as is and replaced it with a better mechanism when it’s available, either panama or better pending reference enqueuing that you and Per discussed. Given the time we have, I think it’s likely post JDK 9 timeframe. Minimizing contention and interference of allocating threads and ReferenceHandler may worth exploring. Mandy
