RFR: JDK-8202788: Explicitly reclaim cached thread-local direct buffers at thread exit

[Tony Printezis]

Peter,

In my proposal, you can register the exit hook in the ThreadLocal c’tor, so
it’s almost as nice as Alan’s in that respect (and it doesn't require an
extra field per ThreadLocal plus two extra fields per JdkEntry). :-)

But, I do like the addition of the JdkEntry list to avoid unnecessarily
iterating over all the map entries (which was my main concern with Alan’s
original webrev). I’ll be totally happy with a version of this.

Tony


—————
Tony Printezis | @TonyPrintezis | tprintezis at twitter.com


On May 12, 2018 at 6:44:08 AM, Peter Levart (peter.levart at gmail.com) wrote:

Hi,

On 05/11/18 16:13, Alan Bateman wrote:

On 08/05/2018 16:07, Tony Printezis wrote:

Hi all,

Following the discussion on this a few weeks ago, here’s the first version
of the change:

http://cr.openjdk.java.net/~tonyp/8202788/webrev.0/

I think the consensus was that it’d be easier if the exit hooks were only
available within java.base. Is it enough that I added the functionality to
the jdk.internal.misc package? (And is jdk.internal.misc the best place for
this?)

I’ll also add a test for the new functionality. But let’s first come up
with an approach that everyone is happy with. :-)

Peter's approach in early April was clean (and we should come to the
getIfPresent discussion) but it adds a field to Thread for the callback
list. If I read your approach correctly, you are avoiding that by
maintaining an array of hooks in ThreadLocalExitHooks.

Another approach to try is a java.base-internal ThreadLocal that defines a
method to be invoked when a thread terminates. Something like the
following:
   http://cr.openjdk.java.net/~alanb/8202788/webrev/index.html

-Alan


>From the API perspective, Alan's suggestion is the most attractive one as
it puts the method to where it belongs - into the ThreadLocal instance. But
the implementation could be improved a bit. I don't like the necessity to
iterate over all ThreadLocal(s) to filter out JdkThreadLocal(s). There
might be a situation where there's plenty of ThreadLocal(s) registered per
exiting thread which would produce a spike in CPU processing at thread exit.

The way to avoid this would be to maintain a separate linked list of
entries that contains just those with JdkThreadLocal(s). Like in this
modification of Alan's patch, for example:

    http://cr.openjdk.java.net/~plevart/jdk-dev/DBBCache_Cleanup/webrev.01/

(Only ThreadLocal class is modified from Alan's patch)

What do you think?


Regards, Peter