RFR: 8319048: Monitor deflation unlink phase prolongs time to safepoint [v7]

[Aleksey Shipilev]

On Mon, 20 Nov 2023 10:39:16 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> src/hotspot/share/runtime/synchronizer.cpp line 108:
>> 
>>> 106:         unlinked_list->append(next);
>>> 107:         next = next_next;
>>> 108:         if (unlink_batch++ >= (size_t)MonitorUnlinkBatch) {
>> 
>> Previously we `break` on the `MonitorDeflationMax` which is a implicit limit of the number of deflations. Not as `MonitorUnlinkBatch` could be `MonitorDeflationMax-1` then we would could walk more links than needed. 
>> 
>> Overall I am unsure why we do not pass in deflated_count to unlink and break when the `unlink_count == deflated_count`. Maybe I am missing something and there can be more deflated monitors than  the deflation_count we get. If so it seems like both this and the previous implementation would miss unlinking deflated monitors.
>
> Walking more links is fine. Good suggestion to just pass the `deflated_count` here! Let me do that.

Actually, now that we have precise `deflated_count`, we might cap the next batch at `deflated_count - unlinked_count` to avoid extra walk, without extra overhead.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/16412#discussion_r1399002732