Integrated: 8296419: [REDO] JDK-8295319: pending_cards_at_gc_start doesn't include cards in thread buffers

[Kim Barrett]

On Wed, 9 Nov 2022 04:57:38 GMT, Kim Barrett <kbarrett at openjdk.org> wrote:

> Let's try this again.  This is the original JDK-8295319 change, plus a couple
> of small additions to fix problems with that original change.
> 
> The description of the original change was:
> 
> -----
> 
> Please review this change to G1 to include the per-thread buffers in the
> number of pending cards at the start of a young GC.
> 
> DCQS::concatenate_logs has been renamed to concatenate_logs_and_stats, and now
> also merges the per-thread refinement stats during the thread walk to flush
> buffers. That replaces the separate thread walk to merge and record these
> stats earlier in the GC. The merged stats and related info don't seem to be
> needed until after the buffer flushing.
> 
> Also, when abandoning dirty card buffers and stats because of a full GC, fixed
> to also abandon any buffers in the paused buffers lists.
> 
> -----
> 
> The problem was that by moving log concatenation earlier, it no longer
> followed the call to retire_tlabs().  Among other things, that function
> calls flush_deferred_card_mark_barrier() on each thread, possibly adding cards
> to the thread's dirty card queue.  Doing that after the log concatenation may
> lead to unprocessed cards in thread queues, with chaos ensuing.
> 
> To fix this, the call to retire_tlabs has also been moved, so that it once
> again precedes the log concatenation.  Also added (debug-only) verification
> that all thread dirty card queues are empty at the end of
> pre_evacuate_collection_set.
> 
> A possible followup is to refactor those two operations, which each do a
> (single-threaded) walk of all threads.  We could combine that into a single
> walk.  We could also parallelize it if that seems warranted, though the
> per-thread work is usually pretty small, so might not be worth parallelizing.
> 
> There might be an opportunity to do some similar refactoring for fullgc and
> log abandonment, though there the two operations are done far away from each
> other.
> 
> Testing:
> mach5 tier1-6

This pull request has now been integrated.

Changeset: 1fb9dadb
Author:    Kim Barrett <kbarrett at openjdk.org>
URL:       https://git.openjdk.org/jdk/commit/1fb9dadb6aadb5a12a3b14798d1cd58ad50ba0b8
Stats:     183 lines in 9 files changed: 74 ins; 71 del; 38 mod

8296419: [REDO] JDK-8295319: pending_cards_at_gc_start doesn't include cards in thread buffers

Reviewed-by: tschatzl, sjohanss

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

PR: https://git.openjdk.org/jdk/pull/11053