RFR: 8315362: NMT: summary diff reports threads count incorrectly [v5]

[Thomas Stuefe]

On Tue, 10 Oct 2023 13:55:05 GMT, Evgeny Ignatenko <duke at openjdk.org> wrote:

>> 8315362: NMT: summary diff reports threads count incorrectly
>
> Evgeny Ignatenko has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Not check exact amount of added threads

There are some things I would like to see changed:

- The malloc tracker should have no knowledge of threads or thread count. Malloc has nothing to do with threads. This is a partly pre-existing issue.

- "track thread stacks as VM" - resp. its negation for AIX - was a workaround we did for AIX because AIX has no page-aligned thread stacks. I would love to get rid of that hack altogether. If not-being-page-aligned is the only issue, one simple way would be to, on AIX, align the boundaries of thread stacks up/down so that they start at whole page boundaries (in `ThreadStackTracker::new_thread_stack()`). It would be a trivial loss in fidelity, but we could get rid of this whole mechanism. @gerard-ziemski maybe? You like disentangling stuff :-)


For this patch, my proposal would be:

- make the setting of `volatile size_t ThreadStackTracker::_thread_count` in `ThreadStackTracker::(new|delete)_thread_stack()` unconditional. So, lets do this for AIX too.

- now we can always use `ThreadStackTracker::thread_count()` unconditionally when snapshoting, which removing dependence on "track_as_vm"

- Let the new `_thread_count` snapshot value, and the snapshoting itself, live in `MemBaseline`, not in the malloc snapshot. That's a better fit. Let's remove `MallocSnapshot::thread_count()` completely.

Side note, `size_t` was always an odd value for thread count. We use `unsigned` in other places since for thread counts that is enough. Would be nice to have this changed too.

Thanks!

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

PR Review: https://git.openjdk.org/jdk/pull/15490#pullrequestreview-1668233486