fix(sdk): Introduce SubscribersHandle and SubscriberHandle

[Hywan]

This PR is twofold.

When a cache (from the Event Cache) has no more subscribers, it auto-shrinks to its last chunk to save memory. That's great! The way it counts/tracks the number of subscribers isn't efficient though. It uses an Arc<AtomicU32>. Arc in itself is already an atomic counter. We can use Arc<()> directly, saving the cost of more atomic operations.

The first main patch introduces SubscribersHandle that is owned by the cache state (like RoomEventCacheState). When a subscriber is created, SubscribersHandle::new_subscriber_handle (note the singular form) is called, and produces a SubscriberHandle (still singular form).

A SubscriberHandle is a thin type around Weak<()>. The state has an Arc, the subscribers have a Weak. The number of subscribers is calculated by (Arc|Weak)::weak_count. That way, no need to fetch_add or load to calculate the count.

The second main patch removes the sending of the VectorDiffs updates when the cache is shrunk. The documentation said:

Hey, fun stuff: we shrunk the linked chunk, so there shouldn't be any subscribers, right? RIGHT? Especially because the state is guarded behind a lock.

However, better safe than sorry, and it's cheap to send an update here, so let's do it!

I think that's an error. Two situations here:

1. No race, no subscribers have been registered, so it's safe to do nothing,
2. A race, a subscriber has been created meanwhile, we must not send the diff to it, otherwise it can create an invalid state.

Note that a race shouldn't be possible as we have acquired an exclusive access to the state, ensuring no subscriber can be created, so technically the diff should be sent to nowhere, but it's really important to not send it :-].

---

• Address [meta] Event Cache layers refactoring #6152

---

• I've documented the public API changes in the appropriate changelog files (see Writing changelog entries).
• This PR was made with the help of AI.

Signed-off-by:

[codecov]

Codecov Report

❌ Patch coverage is 98.33333% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 89.87%. Comparing base (e8673e4) to head (d7a6535).
⚠️ Report is 9 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...ates/matrix-sdk/src/event_cache/caches/room/mod.rs	88.88%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6684      +/-   ##
==========================================
- Coverage   89.88%   89.87%   -0.02%     
==========================================
  Files         396      396              
  Lines      110264   110307      +43     
  Branches   110264   110307      +43     
==========================================
+ Hits        99115    99142      +27     
- Misses       7385     7386       +1     
- Partials     3764     3779      +15     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

[codspeed-hq]

Merging this PR will not alter performance

✅ 50 untouched benchmarks

---

_{Comparing Hywan:fix-sdk-event-cache-shrink-all-caches (d7a6535) with main (4bd5e81)}

[poljar]

I think this is mostly fine.

I wonder if we couldn't just use the strong_count() on the auto_shrink_sender() to get the subscriber count.

Also, how come no tests needed to be modified for the last patch? 🙈

[Hywan]

I wonder if we couldn't just use the strong_count() on the auto_shrink_sender() to get the subscriber count.

Actually, we need to count the number of subscribers, so the weak_count(). The strong_count() would always give 1, as it lives in RoomEventCacheState only.

Also, how come no tests needed to be modified for the last patch? 🙈

Because I struggle to see how to test it. The auto-shrink happens if and only if no more subscriber exists. How to test nothing is sent if we can't receive it…? Hmm, now I'm giving it a second thought, it might be possible if we write the test at the correct place as a unit test, so that we can have access to the private fields. Let me try.