2022-02-01 15:22:30 Two problems.
- If a stop action panics (in `_terminate_facet`), the Facet is dropped before its outbound
handles are removed. With the code as it stands, this leaks assertions (!!).
- The logic for removing an outbound handle seems to be running in the wrong facet context???
(See `f.outbound_handles.remove(&handle)` in the cleanup actions
- I think I need to remove the for_myself mechanism
- and add some callbacks to run only on successful commit
2022-02-02 12:12:33 This is hard.
Here's the current implementation:
- assert
- inserts into outbound_handles of active facet
- adds cleanup action describing how to do the retraction
- enqueues the assert action, which
- calls e.assert()
- retract
- looks up & removes the cleanup action, which
- enqueues the retract action, which
- removes from outbound_handles of the WRONG facet in the WRONG actor
- calls e.retract()
- _terminate_facet
- uses outbound_handles to retract the facet's assertions
- doesn't directly touch cleanup actions, relying on retract to do that
- if one of a facet's stop actions panics, will drop the facet, leaking its assertions
- actually, even if a stop action yields `Err`, it will drop the facet and leak assertions
- yikes
- facet drop
- panics if outbound_handles is nonempty
- actor cleanup
- relies on facet tree to find assertions to retract
Revised plan:
- ✓ revise Activation/PendingEvents structures
- rename `cleanup_actions` to `outbound_assertions`
- remove `for_myself` queues and `final_actions`
- add `pre_commit_actions`, `rollback_actions` and `commit_actions`
- ✓ assert
- as before
- but on rollback, removes from `outbound_handles` (if the facet still exists) and
`outbound_assertions` (always)
- marks the new assertion as "established" on commit
- ✓ retract
- lookup in `outbound_assertions` by handle, using presence as indication it hasn't been
scheduled in this turn
- on rollback, put it back in `outbound_assertions` ONLY IF IT IS MARKED ESTABLISHED -
otherwise it is a retraction of an `assert` that has *also* been rolled back in this turn
- on commit, remove it from `outbound_handles`
- enqueue the retract action, which just calls e.retract()
- ✓ _terminate_facet
- revised quite a bit now we rely on `RunningActor::cleanup` to use `outbound_assertions`
rather than the facet tree.
- still drops Facets on panic, but this is now mostly harmless (reorders retractions a bit)
- handles `Err` from a stop action more gracefully
- slightly cleverer tracking of what needs doing based on a `TerminationDirection`
- now ONLY applies to ORDERLY cleanup of the facet tree. Disorderly cleanup ignores the
facet tree and just retracts the assertions willy-nilly.
- ✓ facet drop
- warn if outbound_handles is nonempty, but don't do anything about it
- ✓ actor cleanup
- doesn't use the facet tree at all.
- cleanly shutting down is done elsewhere
- uses the remaining entries in `outbound_assertions` (previously `cleanup_actions`) to
deal with retractions for dropped facets as well as any other facets that haven't been
cleanly shut down
- ✓ activate
- now has a panic_guard::PanicGuard RAII for conveying a crash to an actor in case the
activation is happening from a linked task or another thread (this wasn't the case in the
examples that provoked this work, though)
- simplified
- explicit commit/rollback decision
- ✓ Actor::run
- no longer uses the same path for crash-termination and success-termination
- instead, for success-termination, takes a turn that calls Activation::stop_root
- this cleans up the facet tree using _terminate_facet
- when the turn ends, it notices that the root facet is gone and shuts down the actor
- so in principle there will be nothing for actor cleanup to do
2022-02-04 13:52:34 This took days. :-(
Tony Garnock-Jones
f88592282d