Finalizers which suspend are terminated immediately

[halorgium]

When a finalizer runs and it's Task suspended, the finalizer is killed immediately after that time by the Actor#cleanup code.
In supervision groups, the finalizer is responsible for terminating other actors, which uses a SyncCall which suspends the task.
The main problem is that there is no more main loop to handle responses to method calls after the shutdown has started.

#<Thread:0x37e247e2 run>: nil: "start finalizer"
#<Celluloid::InternalPool::Thread:0x34c7da23 run>: "cd0813e7-6f56-4849-89ec-000000010000": "terminate #<Celluloid::TaskFiber:0x180c @type=:finalizer, @status=:callwait>"

This has implications and confusions on @jruby where Threads seem to stick around if you have ConditionVariables in use. See #203

[halorgium]

Marking the Task as unsuspendable and raising if it is attempted.
But why is it in a Task to begin with?

[tarcieri]

That's a good question... I was digging through the commit history trying to find why I made that change originally, and couldn't find it.

I think it'd make more sense to either run the finalizer in exclusive mode, or as you said, raise if anything attempts to suspend the current task.

[jjyr]

seems run finalizer in exclusive is better, user can just care the actor is alive or dead, not other.

[halorgium]

@jjyr i would prefer to be able to detect the attempt to suspend.
It is possible to deadlock if we use exclusive.

What do you mean by?

user can just care the actor is alive or dead

[jjyr]

@halorgium i mean when i call the actor, i expected it be alive or dead, but not unsuspend
run finalizer in exclusive make i just call a dead actor, i think to introduce a new state(or just a new error) is not necessary
but i'm not sure about the deadlock

[halorgium]

@jjyr the main problem I foresee is that at some point we will have to suspend the finalizer and allow new messages to arrive and be handled in the case of shutting down a tree of actors.

[jjyr]

@halorgium yes..i did not notice that...:neutral_face:

[halorgium]

@jjyr do not worry, I am glad you have been providing feedback and looking at the code.
A lot of things in @celluloid are still uncertain and probably incorrect.
The best place to have some of these discussions might be in IRC. I hang out there quite a bit 😀

[tarcieri]

A lot of things in @celluloid https://github.com/celluloid are still
uncertain and probably incorrect.

A lot of things in Akka and Erlang are still uncertain and probably
incorrect ;)

[halorgium]

This is also true.

[halorgium]

Task termination now logs, so users should make their finalizers exclusive if they need this behavior.