[Chicken-users] Obscure bug with srfi-18, mailbox and coops

[Théo Cavignac]

Hello,

I am currently trying to write a multi-format (JSON and msgpack) multi-transport (TCP, Unix ports, file ports) RPC server and client egg for Chicken. I have a decent architecture for my program (I think) but I currently face an inexplicable bug that randomly appear and that I cannot explain.

I used a simple actor model to handle concurrency in a clean way: actors communicate through message passing and handle their internal states in a single thread so that race conditions are impossible. To implement this model I used srfi-18 threads, mailbox and COOPS classes. An actor is an instance of a class inheriting from the <actor> base class I defined. It loops over its mailbox and run actions for each message it received.

The software isn't complete yet but I am working on the test suite to validate the features before going any further. Here is my problem:

When I am running the test suite I run the server on a separate thread (for now each actor of the server is on another thread too) and I try to send messages. however, I got the following error:

Warning (#<thread: sch>): in thread: (mailbox-send!) bad argument type - not a mailbox: #f

    Call history:

    src/main/actor.scm:37: mailbox#mailbox-send!          <--

However, the given line is perfectly fine:

36  (define-method (send (self <actor>) msg data)
37    (mailbox-send! (slot-value self 'private-mailbox) (cons msg data)))

And I see no way that the (slot-value self 'private-mailbox) is different from its original value:

(define-class <actor> () ((private-mailbox (make-mailbox)) (private-continue #t)))

Indeed, this is a "virtual" class (never instantiated directly but supposed to be inherited) and the child classes never access or overwrite private-* attributes (that why they got this prefix).

Here it appears to be the one-shot slot from the scheduler class that correspond to this #f but I got the same bug with different actors and different slots from the corresponding classes.

I was then wondering if I got something wrong or if there is a deeper bug in the combo srfi-18/coops/mailbox.

If you want to have a look at the complete code you can find it at https://github.com/Lattay/chicken-meta-rpc. You may want to look at src/main/actor.scm (definition of the actor base class) or to src/main/server.scm (definitions of the actual actors).

If you want to try it you may run `make test`. Dependencies are matchable, mailbox, srfi 18 and 69 and coops.

Hoping someone could help, thank you for reading so far.

PS: I don't ask for a code review (do it if you feels like it though) since it is not a fun thing to do but I want to check if this bug ring a bell for some of you.