CL-MTGNET is a client library for the MTGNet RPC protocol. The library handles all the details of marshalling and unmarshalling, including matching out-of-order responses to the appropriate request.
This library depends on the CL-NETSTRING+ library, as well as the quicklisp-installable CL-JSON and TRIVIAL-UTF-8 libs.
Remote methods are treated like FFI calls: the client creates definitions for the methods it wishes to use, and can then call those methods as if they were local functions.
Most clients will want to use the high-level macros to define method functions that handle the details of calling. The method definition macro is relatively simple:
(defmacro define-rpc-method ((method &optional service &key notification) &body args))
service are symbols naming the method and the service
method belongs to.
define-rpc-method will create a function named
<method>, which takes a socket stream, a symbol naming the service, and the arguments of the method. If
service is provided, the function will be named
<service>-<method> and will not take a service symbol as a parameter. Note that both
service symbols will be processed by CL-JSON, so the standard name-mangling applies (e.g.
describe-method will become
describeMethod when sent to the server).
notification is not supplied or is
nil, the function will return a future object that will return the value of the method (or signal an error) when
wait is called on it. If
notification is not
nil, the function will return immediately with no value (since no response is expected from the server).
;; Define a method without specifying a service (define-rpc-method ;; Method name (describe-method) ;; Method arguments method-name) ;; Invoke the method on the 'server' service ;; (wait (describe-method sock 'server "version")) => [...] ;; Define a method on a specific service (define-rpc-method (describe-method server) method-name) ;; (wait (server-describe-method "version")) => [...] ;; Define a notification method (define-rpc-method (poke server :notification t) msg) ;; Note the lack of a WAIT call here ;; (server-poke "Poked for no reason!") => No value
The MTGNet protocol features the ability to send requests to the server in a batch, with the guarantee that the server will hold the responses for those calls until they're all ready before returning them. Clients can submit a batch of calls using the
;; With STREAM, DESCRIBE-METHOD and VERSION defined elsewhere... (let (describe-future version-future) (with-batch-calls (setf describe-future (server-describe-method sock "version") version-future (server-version sock))) (values (wait description-future) ;; This will return immediately (wait version-future)))
Any method calls made within a
with-batch-calls macro won't be submitted to the server until the end of the block. Note: since none of the calls in a
with-batch-calls block will be submitted until the block is left, calling
wait on one of the returned futures will deadlock.
The non-notification functions created with
define-rpc-method return a future object that will hold the return value of the method or signal an error if there was one.
wait can be used to retrieve a completion value or signal an error: it takes a future object, and returns when that future completes.
Right now the underlying IO mechanism isn't really asynchronous.
wait drives the IO processing, so if a lot of requests are made and
wait is never called, the socket buffer could fill and cause the server to block. This shouldn't be an issue under most workloads, however.
- Use an asynch IO mechanism to avoid the need to call
- Add timeouts and keepalive options for rpc methods.
- Accept strings as well as symbols for method and service names.
- Use typespec values from the server to do typechecking where possible.
- This could be a performance issue, so this may need to be a disable-able feature.
- Allow clients to specify non-mangled names?
- Document signaled conditions once they're all in place.
- Give the client some control over the encoding of arguments passed to methods.
- Either allow client to supply pre-encoded values (potential badness here), or supply encoder functions.
- Matthew Stickney <email@example.com>