Up

module Uq_mt

: sig

Using engines in multi-threaded programs

Monitors can be used to coordinate access to a resource from several threads so that only one thread comes into contact with the resource. Of course, the resources we talk about must have an engine interface. For example, the resource could be an RPC client. In contrast to a critical section (protected with mutexes), the accesses are not serialized, but simply pushed onto the same event system, and run there concurrently. Effectively, the monitor translates heavy-weight threading (kernel threads) into light-weight threading (engines).

#
type monitor

A thread monitor governs which threads have access to a set of engines running together on an event system

#
val create_monitor : Unixqueue.event_system -> monitor

Creates a new monitor. The passed event system becomes the inner event system of the monitor. All engines attached to the monitor will use this event system.

#
val monitor_run : monitor -> (Unixqueue.event_system -> 'a -> 'b Uq_engines.engine) -> 'a -> 'b

let result = monitor_run mon f outer_esys arg: Runs another engine within the monitor. The engine inner_e is created as

let inner_e = f inner_esys arg

where inner_esys is the event system of the monitor. It is obligatory that inner_e uses this event system.

When inner_e reaches a final state, the thread leaves the monitor, and the result of inner_e is returned (or the recorded exception is thrown).

The "special effect" of the monitor is that several threads can share the same monitor, even at the same time. All parallel running engines inner_e will use the same event system inner_esys. Of course, this also means that they run in the same thread. Note that it is possible that the monitor switches the thread it uses for executing the event system (and the engines). In particular, this can happen when one of the inner engines finishes, and the thread that put this engine into the monitor was chosen as the thread for executing the event system. When this happens, another thread waiting for the monitor will be selected for running the event system.

If an inner engine establishes configurations that are not bound to the lifetime of the inner engine, there is no guarantee that these additional events and handlers will be considered when the inner engine ends. For example, if the engine starts a timer to run an action later, it is not said that the action is carried out at all if the engine terminates before.

Exceptions falling through to the caller of the event system are logged and otherwise ignored.

It is possible to use monitors in single-threaded programs - this is treated as if only one thread was using the monitor in a multi-threaded program.

#
val monitor_do : monitor -> ('a -> 'b) -> 'a -> 'b

let result = monitor_do mon f arg: Just runs f arg in the scope of the monitor, and returns the result.

#
val monitor_async : monitor -> ('a -> ((unit -> 'b) -> unit) -> unit) -> 'a -> 'b

let result = monitor_async f arg: For calling RPC-style asynchronous clients. If f is called like f arg emit, the result is passed back by calling emit (fun () -> result).

#

Example: Threads sharing an RPC client

In this example we want to achieve that several threads can use the same RPC client. This type of client is not thread-safe, but it is able to run asynchronously. Because of this, the requests coming from several threads can be concurrently handled.

Prepare the RPC client and the monitor: We assume that M_clnt is the generated RPC module with the client, and that we access program P at version V.

      let esys = Unixqueue.create_unix_event_system
      let rpc = M_clnt.P.V.create_client ~esys connector protocol
      let mon = Uq_mt.create_monitor esys

It is essential that the RPC client and the monitor use the same esys.

Next, let's assume we are in a spawned thread, and we want to call an RPC procedure M_clnt.P.V.proc. The trick here is to use the asynchronous variant M_clnt.P.V.proc'async, and to call it via the monitor:

      let result = Uq_mt.monitor_async mon (M_clnt.P.V.proc'async rpc) arg

The procedure is called with argument arg and returns the result. The monitor waits until the asynchronous call is done, so this appears as a synchronous call to the user.

If you need to call functions to configure and control the RPC client, use monitor_do:

      let sockaddr = Uq_mt.monitor_do mon Rpc_client.get_peer_name rpc
end