module Unixqueue

: sig

Unixqueues are one of the two forms of system event loops provided by Ocamlnet. Besides Unixqueue, there is also pollset (see [root:Netsys_pollset]). The pollsets are much simpler (there is no queuing of events), and nowadays Unixqueue bases upon pollset, and extends its functionality. Historically, however, Unixqueue precede pollset, and there are still implementations of the former in Ocamlnet not using pollset as its base data structure.

The common idea of both data structures is the generalization of watching for events, as it is also provided by the Unix.select function. Note, however, that recent implementations no longer use Unix.select, but better system interfaces for the same.

When there is something to do for a file descriptor (reading, writing, accepting out-of-band data), this is called an event, and the task of Unixqueue is to check when events happen, and to tell some consumer about the events.

There are three further types of events: Timeout events, signal events, and user-defined events.

The events are queued up, and they are presented to event handlers that may process them.

You can describe what types of event conditions are watched by adding resources. You can think a resource being a condition (bound to a real resource of the operating system) for which events are generated if the condition becomes true. Currently, only file descriptors and timers are supported as resources.

Relation to other modules. This module is thought as the primary interface to Unixqueues. If there isn't any specialty one has to deal with, just use this module:

  • It defines all required types like group, wait_id, etc. Note that these types are reexported from Unixqueue_util. Please consider this as implementation detail, and don't use it in your code.
  • It defines a standard implementation standard_event_system, which is a good default implementation, although it might not be the best available for all purposes.
  • It defines a set of access functions like add_event which simply call the methods of the event system object of the same name. Note that these functions work for all event system implementation, not only for standard_event_system.

There are further modules that have to do with Unixqueue:

  • [root:Unixqueue_pollset] is the implementation behind standard_event_system. If you want to use other pollsets than the standard one, it is possible to create Unixqueues on top of these by using this module directly.
  • [root:Unixqueue_select] is the historic default implementation. It calls directly Unix.select. It is still available because it serves as a reference implementation for now.
  • Unixqueue_util is an internal module with implementation details. Please don't call it directly.
  • [root:Uq_gtk] is an implementation of Unixqueue mapping to the GTK event loop. Useful for multiplexing event-based I/O and GTK graphics operations.
  • [root:Uq_tcl] is an implementation of Unixqueue mapping to the TCL event loop. Useful for multiplexing event-based I/O and event-based code written in TCL (especially TK).

Thread safety. The default implementation of Unixqueue is thread-safe, and operations can be called from different threads. For other implementations, please look at the modules implementing them.

Types and exceptions

type group = Unixqueue_util.group

A group is an abstract tag for a set of events, resources, and event handlers. Usually every event handler creates a new group, and all events and resources processed by the handler are members of this group.

exception Abort of (group * exn)

Event handlers can raise this exception to cancel a group of handlers, events, and resources. If an abort action is defined for the group, it will be executed. Next, all members of the group are removed from the event system.

First argument is the group. The second argument is an arbitrary exception (must not be Abort again) which is passed to the abort action.

Abort handlers are a questionable feature of Unixqueues. You can also call the clear operation, and raise the exception directly. Do not use in new code!

type wait_id = Unixqueue_util.wait_id

A wait identifier is used to distinguish between several timers, see type operation.

type operation = Unixqueue_util.operation =
# | Wait_in of Unix.file_descr
(*wait for input data*)
# | Wait_out of Unix.file_descr
(*wait until output can be written*)
# | Wait_oob of Unix.file_descr
(*wait for out-of-band data*)
# | Wait of wait_id
(*wait only for timeout*)

An operation specifies the condition to wait for. Every kind of operation may have an associated timer (not only Wait).

type event = Unixqueue_util.event =
# | Input_arrived of (group * Unix.file_descr)
(*Input data has arrived*)
# | Output_readiness of (group * Unix.file_descr)
(*Output is possible now*)
# | Out_of_band of (group * Unix.file_descr)
(*OOB data has arrived*)
# | Timeout of (group * operation)
(*A timer has expired*)
# | Signal
(*A signal has happened*)
# | Extra of exn
(*User-generated event*)
# | Immediate of (group * (unit -> unit))
(*immediate event*)

An event is triggered when the condition of an operation becomes true, when a signal happens, or when the event is (artificially) added to the event queue (add_event, below). The events resulting from an operation carry the group of the resource with them.

The event Signal is triggered when the EINTR condition is caught; this normally means that a signal has just been delivered. The generation of Signal events should be considered as unreliable, not every signal delivery can be detected. Reasons for the unrealiability are that user-supplied code happens to get the EINTR condition and not the Unixqueue event loop, and that there are known race conditions in the O'Caml signal handling routines that may cause signals to be lost. However, it can be expected that almost all signals will trigger Signal.

The event Extra can only be artificially added to the queue, and the argument of Extra is an exception value that distinguishes between several kinds of user-generated events.

The event Immediate(g,f) also can only be artificially added to the queue. In contrast to other events, it is not passed to handlers when the event is processed. Instead, an immediate event is processed by calling f(). This is a more direct way of notification, and it is not necessary to define a handler. Even an immediate event is member of a group g, and if the clear function is called for g, the callback function f will no longer be called.

class type event_system =
method new_group : unit -> group
method new_wait_id : unit -> wait_id
method exists_resource : operation -> bool
method add_resource : group -> operation * float -> unit
method add_weak_resource : group -> operation * float -> unit
method add_close_action : group -> Unix.file_descr * (Unix.file_descr -> unit) -> unit
method add_abort_action : group -> (group -> exn -> unit) -> unit
method remove_resource : group -> operation -> unit
method add_handler : group -> (event_system -> event Equeue.t -> event -> unit) -> unit
method add_event : event -> unit
method clear : group -> unit
method run : unit -> unit
method is_running : bool
method when_blocking : (unit -> unit) -> unit

The event_system manages events, handlers, resources, groups, etc. It is now a class type, and you may invoke the operations directly for the class. The operations are still available as functions (below).

A resource is an operation with an optional timer. The operation describes the condition to watch for, and the timer defines the maximum period of time for that. If the condition becomes true, an Input_arrived, Output_readiness, or Out_of_band event will be triggered. If the timer expires, a Timeout event will be generated. After the event the resource remains active, and the timeout period begins anew.

A resource is usually bound to a file descriptor. It is allowed to watch the same descriptor for several different conditions, but it is forbidden to watch the same descriptor for the same kind of condition several times.

As a special case, the operation Wait is not bound to a file descriptor, but simply starts a timer. The argument of Wait can be used to distinguish between several timers that are active at the same time.

Event handlers get the events one after the other, and process them. When a handler is called for an event, there are several possible reactions: (1) The handler can return normally, which means that the event has been accepted, and will not be passed to any other handler. (2) The handler can raise Equeue.Reject, which means that the handler cannot process the event, and that another handler should get it. (3) The handler can raise Equeue.Terminate which means that the event has been accepted, and that the handler is terminated (it will never be called again). (4) The handler can raise Abort which means that the event is deferred, and that a special abort mechanism is triggered (see the description for Abort above), this is also terminates the handler. The deferred event will again be processed in the future. (5) The handler can raise any other exception. This causes that the event is deferred, and the exception falls through to the caller of run.

Groups are used to simplify the association of events to handlers, and to simplify the termination of handlers (see clear). If an event is associated with a group, only handlers associated with the same group will get them.

There is a special Close handler which is useful to close file descriptors no longer needed. It is called when all resources are removed from the event system dealing with the file descriptor. The close handler should close the descriptor. Note that close handlers are only useful under certain circumstances.

Creating event systems

class standard_event_system : unit -> event_system

The standard implementation of an event system. It uses Unixqueue_pollset.pollset_event_system on top of Netsys_pollset_generic.standard_pollset.

val standard_event_system : unit -> event_system

Create a new, empty, standard event system

class unix_event_system : unit -> event_system

An alternate name for standard_event_system, provided for backward compatibility.

val create_unix_event_system : unit -> event_system

An alternate name for standard_event_system, provided for backward compatibility.

class performance_event_system : unit -> event_system

The implementation using Netsys_pollset_generic.performance_pollset.

val performance_event_system : unit -> event_system

Same as function

Using event systems

The following functions work for all kinds of event systems, not only for the ones returned by standard_event_system.

val new_group : event_system -> group

Create a new, empty group for the event system

val new_wait_id : event_system -> wait_id

Create a new unique wait identifier

val exists_resource : event_system -> operation -> bool

Find out if a specific resource already exists (or better: is already watched by an operation).

val add_resource : event_system -> group -> operation * float -> unit

Add a resource such that it is watched for conditions described by the operation for the period given by the float number. A negative number means that the resource is watched for an infinite period. The resource becomes a member of the group.

You cannot add the same operation several times; if you try it the second operation is silently dropped.

The resource remains even if it has generated an event. The timeout period starts again in this case.

val add_weak_resource : event_system -> group -> operation * float -> unit

Similar to add_resource, but the resource is weak. Such resources do not keep the event system running when only weak resources remain. Normally, Unixqueue.run returns to the caller not before all resources are removed and all events are processed. Weak resources do not count for this condition, i.e. Unixqueue.run also returns when there are only weak resources left. As an example, weak resources can be used to time out unused file descriptors.

Weak resources can be removed with remove_resource.

New in Ocamlnet 3.

val add_close_action : event_system -> group -> Unix.file_descr * (Unix.file_descr -> unit) -> unit

A close action is added for the file descriptor. The action callback (which gets the descriptor as argument) is called when there is not any watched resource remaining for this descriptor.

This may be useful if the descriptor can be closed in this case.

The close action becomes member of the passed group. The only effect of this is that the action is removed when the clear function is called.

You can only add (set) one close action for every descriptor.

Of course, the idea is to do add_close_action ... Unix.close. Note that there is a problem with multi-threaded programs, and this construct must not be used there. In particular, the close action is called from remove_resource or clear, but it is possible that the event system is running, so a watched descriptor might be closed. This has undesired effects. What you should better do is to delay the closure of the descriptor to a sane moment, e.g. by calling

Unixqueue.once esys g 0.0 (fun () -> Unix.close fd)

from the close action.

val add_abort_action : event_system -> group -> (group -> exn -> unit) -> unit

An abort action is added to the group. The action callback is called when an arbitrary handler raises Abort(g,exn) where g is the group the abort action is member of. In this case, the callback function is invoked with the group and exn as arguments. After that, the group is cleared.

You can only add (set) one abort action for every group.

val remove_resource : event_system -> group -> operation -> unit

Removes the operation from the watch list of the group. It is an error if the operation is member of another group. If the operation cannot be found at all, the exception Not_found will be raised.

The removal of resources may trigger close actions.

val add_handler : event_system -> group -> (event_system -> event Equeue.t -> event -> unit) -> unit

Add an event handler that is associated to the given group. There may be several handlers for a group.

The handler callback function is invoked when there is an event that could be processeable by the handler. As outlined above, the callback function can accept or reject the event, it can terminate itself, and it can abort the whole group.

val add_event : event_system -> event -> unit

Add an additional event. The event will be processed after the current list of events is done.

val clear : event_system -> group -> unit

Terminate the whole group. This means that the handlers of the group are not called any longer, and that all resources and actions are removed. It is possible that there are pending events after termination, but these will be usually be dropped because there is no handler for them.

When a group is terminated, it is not allowed to refer to the group any longer. Functions will raise Invalid_argument if this is tried nevertheless.

val run : event_system -> unit

Starts the event loop. This means that the resources are watched, and that events are generated, and that handlers are called.

The event loop returns normally when there are not any resources and not any events in the queue. The loop raises Equeue.Out_of_handlers if there are resources but no handlers to process their events. It is possible that exceptions raised from handlers fall through to the run call.

After the exception is caught and processed, the event loop can be restarted.

val is_running : event_system -> bool

Whether the event loop is running

val once : event_system -> group -> float -> (unit -> unit) -> unit

Arranges that the callback function is called once after the passed period of time (the float argument) has elapsed.

The arrangement is member of the passed group. By clearing the group, the timer is deleted, too.

val weak_once : event_system -> group -> float -> (unit -> unit) -> unit

Same as once, but the timer does not keep the event system running if it is the only remaining resource.

val epsilon : event_system -> (unit -> unit) -> unit

The execution of the function is pushed onto the event queue (minimal delay)


module Debug : sig

This module controls debugging of all Unixqueue_* modules

val enable : bool Pervasives.ref

Enable [root:Netlog] debugging

val set_debug_mode : bool -> unit

Sets enable.

val set_debug_target : Equeue.Debug.debug_target -> unit

Restricts debugging to this target.