TIP 175: Add an -async Option to [open]

Login
Bounty program for improvements to Tcl and certain Tcl packages.
Tcl 2017 Conference, Houston/TX, US, Oct 16-20
Send your abstracts to tclconference@googlegroups.com
by Aug 21.
Author:         Neil Madden <nem@cs.nott.ac.uk>
State:          Withdrawn
Type:           Project
Vote:           Pending
Created:        15-Mar-2004
Post-History:   
Tcl-Version:    9.0

Abstract

This TIP propose to add an -async option to the [open] command, with identical semantics to the -async option to the [socket] command.

Rationale

With the introduction of Virtual File Systems (VFS, see [17]), it is now possible to use [open] to access resources which are not available in the local file system. However, initial access to these resources may take some time (for instance, in an HTTP VFS, this requires an HTTP GET request to a remote server). Currently, while this request is being processed, the Tcl application will block waiting for a response. It may take several seconds before the open call returns, and control is passed back to the application. Delays of this length can cause problems, especially in applications with a Tk graphical user interface, which will not repond to events until the call has completed. Tcl provides mechanisms to avoid this initial blocking for the socket command, and in the HTTP package. However, this functionality is missing from the open command. For VFS to become useful for writing network code, this functionality needs to be added.

Proposed Change

This TIP proposes adding an -async option to the open command. The new syntax for the open command would be:

 open ?-async? ?--? name ?access? ?permissions?

The "--" marker is also proposed to signify "end of options". This is needed as name can be anything, including the string "-async", and so some mechanism is needed to distinguish between the -async option, and the name. This is in line with other commands which take multiple options.

When the -async option is present, the open command will return immediately, without waiting for the underlying channel to be created and connected. If the channel is configured for blocking mode, a gets or flush that is done on the channel before the connection is made will wait until the connection is completed, or fails. If the channel is configured for nonblocking mode, then a gets or flush on the channel before the connection has been made will return immediately and fblocked on the channel will return 1. This is exactly the same behaviour of the current implementation of the -async option to the socket command.

In addition, each VFS should support the -error option to fconfigure to allow scripts to determine the cause of an error when asynchronous opening fails. Furthermore, there needs to be a mechanism for notifying a script when an asynchronous open fails. This could be achieved by firing any fileevents (readable or writeable) which are registered on the resulting channel. This is the behaviour of [socket -async].

In order for this TIP to be implemented, there would have to be changes to the VFS layer in Tcl. There are at three ways to accomplish this:

  1. Extend Tcl_FSOpenFileChannelProc to accept an async argument, of type int.

  2. Extend the Tcl_Filesystem structure to add a new callback, alongside T_FSOFCP. For instance, Tcl_FSOpenFileChannelAsyncProc.

  3. (Ab)use the mode parameter, and add a mask such as TCL_ASYNC.

Of these, option 3 is the easiest in terms of maintaining compatibility, but is not the nicest implementation.

Reference Implementation

A reference implementation does not yet exist. The TIP author would be willing to implement any changes needed, but currently is not familiar with the internal workings of the VFS system.

Questions

Should the -async option be supported when a command pipeline is being created?

Should a more general option passing mechanism be implemented? It is possible that some VFS implementations may require more information than is currently available in order to correctly create a channel (the -myaddr and -myport options to socket spring to mind). Would it be more sensible to consider a proposal for a more general option passing mechanism, to avoid further updates to the API in the future? This TIP does not propose such a mechanism.

Copyright

This document has been placed in the public domain.

History