Programmable Widgets and XSAR
Since we will be creating new display widgets for our project (e.g, the form and list widgets), it might be worthwhile to consider some alternative approaches in the design of the XSAR widgets. I had some thoughts on the subject that might prove useful. (Jeff: "That's a change - on both counts!") This memo suggests that we investigate the possibility of making our widgets user-programmable, a capability that would have the following benefits:
The TPOCC display subsystem is flexible in many ways and rigid in others. Ironically, the very mechanisms that give the TPOCC display system its flexibility - the resource database and UIL page definitions - are also the source of the display system's inflexibilities. With resource and UIL files, you can define the look of a display page, but you can't define its behavior beyond what is currently hard-wired in the widget's C code. This (i) limits the usefulness of a given widget with respect to the range of applications it supports and (ii) increases maintenance costs, since enhancing or correcting the functionality of a widget necessitates digging into the internals of the widget's code. The limited versatility of a widget with hard-coded behavior could pose a problem for XSAR. In particular, can we design a generic form widget that will support all of the forms planned for the XSAR operator interface, with the layout and function of each form specified solely through resource database entries and UIL page definitions? Or, must we design a whole slew of specialized widgets (or programs)?
With everyone and his/her brother working on the XSAR display system, maintaining hard-coded widgets would not seem to be a problem, at least not for manpower reasons. The problems that could occur in general might best be illustrated by example.
GSMS is a spacecraft monitoring system that is not part of a mission's control center; it is a separate entity, running on a separate computer, and connected to the POCC via the network. GSMS needs to display and access system variables of two types: its own GSMS-specific system variables (local to the GSMS computer) and the target mission's system variables (found on the POCC's computers). Consequently, GSMS must talk to two data servers: the local GSMS data server and the remote mission's data server, each of which has a different mission name and, thus, a different server name.
TPOCC does not currently support GSMS's multi-mission mode of operation. A
proposal to add mission-level identification of system variables was
rejected, probably because of the changes that would have been required in
a number of TPOCC subsystems. For the display system, the declaration and
handling of a new
mission resource would have to be added to
most of the TPOCC widgets, the TPOCC WML definitions and UIL compiler would
have to be updated to reflect the new resource, and existing resource files
and (possibly) UIL files would have to be changed.
Aside: Actually, the problem above is not so much a problem with hard-wired widgets as it is a problem of resources being specified at too low a level. Replacing the current
offsetresources by a single
variableresource would have eliminated the multi-mission problem as far as Display is concerned. The
variableresource would specify the desired system variable, perhaps in a format similar to that used by TSTOL and the TPOCC configuration monitor:mnemonic#process@host[offset:count]
The mnemonic field is mandatory; the other fields are optional, including the process name for unique mnemonics. The data services library would extract the host name and forward the data request to the targeted data server; the data server would parse and provide defaults for the remaining fields in the variable specification. Adding mission-level identification of system variables could be accomplished by allowing an optional "mission:" field preceding the mnemonic name. The mission field would be processed by the data services library - no changes to the display subsystem (either in the code or in the data files) would be needed.
Some might argue that the X resource database provides defaults for unspecified values. This is true, but the defaults for some fields (e.g., process names and data types) are more sensibly supplied by the data services subsystem.
Speaking of resources, a redesign of the data services interface would eliminate the need to specify data types and type sizes in page definitions, a peculiar requirement for a user-friendly interface. And the data rate, currently expressed in tenths of seconds, ought to be specified as a user-friendly real number representing the seconds and/or fractions of seconds between data points - DEC's new Alpha-based workstations will make the data server's 1/10-second timing granularity obsolete!
As Mike mentioned in one of our meetings, this problem with simultaneously processing and displaying data from multiple missions affects the XSAR design. Our requirement to support the decommutation and display of real-time and playback data at the same time is effectively a matter of handling two missions: real-time and playback.
We uncovered another problem at one of our recent TS/CS meetings: a UIL-based display system doesn't respond easily to database and configuration updates. The old TPOCC page description language (TDL) allowed resource values to be retrieved - at run-time - from the Unify database or from environment variables, a capability that XSAR would find useful. For example, when you bring up a page, a generic, UIL-specified page could be read in and particular resource values retrieved from our "database": the ADT and the GDT.
In the old TDL, a resource value could be expressed as a literal string, as
%dbms (table, key, field)", or as
%env (variable)". Adding a similar capability to the
current TPOCC display program would require changing all the resources of
interest to strings and modifying the widget creation functions to check
for and resolve dynamic references in these resource values. (Perhaps you
could just specify the resources as strings in UIL and let the resource
converters perform the value resolution.) It might be argued that doing
this subverts the validity checks of the UIL compiler, but these checks are
solely syntactic in nature; the semantic checks have always been left to
the data server.
As briefly discussed in the introduction, user-programmable widgets could benefit the XSAR project in the following ways:
Making a widget user-programmable requires that an extensible, command language interpreter (or two or more) be embedded in the display system. The benefits of an extensible command language are evident in TPOCC's TSTOL interpreter. TSTOL implements a generic set of STOL directives; mission-specific directives are defined and programmed using TSTOL itself. TSTOL's programmability allows users to extend the lexical structure, syntax, and semantics of the language - without changing the executable program.
The TAE Command
Language (TCL) is similar to TSTOL. One feature of TCL is that it gives the
user access, at the TCL level, to TAE's
Wpt functions for creating
and controlling display widgets. This allows you, for instance, to test new
or updated widgets without having to write special test programs. Neither
TSTOL nor TCL are embeddable languages and the requirements that guided their
design make them unlikely candidates for that role.
Embedded command languages have proven their utility in a number of interactive applications. The Emacs editor uses a LISP-like internal language; the VAX/VMS EVE editor is programmed using the Pascal-like TPU language. Cadre's teamwork CASE tool allows users to define their own menus using the Teamwork Extensibility Language (TEL - my abbreviation). An arbitrarily-complex TEL expression can be assigned to a menu item; when that item is selected, its action expression is evaluated. With TEL, you can pop up a dialog box to query the user for more information, iteratively apply an action to multiple selections, store and retrieve variables, execute shell (DCL) commands, and so on. Although he found the language itself awkward, Tom Nicinski used TEL to couple two stand-alone applications: at any level and location in a dataflow diagram (built by Teamwork), one could switch to editing the relevant architecture-flow diagram (built by xfig).
Programmers developing X Windows-based, display applications have not been left out in the cold; they have several languages to choose from. One such language is WINTERP (Widget INTERPreter), developed by Niels Mayer at HP's Human-Computer Interaction Laboratory. WINTERP is built on top of XLISP, a public domain, object-oriented version of LISP. WINTERP is not really an embedded language; the WINTERP interpreter runs as a network server process, separate from its client application(s). WINTERP's advantages: it's free (as part of the MIT X11 distribution) and it knows about Motif. WINTERP's disadvantages: it's programmed in XLISP (a problem for those not familiar with LISP), it runs in a separate process, and it looks like it would be difficult to port to VMS.
Aside: Running the command language interpreter in a separate process might not be such a bad idea. For example, the Wafe (Widget [Athena] Front End) package provides windowing capabilities to programs written in any language that supports standard I/O: shell scripts, awk(1) scripts, C, FORTRAN, Prolog, etc. The Wafe program runs as a co-process of the target application, accepting Tcl commands to create and interact with Athena widgets:network I/O X Windows <-------> Wafe <--------> Application standard I/O
(Tcl is an embedded command language described later.) Wafe spawns the application program; the two processes communicate through the application's standard input and output channels (
SYS$OUTPUT). The Wafe scheme is advantageous in several ways. First, a Wafe-like interface makes an application independent of the windowing system; you could conceivably have a Motif Wafe, an Open Look Wafe, a Windows Wafe, etc. Second, Wafe allows an application to be coded in the most suitable programming language; you don't have to wait for (or write your own) awk or Prolog bindings to the X libraries. Finally, an application can download Tcl procedures to the Wafe "server", thus offloading display-specific processing; sounds like News, doesn't it?!
IXI's Deskterm product provides existing, character-based applications with graphical, Motif interfaces. No changes need to be made in the applications; you don't even need their source code! I've only seen it in advertisements, but I suspect that Deskterm is a Wafe-like interface process that you "program" (in some command or scripting language) to recognize and respond to the output of the target application. For example, you could instruct Deskterm to update field A when it sees the cursor-positioning sequence for field A. Or, if the FORMAT option on the character-based "screen" is highlighted, Deskterm should pop-up the FORMAT menu. Operator input can likewise be transformed into what the character-based application expects. The techniques used by IXI are not new: graphical interfaces to dumb-terminal, on-line services like Genie have been implemented on a variety of personal computers and many terminal programs use scripts to control the exchange of information through modems.
Unlike Wafe, AT&T's Plan 9 windowing system, 8½, skips the intermediate process altogether and connects an application's standard I/O channels directly to the display server. Window-related events and commands are embedded in the I/O stream. A program can use or not use the windowing system capabilities available to it. Some of the programs for my Commodore 128 ($250 a few years ago) operate in a similar fashion. If the program sees a mouse event (e.g., the user moves the mouse), then features such as pull-down menus are enabled; otherwise, the program simply acts in "dumb terminal" mode.
Back in the early days of TPOCC, some thought was given to running the TPOCC display system on the real-time, front-end computer. Unfortunately, the various real-time operating systems that were evaluated at that time did not support X client libraries. Imagine, however, running the display program as a real-time process on the front end, talking to a Wafe or WINTERP server on the workstation ...
The Widget Creation Library (Wcl) is another language that was praised in a USENET posting. Called by one user "a poor man's UIL", Wcl allows you "to specify all the static properties ... and some of the dynamics" of a display interface via the X resource database. Wcl is a library layered over the X Toolkit. Its ability to control the behavior of an application is limited to specifying a string to be passed to callback routines when they are invoked. Wcl's strong points are that it has been ported to VMS and that it allows you to write display programs that are, to a large extent, independent of the widget set being used (e.g., Motif or Open Look).
A number of other embedded command languages are available, most of which are based on some variant of LISP. One language which is not is the Tool Command Language (Tcl). Tcl (no relation to TAE's TCL) can be characterized as follows:
The power of Tcl is evident from the fact that a complete X11 toolkit, Tk, was written based on Tcl. Tk provides most of the capabilities of the standard X11 toolkit, Xt. Although we won't be using the Tk toolkit, its use of Tcl illustrates how Tcl could be used in a programmable GUI:
object_type pathname [attributes]Executing that Tcl command creates a new instance of the widget, identified by its pathname (in standard Xt dot notation).
object method arg1 arg2 ...Some typical methods include
configure(set or change a widget's resources),
flash("blink" the widget a few times),
invoke(execute the Tcl command assigned to this object).
.allow invoke" command to the
/ALLOWbutton on the critical command verification panel; the F12 key could send a "
.cancel invoke" command to the
/CANCELbutton. (The buttons would, in turn, execute their commands to send the appropriate directives to TSTOL.)
sendcommand is defined which allows an application to easily exchange information with or control other applications. Since
sendis implemented using X11 window properties, the participating applications must all be X11-based and have embedded Tcl interpreters.
Tk does not attempt to implement the inheritance mechanisms favored by other toolkits; instead, composition is the preferred method for reusing code in new widget types. As the authors of the C++-based, InterViews toolkit say,
"A rich set of primitive and composition objects promotes flexibility, while composition itself represents a powerful way to specify sophisticated and diverse interfaces."
In a sense, composition is a form of delegation, considered by some to be a superior object-oriented programming paradigm. Rather than inheriting capabilities from a parent class, an object delegates responsibilities it can't fulfill to other objects.
Currently, the TPOCC widgets comprise a forest of widget type hierarchies, because different widgets are descended from different Motif "seeds":
XmLabel XmPrimitive XmScrolledWindow | | | XtpRealTimeLabel XtpRTPrim XtpScrollingText | | | XtpAscii, XtpLGraph XtpRealTimeScrollingText XtpNumeric, | | ... XtpXTGraph, XtpCpage, XtpXYGraph XtpEventWindow
By combining inheritance with delegation, a single inheritance hierarchy could conceivably be created for the TPOCC widgets (delegation is indicated by horizontal branches):
TPOCC Core Widget Class / \ Motif Class(es) -- Form List -- Motif Class(es)
Structuring the widget classes in this fashion would have a number of advantages. First, common characteristics of the widget classes could be collected in abstract classes, a technique commonly used in languages that support multiple inheritance. For example, the data communications aspects (those dealing with the data server) of the widgets could be isolated in a separate class. Then, a real-time widget that needs to request data from the data server could delegate that function to a data communication object.
Mike recently discovered another advantage of having a single inheritance hierarchy. I'm not familiar with all the details; I wasn't following the discussion closely and, once Fred got involved, I couldn't hear the forest for the trees! From what I understand, Mike wanted to add a flag to control the update of a display object, but, since the object was a Motif widget, that was out of the question. Wrapping a TPOCC "core" widget around the Motif widget gives us a hook upon which to hang new resources like the update flag. The flag could be stored in the callback code or, if Tcl commands can be assigned to the TPOCC core widget, in a Tcl variable.
The remainder of this memo discusses how we might use an embedded command language like Tcl in XSAR's form and list widgets. Before continuing, it should be pointed out that using an embedded command language does not preclude the use of the X resource database or of UIL page definitions. The Tcl commands associated with a widget would be defined as resources of the widgets, specified in the user's resource file or in the display's UIL files.
A generic list widget is particularly interesting because it could be used in many different ways on XSAR: to view reports stored in files, to monitor scrolling real-time data (e.g., events), to display tables of information, etc. List widgets could be used both to select and display report files. For example, John Ousterhout, in his paper on the Tk toolkit, presents a directory/file browser implemented by a 16-line Tcl script:
examineprocedure to a function key.
Short and sweet! Seen in action, the browse script brings up a window
containing a list of the files in your current directory. Scrolling up and
down the list, you can select the file you want to examine. When the
defined function key is hit, the
examine procedure is
automatically called. If the selected file is a directory,
examine recursively invokes the browse script to bring up the
subdirectory's listing in another window. If the selected file is an
examine spawns a process to edit the file.
The events/alarms subsystem (whether MPOS or TPOCC, I'm not sure) will be the data source for many of XSAR's displays; e.g., the Telemetry Validation "Table". The list widget is a natural for any events-based display page. Assume the generic XSAR list widget has the following methods (callbacks), each with an associated Tcl command:
initialize- initialize the list object.
select- process the selected list item.
popup- popup another display object.
insert/append- insert or add an item to the list.
update- process a real-time data update.
A method, when it is called, executes the Tcl command assigned to it, as
well as any internal processing that needs to be done. Some commands will
be provided with additional information; for example,
needs to pass the incoming data value to its Tcl command.
A general-purpose, XSAR events/alarms page could then be implemented using the generic list widget and the following Tcl command assignments:
initialize- execute a Tcl script to:
popup- pop up the class selection "menu".
insert/append- add a line of text to the list.
appendto add the incoming event/alarm message to the scrolling list of messages.
To change the class of events and alarms being viewed, the operator would click on the right mouse button and a class selection menu would pop up on the screen. This popup could also be implemented as a list, with the following Tcl definitions:
initialize- execute a Tcl script to:
select- highlight the current item and call
popup. (A real implementation would support the selection of multiple items.)
popup- display an APPLY/CANCEL dialog window. APPLY effects the class change; CANCEL leaves everything as is.
insert/append- add a line of text to the list.
Actually changing the event/alarm classes could be accomplished in either of two ways:
updatemethod's Tcl command filter the incoming messages.
Note that the latter approach could be used to give any dynamic, scrolling list display an input filtering capability - without changing the generic list widget's code. This might come in handy if, for example, we are forced to read messages from the MPOS Events/Alarms server and it doesn't support message filtering on its end.
Programming widgets in Tcl doesn't eliminate the need to write C code, but it does provide a clear separation between the display aspects of the widget and the functional aspects. New Tcl commands (to initiate data and events streams, for instance) must be implemented in C and "registered" with the Tcl interpreter. With some forethought, these C routines could be general purpose functions usable by other, non-display subsystems.
The XSAR display system will include a number of tabular displays, which, not surprisingly, could be handled by a generic list widget that allows an optional, non-scrolling header at the top of its list. The header could be specified as a single, text-string resource with embedded new-lines (if necessary); the widget would break the string into separate lines and determine how many non-scrolling lines (0 to N) to allocate for the header.
Despite the fact that a list is a list, a list object need not be limited to sequential updates. For example, TPOCC's command page (CPAGE) is essentially a random-access list of the current contents of the command buffer; each CPAGE update received by the display program specifies the associated command's position in the displayed list. If XSAR's generic list widget employed a similar scheme, it could even be used to implement screens such as the Telemetry Archive Log Display, a 4-line list of the 4 different kinds of major frame log. (I don't recommend this approach for this particular display, since it moves some of the display functionality [the formatting of the lines] into the Telemetry History Manager.)
Aside: As of Release 6, TPOCC did not have a generic list widget. Consequently, people had to resort to using the CPAGE for non-command, list-oriented displays. The XSAR generic list widget could be used for these displays, as well as for the command and events pages. The somewhat redundant CPAGE and EVENT data objects received from the data and events servers could be replaced by a single "list string" containing the following items of information:
- position information (e.g., CPAGE line numbers),
- display attributes (e.g., to provide for displaying events of different severities in different colors), and
- the text to be displayed.
The position information and display attributes should be encapsulated so as to allow for future additions without affecting existing software. With some effort, control sequences embedded in the string text could be recognized and used to change display attributes in mid-string.
After giving the matter some thought, I'm not really sure what a generic
form widget would look like, or even if one is needed. Presumably, it
would be some kind of container widget, a subclass of the Motif
XmManager metaclass (itself a subclass of the Xt Composite
metaclass). The generic form widget might support the following methods,
each of which has a Tcl command resource:
initialize- initialize the form.
process- process the current contents of the form.
terminate- destroy the form.
initialize method is responsible for laying out and
configuring the display and input objects of which the form is composed.
The layout of the form could be specified in UIL or programmed in a Tcl
script invoked by the method's Tcl command. Regardless of how the form is
initialize's Tcl procedure could also perform such
non-display functions as requesting command controller status for command
terminate method's Tcl command is executed when the
operator exits the form. This command performs any wrap-up activities
required for the form; e.g., relinquishing command controller status or
saving current field values (to be recalled when the form is displayed
process method is responsible for collecting and
processing the values entered on the form. This can all be done by a Tcl
procedure assigned to this method. If the form is built from TPOCC/XSAR
input objects, individual field values could be stored by the Tcl commands
bound to those objects; the form's
process method would only
be called when the operator has completed the form (e.g., has hit the SEND
button on an XSAR command processing panel). Motif input objects have no
Tcl command execution capability, so the activation callback defined for
such an object could pass the field name and value to the form's
process method. The
process method would then
have to remember the different fields' values until the form is complete,
at which time the processing of the form as a whole could take place. (Or,
now that I think about it, the Motif object's callback could execute
a Tcl command; it's just that the Motif widgets have no explicit Tcl
It is interesting and, perhaps, instructive to see how a form is created - entirely from a Tcl script - using the Tk X Windows toolkit. For example, the XSAR "Change Telemetry Replay Parameters" form consists of 7 input fields arranged vertically down the page and a row of 6 buttons along the bottom of the form. The Tcl script to generate and process this form would work as follows:
toplevelcontainer object (similar to the Xt
TopLevelShellwidget, except that the Tk widget can have more than one child).
framecontainer object (similar to Motif's
XmFormwidget) to hold all of the operator input fields. Call this the input region frame.
frameobject for the field's label and input objects.
labelobject for the field's label.
entryobject (a single-line,
XmText-like input widget) for the operator's input.
frameto hold the buttons.
buttonobject for each of the 6 buttons.
(This example is based on a demo script included in the Tk distribution.) The Tk/Tcl commands that create a widget also specify its resources; e.g, button labels, Tcl command bindings, fonts, default text, etc. In all, about 50 or 60 lines of Tk-extended Tcl would suffice to build the XSAR "Change Telemetry Replay Parameters" form. Compare this to the number of lines of UIL required to layout the page (many more, I assume and hope!). Also, compare the ease with which the Tcl script-based form can be changed. Modifying the look or behavior of existing fields and buttons - or adding entirely new ones - is a simple matter of editing the script file and bringing up the form again. There is no need to recompile a UIL file (if a change was made in the UIL file) or to refresh the Xrm database (if the change was effected in a defaults file).
Two questions that must be asked when considering design changes to any
system are how the changes affect the existing system and how the changes
will affect performance. Embedding Tcl in the display program would have
virtually no impact on the current TPOCC display software. The main
display routine would need to create the Tcl command interpreter (by
Tcl_CreateInterp()) and register the XSAR command
extensions (by calling, for each new command,
Tcl_CreateCommand()), but the other display code would remain
as-is. Tcl capabilities could be retrofitted into the existing TPOCC
widgets by adding new Tcl command resources to the widget structures and by
Tcl_Eval() call (to evaluate the commands) in the
Within the context of an interactive application, the performance of an embedded command language like Tcl is almost a non-issue. Countless users of programmable editors (Emacs, EVE, Brief, etc.), on computers ranging from PCs to VAXes to Crays, voice no complaints about the speed of their systems. John Ousterhout says that "the Tcl interpreter", averaging 160 microseconds per command on a DECstation 3100, "is fast enough to execute many hundreds of Tcl commands within a human response time". Inverting the 160-microsecond figure puts things into a little more perspective: 6250 commands per second. And the XSAR telemetry replay form described earlier was created by 50-60 lines of Tcl.
Executing Tcl commands as part of a real-time widget's update callback might be of concern if large volumes of data are being displayed or if data is coming in at a high rate. The solution to this problem, of course, is to specify an empty Tcl command for the update callback in the page definition.
In summary, user-programmable widgets would benefit the XSAR project in the following ways:
User-programmable widgets also hold out the potential of more fully isolating the display system kernel from mission-specific functionality. TPOCC programs like TSTOL and Display achieve their mission-independence by attempting to be prepared for every eventuality. A better approach might be one similar to that used by the state manager and telemetry decommutator: link the core display system with a mission-specific library of Tcl-accessible, C functions, including an initialization function that would register the mission's new Tcl commands with the Tcl interpreter. (Storing mission-specific code in a dynamically-linked, shared library would enable the creation of a truly generic display program, decommutator, or whatever. Loading and linking object modules at run-time is supported under UNIX, VMS, and VxWorks.)
While decoupling an application from its user interface is usually considered a desirable goal, some people have pointed out that a user interface that doesn't know anything about its underlying application is liable to be a user-unfriendly interface. User-programmable widgets let you have your cake and eat it too: languages like Tcl make it possible to embed mission-specific knowledge in the user interface while, at the same time, keeping it separate from the hard-coded heart of the display system.