<!DOCTYPE linuxdoc PUBLIC "-//XFree86//DTD linuxdoc//EN">
  <article>

    <!-- Title information -->
      <title>Distributed Multihead X design
      <author>Kevin E. Martin, David H. Dawes, and Rickard E. Faith
      <date>29 June 2004 (created 25 July 2001)
      <abstract>
        This document covers the motivation, background, design, and
        implementation of the distributed multihead X (DMX) system.  It
        is a living document and describes the current design and
        implementation details of the DMX system.  As the project
        progresses, this document will be continually updated to reflect
        the changes in the code and/or design.  <it>Copyright 2001 by VA
        Linux Systems, Inc., Fremont, California.  Copyright 2001-2004
        by Red Hat, Inc., Raleigh, North Carolina</it>
      </abstract>

      <!-- Table of contents -->
    <toc>

<!-- Begin the document -->
<sect>Introduction

<sect1>The Distributed Multihead X Server

<p>Current Open Source multihead solutions are limited to a single
physical machine.  A single X server controls multiple display devices,
which can be arranged as independent heads or unified into a single
desktop (with Xinerama).  These solutions are limited to the number of
physical devices that can co-exist in a single machine (e.g., due to the
number of AGP/PCI slots available for graphics cards).  Thus, large
tiled displays are not currently possible.  The work described in this
paper will eliminate the requirement that the display devices reside in
the same physical machine.  This will be accomplished by developing a
front-end proxy X server that will control multiple back-end X servers
that make up the large display.

<p>The overall structure of the distributed multihead X (DMX) project is
as follows: A single front-end X server will act as a proxy to a set of
back-end X servers, which handle all of the visible rendering.  X
clients will connect to the front-end server just as they normally would
to a regular X server.  The front-end server will present an abstracted
view to the client of a single large display.  This will ensure that all
standard X clients will continue to operate without modification
(limited, as always, by the visuals and extensions provided by the X
server).  Clients that are DMX-aware will be able to use an extension to
obtain information about the back-end servers (e.g., for placement of
pop-up windows, window alignments by the window manager, etc.).

<p>The architecture of the DMX server is divided into two main sections:
input (e.g., mouse and keyboard events) and output (e.g., rendering and
windowing requests).  Each of these are describe briefly below, and the
rest of this design document will describe them in greater detail.

<p>The DMX server can receive input from three general types of input
devices: "local" devices that are physically attached to the machine on
which DMX is running, "backend" devices that are physically attached to
one or more of the back-end X servers (and that generate events via the
X protocol stream from the backend), and "console" devices that can be
abstracted from any non-back-end X server.  Backend and console devices
are treated differently because the pointer device on the back-end X
server also controls the location of the hardware X cursor.  Full
support for XInput extension devices is provided.

<p>Rendering requests will be accepted by the front-end server; however,
rendering to visible windows will be broken down as needed and sent to
the appropriate back-end server(s) via X11 library calls for actual
rendering.  The basic framework will follow a Xnest-style approach.  GC
state will be managed in the front-end server and sent to the
appropriate back-end server(s) as required.  Pixmap rendering will (at
least initially) be handled by the front-end X server.  Windowing
requests (e.g., ordering, mapping, moving, etc.) will handled in the
front-end server.  If the request requires a visible change, the
windowing operation will be translated into requests for the appropriate
back-end server(s).  Window state will be mirrored in the back-end
server(s) as needed.

<sect1>Layout of Paper

<p>The next section describes the general development plan that was
actually used for implementation.  The final section discusses
outstanding issues at the conclusion of development.  The first appendix
provides low-level technical detail that may be of interest to those
intimately familiar with the X server architecture.  The final appendix
describes the four phases of development that were performed during the
first two years of development.

<p>The final year of work was divided into 9 tasks that are not
described in specific sections of this document.  The major tasks during
that time were the enhancement of the reconfiguration ability added in
Phase IV, addition of support for a dynamic number of back-end displays
(instead of a hard-coded limit), and the support for back-end display
and input removal and addition.  This work is mentioned in this paper,
but is not covered in detail.

<!-- ============================================================ -->
<sect>Development plan

<p>This section describes the development plan from approximately June
2001 through July 2003.

<sect1>Bootstrap code

<p>To allow for rapid development of the DMX server by multiple
developers during the first development stage, the problem will be
broken down into three tasks: the overall DMX framework, back-end
rendering services and input device handling services.  However, before
the work begins on these tasks, a simple framework that each developer
could use was implemented to bootstrap the development effort.  This
framework renders to a single back-end server and provides dummy input
devices (i.e., the keyboard and mouse).  The simple back-end rendering
service was implemented using the shadow framebuffer support currently
available in the XFree86 environment.

<p>Using this bootstrapping framework, each developer has been able to
work on each of the tasks listed above independently as follows: the
framework will be extended to handle arbitrary back-end server
configurations; the back-end rendering services will be transitioned to
the more efficient Xnest-style implementation; and, an input device
framework to handle various input devices via the input extension will
be developed.

<p>Status: The boot strap code is complete.  <!-- August 2001 -->


<sect1>Input device handling

<p>An X server (including the front-end X server) requires two core
input devices -- a keyboard and a pointer (mouse).  These core devices
are handled and required by the core X11 protocol.  Additional types of
input devices may be attached and utilized via the XInput extension.
These are usually referred to as ``XInput extension devices'',

<p>There are some options as to how the front-end X server gets its core
input devices:

<enum>
    <item>Local Input. The physical input devices (e.g., keyboard and
    mouse) can be attached directly to the front-end X server.  In this
    case, the keyboard and mouse on the machine running the front-end X
    server will be used.  The front-end will have drivers to read the
    raw input from those devices and convert it into the required X
    input events (e.g., key press/release, pointer button press/release,
    pointer motion).  The front-end keyboard driver will keep track of
    keyboard properties such as key and modifier mappings, autorepeat
    state, keyboard sound and led state.  Similarly the front-end
    pointer driver will keep track if pointer properties such as the
    button mapping and movement acceleration parameters.  With this
    option, input is handled fully in the front-end X server, and the
    back-end X servers are used in a display-only mode.  This option was
    implemented and works for a limited number of Linux-specific
    devices.  Adding additional local input devices for other
    architectures is expected to be relatively simple.

    <p>The following options are available for implementing local input
    devices:

    <enum>
        <item>The XFree86 X server has modular input drivers that could
        be adapted for this purpose.  The mouse driver supports a wide
        range of mouse types and interfaces, as well as a range of
        Operating System platforms.  The keyboard driver in XFree86 is
        not currently as modular as the mouse driver, but could be made
        so.  The XFree86 X server also has a range of other input
        drivers for extended input devices such as tablets and touch
        screens.  Unfortunately, the XFree86 drivers are generally
        complex, often simultaneously providing support for multiple
        devices across multiple architectures; and rely so heavily on
        XFree86-specific helper-functions, that this option was not
        pursued.


        <item>The <tt/kdrive/ X server in XFree86 has built-in drivers that
        support PS/2 mice and keyboard under Linux.  The mouse driver
        can indirectly handle other mouse types if the Linux utility
        <tt/gpm/ is used as to translate the native mouse protocol into
        PS/2 mouse format.  These drivers could be adapted and built in
        to the front-end X server if this range of hardware and OS
        support is sufficient.  While much simpler than the XFree86
        drivers, the <tt/kdrive/ drivers were not used for the DMX
        implementation.

        <item>Reimplementation of keyboard and mouse drivers from
        scratch for the DMX framework.  Because keyboard and mouse
        drivers are relatively trivial to implement, this pathway was
        selected.  Other drivers in the X source tree were referenced,
        and significant contributions from other drivers are noted in
        the DMX source code.
    </enum>

    <item>Backend Input.  The front-end can make use of the core input
    devices attached to one or more of the back-end X servers.  Core
    input events from multiple back-ends are merged into a single input
    event stream.  This can work sanely when only a single set of input
    devices is used at any given time.  The keyboard and pointer state
    will be handled in the front-end, with changes propagated to the
    back-end servers as needed.  This option was implemented and works
    well.  Because the core pointer on a back-end controls the hardware
    mouse on that back-end, core pointers cannot be treated as XInput
    extension devices.  However, all back-end XInput extensions devices
    can be mapped to either DMX core or DMX XInput extension devices.

    <item>Console Input.  The front-end server could create a console
    window that is displayed on an X server independent of the back-end
    X servers.  This console window could display things like the
    physical screen layout, and the front-end could get its core input
    events from events delivered to the console window.  This option was
    implemented and works well.  To help the human navigate, window
    outlines are also displayed in the console window.  Further, console
    windows can be used as either core or XInput extension devices.

    <item>Other options were initially explored, but they were all
    partial subsets of the options listed above and, hence, are
    irrelevant.

</enum>

<p>Although extended input devices are not specifically mentioned in the
Distributed X requirements, the options above were all implemented so
that XInput extension devices were supported.

<p>The bootstrap code (Xdmx) had dummy input devices, and these are
still supported in the final version.  These do the necessary
initialization to satisfy the X server's requirements for core pointer
and keyboard devices, but no input events are ever generated.

<p>Status: The input code is complete.  Because of the complexity of the
XFree86 input device drivers (and their heavy reliance on XFree86
infrastructure), separate low-level device drivers were implemented for
Xdmx.  The following kinds of drivers are supported (in general, the
devices can be treated arbitrarily as "core" input devices or as XInput
"extension" devices; and multiple instances of different kinds of
devices can be simultaneously available):
    <enum>
        <item> A "dummy" device drive that never generates events.

        <item> "Local" input is from the low-level hardware on which the
        Xdmx binary is running.  This is the only area where using the
        XFree86 driver infrastructure would have been helpful, and then
        only partially, since good support for generic USB devices does
        not yet exist in XFree86 (in any case, XFree86 and kdrive driver
        code was used where possible).  Currently, the following local
        devices are supported under Linux (porting to other operating
        systems should be fairly straightforward):
        <itemize>
            <item>Linux keyboard
            <item>Linux serial mouse (MS)
            <item>Linux PS/2 mouse
            <item>USB keyboard
            <item>USB mouse
            <item>USB generic device (e.g., joystick, gamepad, etc.)
        </itemize>

        <item> "Backend" input is taken from one or more of the back-end
        displays.  In this case, events are taken from the back-end X
        server and are converted to Xdmx events.  Care must be taken so
        that the sprite moves properly on the display from which input
        is being taken.

        <item> "Console" input is taken from an X window that Xdmx
        creates on the operator's display (i.e., on the machine running
        the Xdmx binary).  When the operator's mouse is inside the
        console window, then those events are converted to Xdmx events.
        Several special features are available: the console can display
        outlines of windows that are on the Xdmx display (to facilitate
        navigation), the cursor can be confined to the console, and a
        "fine" mode can be activated to allow very precise cursor
        positioning.
    </enum>


<!-- May 2002; July 2003 -->

<sect1>Output device handling

<p>The output of the DMX system displays rendering and windowing
requests across multiple screens.  The screens are typically arranged in
a grid such that together they represent a single large display.

<p>The output section of the DMX code consists of two parts.  The first
is in the front-end proxy X server (Xdmx), which accepts client
connections, manages the windows, and potentially renders primitives but
does not actually display any of the drawing primitives.  The second
part is the back-end X server(s), which accept commands from the
front-end server and display the results on their screens.

<sect2>Initialization

<p>The DMX front-end must first initialize its screens by connecting to
each of the back-end X servers and collecting information about each of
these screens.  However, the information collected from the back-end X
servers might be inconsistent.  Handling these cases can be difficult
and/or inefficient.  For example, a two screen system has one back-end X
server running at 16bpp while the second is running at 32bpp.
Converting rendering requests (e.g., XPutImage() or XGetImage()
requests) to the appropriate bit depth can be very time consuming.
Analyzing these cases to determine how or even if it is possible to
handle them is required.  The current Xinerama code handles many of
these cases (e.g., in PanoramiXConsolidate()) and will be used as a
starting point.  In general, the best solution is to use homogeneous X
servers and display devices.  Using back-end servers with the same depth
is a requirement of the final DMX implementation.

<p>Once this screen consolidation is finished, the relative position of
each back-end X server's screen in the unified screen is initialized.  A
full-screen window is opened on each of the back-end X servers, and the
cursor on each screen is turned off.  The final DMX implementation can
also make use of a partial-screen window, or multiple windows per
back-end screen.

<sect2>Handling rendering requests

<p>After initialization, X applications connect to the front-end server.
There are two possible implementations of how rendering and windowing
requests are handled in the DMX system:

<enum>
    <item>A shadow framebuffer is used in the front-end server as the
    render target.  In this option, all protocol requests are completely
    handled in the front-end server.  All state and resources are
    maintained in the front-end including a shadow copy of the entire
    framebuffer.  The framebuffers attached to the back-end servers are
    updated by XPutImage() calls with data taken directly from the
    shadow framebuffer.

    <p>This solution suffers from two main problems.  First, it does not
    take advantage of any accelerated hardware available in the system.
    Second, the size of the XPutImage() calls can be quite large and
    thus will be limited by the bandwidth available.

    <p>The initial DMX implementation used a shadow framebuffer by
    default.

    <item>Rendering requests are sent to each back-end server for
    handling (as is done in the Xnest server described above).  In this
    option, certain protocol requests are handled in the front-end
    server and certain requests are repackaged and then sent to the
    back-end servers.  The framebuffer is distributed across the
    multiple back-end servers.  Rendering to the framebuffer is handled
    on each back-end and can take advantage of any acceleration
    available on the back-end servers' graphics display device.  State
    is maintained both in the front and back-end servers.

    <p>This solution suffers from two main drawbacks.  First, protocol
    requests are sent to all back-end servers -- even those that will
    completely clip the rendering primitive -- which wastes bandwidth
    and processing time.  Second, state is maintained both in the front-
    and back-end servers.  These drawbacks are not as severe as in
    option 1 (above) and can either be overcome through optimizations or
    are acceptable.  Therefore, this option will be used in the final
    implementation.

    <p>The final DMX implementation defaults to this mechanism, but also
    supports the shadow framebuffer mechanism.  Several optimizations
    were implemented to eliminate the drawbacks of the default
    mechanism.  These optimizations are described the section below and
    in Phase II of the Development Results (see appendix).

</enum>

<p>Status: Both the shadow framebuffer and Xnest-style code is complete.
<!-- May 2002 -->


<sect1>Optimizing DMX

<p>Initially, the Xnest-style solution's performance will be measured
and analyzed to determine where the performance bottlenecks exist.
There are four main areas that will be addressed.

<p>First, to obtain reasonable interactivity with the first development
phase, XSync() was called after each protocol request.  The XSync()
function flushes any pending protocol requests.  It then waits for the
back-end to process the request and send a reply that the request has
completed.  This happens with each back-end server and performance
greatly suffers.  As a result of the way XSync() is called in the first
development phase, the batching that the X11 library performs is
effectively defeated.  The XSync() call usage will be analyzed and
optimized by batching calls and performing them at regular intervals,
except where interactivity will suffer (e.g., on cursor movements).

<p>Second, the initial Xnest-style solution described above sends the
repackaged protocol requests to all back-end servers regardless of
whether or not they would be completely clipped out.  The requests that
are trivially rejected on the back-end server wastes the limited
bandwidth available.  By tracking clipping changes in the DMX X server's
windowing code (e.g., by opening, closing, moving or resizing windows),
we can determine whether or not back-end windows are visible so that
trivial tests in the front-end server's GC ops drawing functions can
eliminate these unnecessary protocol requests.

<p>Third, each protocol request will be analyzed to determine if it is
possible to break the request into smaller pieces at display boundaries.
The initial ones to be analyzed are put and get image requests since
they will require the greatest bandwidth to transmit data between the
front and back-end servers.  Other protocol requests will be analyzed
and those that will benefit from breaking them into smaller requests
will be implemented.

<p>Fourth, an extension is being considered that will allow font glyphs to
be transferred from the front-end DMX X server to each back-end server.
This extension will permit the front-end to handle all font requests and
eliminate the requirement that all back-end X servers share the exact
same fonts as the front-end server.  We are investigating the
feasibility of this extension during this development phase.

<p>Other potential optimizations will be determined from the performance
analysis.

<p>Please note that in our initial design, we proposed optimizing BLT
operations (e.g., XCopyArea() and window moves) by developing an
extension that would allow individual back-end servers to directly copy
pixel data to other back-end servers.  This potential optimization was
in response to the simple image movement implementation that required
potentially many calls to GetImage() and PutImage().  However, the
current Xinerama implementation handles these BLT operations
differently.  Instead of copying data to and from screens, they generate
expose events -- just as happens in the case when a window is moved from
off a screen to on screen.  This approach saves the limited bandwidth
available between front and back-end servers and is being standardized
with Xinerama.  It also eliminates the potential setup problems and
security issues resulting from having each back-end server open
connections to all other back-end servers.  Therefore, we suggest
accepting Xinerama's expose event solution.

<p>Also note that the approach proposed in the second and third
optimizations might cause backing store algorithms in the back-end to be
defeated, so a DMX X server configuration flag will be added to disable
these optimizations.

<p>Status: The optimizations proposed above are complete.  It was
determined that the using the xfs font server was sufficient and
creating a new mechanism to pass glyphs was redundant; therefore, the
fourth optimization proposed above was not included in DMX.
<!-- September 2002 -->


<sect1>DMX X extension support

<p>The DMX X server keeps track of all the windowing information on the
back-end X servers, but does not currently export this information to
any client applications.  An extension will be developed to pass the
screen information and back-end window IDs to DMX-aware clients.  These
clients can then use this information to directly connect to and render
to the back-end windows.  Bypassing the DMX X server allows DMX-aware
clients to break up complex rendering requests on their own and send
them directly to the windows on the back-end server's screens.  An
example of a client that can make effective use of this extension is
Chromium.

<p>Status: The extension, as implemented, is fully documented in
"Client-to-Server DMX Extension to the X Protocol".  Future changes
might be required based on feedback and other proposed enhancements to
DMX.  Currently, the following facilities are supported:
<enum>
    <item>
        Screen information (clipping rectangle for each screen relative
        to the virtual screen)
    <item>
        Window information (window IDs and clipping information for each
        back-end window that corresponds to each DMX window)
    <item>
        Input device information (mappings from DMX device IDs to
        back-end device IDs)
    <item>
        Force window creation (so that a client can override the
        server-side lazy window creation optimization)
    <item>
        Reconfiguration (so that a client can request that a screen
        position be changed)
    <item>
        Addition and removal of back-end servers and back-end and
        console inputs.
</enum>
<!-- September 2002; July 2003 -->


<sect1>Common X extension support

<p>The XInput, XKeyboard and Shape extensions are commonly used
extensions to the base X11 protocol.  XInput allows multiple and
non-standard input devices to be accessed simultaneously.  These input
devices can be connected to either the front-end or back-end servers.
XKeyboard allows much better keyboard mappings control.  Shape adds
support for arbitrarily shaped windows and is used by various window
managers.  Nearly all potential back-end X servers make these extensions
available, and support for each one will be added to the DMX system.

<p>In addition to the extensions listed above, support for the X
Rendering extension (Render) is being developed.  Render adds digital
image composition to the rendering model used by the X Window System.
While this extension is still under development by Keith Packard of HP,
support for the current version will be added to the DMX system.

<p>Support for the XTest extension was added during the first
development phase.

<!-- WARNING: this list is duplicated in the Phase IV discussion -->
<p>Status: The following extensions are supported and are discussed in
more detail in Phase IV of the Development Results (see appendix):
    BIG-REQUESTS,
    DEC-XTRAP,
    DMX,
    DPMS,
    Extended-Visual-Information,
    GLX,
    LBX,
    RECORD,
    RENDER,
    SECURITY,
    SHAPE,
    SYNC,
    X-Resource,
    XC-APPGROUP,
    XC-MISC,
    XFree86-Bigfont,
    XINERAMA,
    XInputExtension,
    XKEYBOARD, and
    XTEST.
<!-- November 2002; updated February 2003, July 2003 -->

<sect1>OpenGL support

<p>OpenGL support using the Mesa code base exists in XFree86 release 4
and later.  Currently, the direct rendering infrastructure (DRI)
provides accelerated OpenGL support for local clients and unaccelerated
OpenGL support (i.e., software rendering) is provided for non-local
clients.

<p>The single head OpenGL support in XFree86 4.x will be extended to use
the DMX system.  When the front and back-end servers are on the same
physical hardware, it is possible to use the DRI to directly render to
the back-end servers.  First, the existing DRI will be extended to
support multiple display heads, and then to support the DMX system.
OpenGL rendering requests will be direct rendering to each back-end X
server.  The DRI will request the screen layout (either from the
existing Xinerama extension or a DMX-specific extension).  Support for
synchronized swap buffers will also be added (on hardware that supports
it).  Note that a single front-end server with a single back-end server
on the same physical machine can emulate accelerated indirect rendering.

<p>When the front and back-end servers are on different physical
hardware or are using non-XFree86 4.x X servers, a mechanism to render
primitives across the back-end servers will be provided.  There are
several options as to how this can be implemented.

<enum>
    <item>The existing OpenGL support in each back-end server can be
    used by repackaging rendering primitives and sending them to each
    back-end server.  This option is similar to the unoptimized
    Xnest-style approach mentioned above.  Optimization of this solution
    is beyond the scope of this project and is better suited to other
    distributed rendering systems.

    <item>Rendering to a pixmap in the front-end server using the
    current XFree86 4.x code, and then displaying to the back-ends via
    calls to XPutImage() is another option.  This option is similar to
    the shadow frame buffer approach mentioned above.  It is slower and
    bandwidth intensive, but has the advantage that the back-end servers
    are not required to have OpenGL support.
</enum>

<p>These, and other, options will be investigated in this phase of the
work.

<p>Work by others have made Chromium DMX-aware.  Chromium will use the
DMX X protocol extension to obtain information about the back-end
servers and will render directly to those servers, bypassing DMX.

<p>Status: OpenGL support by the glxProxy extension was implemented by
SGI and has been integrated into the DMX code base.
<!-- May 2003-->


<!-- ============================================================ -->
<sect>Current issues

<p>In this sections the current issues are outlined that require further
investigation.

<sect1>Fonts

<p>The font path and glyphs need to be the same for the front-end and
each of the back-end servers.  Font glyphs could be sent to the back-end
servers as necessary but this would consume a significant amount of
available bandwidth during font rendering for clients that use many
different fonts (e.g., Netscape).  Initially, the font server (xfs) will
be used to provide the fonts to both the front-end and back-end servers.
Other possibilities will be investigated during development.

<sect1>Zero width rendering primitives

<p>To allow pixmap and on-screen rendering to be pixel perfect, all
back-end servers must render zero width primitives exactly the same as
the front-end renders the primitives to pixmaps.  For those back-end
servers that do not exactly match, zero width primitives will be
automatically converted to one width primitives.  This can be handled in
the front-end server via the GC state.

<sect1>Output scaling

<p>With very large tiled displays, it might be difficult to read the
information on the standard X desktop.  In particular, the cursor can be
easily lost and fonts could be difficult to read.  Automatic primitive
scaling might prove to be very useful.  We will investigate the
possibility of scaling the cursor and providing a set of alternate
pre-scaled fonts to replace the standard fonts that many applications
use (e.g., fixed).  Other options for automatic scaling will also be
investigated.

<sect1>Per-screen colormaps

<p>Each screen's default colormap in the set of back-end X servers
should be able to be adjusted via a configuration utility.  This support
is would allow the back-end screens to be calibrated via custom gamma
tables.  On 24-bit systems that support a DirectColor visual, this type
of correction can be accommodated.  One possible implementation would be
to advertise to X client of the DMX server a TrueColor visual while
using DirectColor visuals on the back-end servers to implement this type
of color correction.  Other options will be investigated.

<!-- ============================================================ -->
<appendix>

<sect>Background

<p>This section describes the existing Open Source architectures that
can be used to handle multiple screens and upon which this development
project is based.  This section was written before the implementation
was finished, and may not reflect actual details of the implementation.
It is left for historical interest only.

<sect1>Core input device handling

<p>The following is a description of how core input devices are handled
by an X server.

<sect2>InitInput()

<p>InitInput() is a DDX function that is called at the start of each
server generation from the X server's main() function.  Its purpose is
to determine what input devices are connected to the X server, register
them with the DIX and MI layers, and initialize the input event queue.
InitInput() does not have a return value, but the X server will abort if
either a core keyboard device or a core pointer device are not
registered.  Extended input (XInput) devices can also be registered in
InitInput().

<p>InitInput() usually has implementation specific code to determine
which input devices are available.  For each input device it will be
using, it calls AddInputDevice():

<descrip>
<tag/AddInputDevice()/ This DIX function allocates the device structure,
registers a callback function (which handles device init, close, on and
off), and returns the input handle, which can be treated as opaque.  It
is called once for each input device.
</descrip>

<p>Once input handles for core keyboard and core pointer devices have
been obtained from AddInputDevice(), they are registered as core devices
by calling RegisterPointerDevice() and RegisterKeyboardDevice().  Each
of these should be called once.  If both core devices are not
registered, then the X server will exit with a fatal error when it
attempts to start the input devices in InitAndStartDevices(), which is
called directly after InitInput() (see below).

<descrip>
<tag/Register{Pointer,Keyboard}Device()/ These DIX functions take a
handle returned from AddInputDevice() and initialize the core input
device fields in inputInfo, and initialize the input processing and grab
functions for each core input device.
</descrip>

<p>The core pointer device is then registered with the miPointer code
(which does the high level cursor handling).  While this registration
is not necessary for correct miPointer operation in the current XFree86
code, it is still done mostly for compatibility reasons.

<descrip>
<tag/miRegisterPointerDevice()/ This MI function registers the core
pointer's input handle with with the miPointer code.
</descrip>

<p>The final part of InitInput() is the initialization of the input
event queue handling.  In most cases, the event queue handling provided
in the MI layer is used.  The primary XFree86 X server uses its own
event queue handling to support some special cases related to the XInput
extension and the XFree86-specific DGA extension.  For our purposes, the
MI event queue handling should be suitable.  It is initialized by
calling mieqInit():

<descrip>
<tag/mieqInit()/ This MI function initializes the MI event queue for the
core devices, and is passed the public component of the input handles
for the two core devices.
</descrip>

<p>If a wakeup handler is required to deliver synchronous input
events, it can be registered here by calling the DIX function
RegisterBlockAndWakeupHandlers().  (See the devReadInput() description
below.)

<sect2>InitAndStartDevices()

<p>InitAndStartDevices() is a DIX function that is called immediately
after InitInput() from the X server's main() function.  Its purpose is
to initialize each input device that was registered with
AddInputDevice(), enable each input device that was successfully
initialized, and create the list of enabled input devices.  Once each
registered device is processed in this way, the list of enabled input
devices is checked to make sure that both a core keyboard device and
core pointer device were registered and successfully enabled.  If not,
InitAndStartDevices() returns failure, and results in the the X server
exiting with a fatal error.

<p>Each registered device is initialized by calling its callback
(dev-&gt;deviceProc) with the DEVICE_INIT argument:

<descrip>
<tag/(*dev-&gt;deviceProc)(dev, DEVICE_INIT)/ This function initializes the
device structs with core information relevant to the device.

<p>For pointer devices, this means specifying the number of buttons,
default button mapping, the function used to get motion events (usually
miPointerGetMotionEvents()), the function used to change/control the
core pointer motion parameters (acceleration and threshold), and the
motion buffer size.

<p>For keyboard devices, this means specifying the keycode range,
default keycode to keysym mapping, default modifier mapping, and the
functions used to sound the keyboard bell and modify/control the
keyboard parameters (LEDs, bell pitch and duration, key click, which
keys are auto-repeating, etc).
</descrip>

<p>Each initialized device is enabled by calling EnableDevice():

<descrip>
<tag/EnableDevice()/ EnableDevice() calls the device callback with
DEVICE_ON:
    <descrip>
    <tag/(*dev-&gt;deviceProc)(dev, DEVICE_ON)/ This typically opens and
    initializes the relevant physical device, and when appropriate,
    registers the device's file descriptor (or equivalent) as a valid
    input source.
    </descrip>

    <p>EnableDevice() then adds the device handle to the X server's
    global list of enabled devices.
</descrip>

<p>InitAndStartDevices() then verifies that a valid core keyboard and
pointer has been initialized and enabled.  It returns failure if either
are missing.

<sect2>devReadInput()

<p>Each device will have some function that gets called to read its
physical input.  These may be called in a number of different ways.  In
the case of synchronous I/O, they will be called from a DDX
wakeup-handler that gets called after the server detects that new input is
available.  In the case of asynchronous I/O, they will be called from a
(SIGIO) signal handler triggered when new input is available.  This
function should do at least two things: make sure that input events get
enqueued, and make sure that the cursor gets moved for motion events
(except if these are handled later by the driver's own event queue
processing function, which cannot be done when using the MI event queue
handling).

<p>Events are queued by calling mieqEnqueue():

<descrip>
<tag/mieqEnqueue()/ This MI function is used to add input events to the
event queue.  It is simply passed the event to be queued.
</descrip>

<p>The cursor position should be updated when motion events are
enqueued, by calling either miPointerAbsoluteCursor() or
miPointerDeltaCursor():

<descrip>
<tag/miPointerAbsoluteCursor()/ This MI function is used to move the
cursor to the absolute coordinates provided.
<tag/miPointerDeltaCursor()/ This MI function is used to move the cursor
relative to its current position.
</descrip>

<sect2>ProcessInputEvents()

<p>ProcessInputEvents() is a DDX function that is called from the X
server's main dispatch loop when new events are available in the input
event queue.  It typically processes the enqueued events, and updates
the cursor/pointer position.  It may also do other DDX-specific event
processing.

<p>Enqueued events are processed by mieqProcessInputEvents() and passed
to the DIX layer for transmission to clients:

<descrip>
<tag/mieqProcessInputEvents()/ This function processes each event in the
event queue, and passes it to the device's input processing function.
The DIX layer provides default functions to do this processing, and they
handle the task of getting the events passed back to the relevant
clients.
<tag/miPointerUpdate()/ This function resynchronized the cursor position
with the new pointer position.  It also takes care of moving the cursor
between screens when needed in multi-head configurations.
</descrip>


<sect2>DisableDevice()

<p>DisableDevice is a DIX function that removes an input device from the
list of enabled devices.  The result of this is that the device no
longer generates input events.  The device's data structures are kept in
place, and disabling a device like this can be reversed by calling
EnableDevice().  DisableDevice() may be called from the DDX when it is
desirable to do so (e.g., the XFree86 server does this when VT
switching).  Except for special cases, this is not normally called for
core input devices.

<p>DisableDevice() calls the device's callback function with
<tt/DEVICE_OFF/:

<descrip>
<tag/(*dev-&gt;deviceProc)(dev, DEVICE_OFF)/ This typically closes the
relevant physical device, and when appropriate, unregisters the device's
file descriptor (or equivalent) as a valid input source.
</descrip>

<p>DisableDevice() then removes the device handle from the X server's
global list of enabled devices.


<sect2>CloseDevice()

<p>CloseDevice is a DIX function that removes an input device from the
list of available devices.  It disables input from the device and frees
all data structures associated with the device.  This function is
usually called from CloseDownDevices(), which is called from main() at
the end of each server generation to close all input devices.

<p>CloseDevice() calls the device's callback function with
<tt/DEVICE_CLOSE/:

<descrip>
<tag/(*dev-&gt;deviceProc)(dev, DEVICE_CLOSE)/ This typically closes the
relevant physical device, and when appropriate, unregisters the device's
file descriptor (or equivalent) as a valid input source.  If any device
specific data structures were allocated when the device was initialized,
they are freed here.
</descrip>

<p>CloseDevice() then frees the data structures that were allocated
for the device when it was registered/initialized.


<sect2>LegalModifier()
<!-- dmx/dmxinput.c - currently returns TRUE -->           
<p>LegalModifier() is a required DDX function that can be used to
restrict which keys may be modifier keys.  This seems to be present for
historical reasons, so this function should simply return TRUE
unconditionally.


<sect1>Output handling

<p>The following sections describe the main functions required to
initialize, use and close the output device(s) for each screen in the X
server.

<sect2>InitOutput()

<p>This DDX function is called near the start of each server generation
from the X server's main() function.  InitOutput()'s main purpose is to
initialize each screen and fill in the global screenInfo structure for
each screen.  It is passed three arguments: a pointer to the screenInfo
struct, which it is to initialize, and argc and argv from main(), which
can be used to determine additional configuration information.

<p>The primary tasks for this function are outlined below:

<enum>
    <item><bf/Parse configuration info:/ The first task of InitOutput()
    is to parses any configuration information from the configuration
    file.  In addition to the XF86Config file, other configuration
    information can be taken from the command line.  The command line
    options can be gathered either in InitOutput() or earlier in the
    ddxProcessArgument() function, which is called by
    ProcessCommandLine().  The configuration information determines the
    characteristics of the screen(s).  For example, in the XFree86 X
    server, the XF86Config file specifies the monitor information, the
    screen resolution, the graphics devices and slots in which they are
    located, and, for Xinerama, the screens' layout.

    <item><bf/Initialize screen info:/ The next task is to initialize
    the screen-dependent internal data structures.  For example, part of
    what the XFree86 X server does is to allocate its screen and pixmap
    private indices, probe for graphics devices, compare the probed
    devices to the ones listed in the XF86Config file, and add the ones that
    match to the internal xf86Screens&lsqb;&rsqb; structure.

    <item><bf/Set pixmap formats:/ The next task is to initialize the
    screenInfo's image byte order, bitmap bit order and bitmap scanline
    unit/pad.  The screenInfo's pixmap format's depth, bits per pixel
    and scanline padding is also initialized at this stage.

    <item><bf/Unify screen info:/ An optional task that might be done at
    this stage is to compare all of the information from the various
    screens and determines if they are compatible (i.e., if the set of
    screens can be unified into a single desktop).  This task has
    potential to be useful to the DMX front-end server, if Xinerama's
    PanoramiXConsolidate() function is not sufficient.
</enum>

<p>Once these tasks are complete, the valid screens are known and each
of these screens can be initialized by calling AddScreen().

<sect2>AddScreen()

<p>This DIX function is called from InitOutput(), in the DDX layer, to
add each new screen to the screenInfo structure.  The DDX screen
initialization function and command line arguments (i.e., argc and argv)
are passed to it as arguments.

<p>This function first allocates a new Screen structure and any privates
that are required.  It then initializes some of the fields in the Screen
struct and sets up the pixmap padding information.  Finally, it calls
the DDX screen initialization function ScreenInit(), which is described
below.  It returns the number of the screen that were just added, or -1
if there is insufficient memory to add the screen or if the DDX screen
initialization fails.

<sect2>ScreenInit()

<p>This DDX function initializes the rest of the Screen structure with
either generic or screen-specific functions (as necessary).  It also
fills in various screen attributes (e.g., width and height in
millimeters, black and white pixel values).

<p>The screen init function usually calls several functions to perform
certain screen initialization functions.  They are described below:

<descrip>
<tag/{mi,*fb}ScreenInit()/ The DDX layer's ScreenInit() function usually
calls another layer's ScreenInit() function (e.g., miScreenInit() or
fbScreenInit()) to initialize the fallbacks that the DDX driver does not
specifically handle.

<p>After calling another layer's ScreenInit() function, any
screen-specific functions either wrap or replace the other layer's
function pointers.  If a function is to be wrapped, each of the old
function pointers from the other layer are stored in a screen private
area.  Common functions to wrap are CloseScreen() and SaveScreen().

<tag/miInitializeBackingStore()/ This MI function initializes the
screen's backing storage functions, which are used to save areas of
windows that are currently covered by other windows.

<tag/miDCInitialize()/ This MI function initializes the MI cursor
display structures and function pointers.  If a hardware cursor is used,
the DDX layer's ScreenInit() function will wrap additional screen and
the MI cursor display function pointers.
</descrip>

<p>Another common task for ScreenInit() function is to initialize the
output device state.  For example, in the XFree86 X server, the
ScreenInit() function saves the original state of the video card and
then initializes the video mode of the graphics device.

<sect2>CloseScreen()

<p>This function restores any wrapped screen functions (and in
particular the wrapped CloseScreen() function) and restores the state of
the output device to its original state.  It should also free any
private data it created during the screen initialization.

<sect2>GC operations

<p>When the X server is requested to render drawing primitives, it does
so by calling drawing functions through the graphics context's operation
function pointer table (i.e., the GCOps functions).  These functions
render the basic graphics operations such as drawing rectangles, lines,
text or copying pixmaps.  Default routines are provided either by the MI
layer, which draws indirectly through a simple span interface, or by the
framebuffer layers (e.g., CFB, MFB, FB), which draw directly to a
linearly mapped frame buffer.

<p>To take advantage of special hardware on the graphics device,
specific GCOps functions can be replaced by device specific code.
However, many times the graphics devices can handle only a subset of the
possible states of the GC, so during graphics context validation,
appropriate routines are selected based on the state and capabilities of
the hardware.  For example, some graphics hardware can accelerate single
pixel width lines with certain dash patterns.  Thus, for dash patterns
that are not supported by hardware or for width 2 or greater lines, the
default routine is chosen during GC validation.

<p>Note that some pointers to functions that draw to the screen are
stored in the Screen structure.  They include GetImage(), GetSpans(),
PaintWindowBackground(), PaintWindowBorder(), CopyWindow() and
RestoreAreas().

<sect2>Xnest

<p>The Xnest X server is a special proxy X server that relays the X
protocol requests that it receives to a ``real'' X server that then
processes the requests and displays the results, if applicable.  To the X
applications, Xnest appears as if it is a regular X server.  However,
Xnest is both server to the X application and client of the real X
server, which will actually handle the requests.

<p>The Xnest server implements all of the standard input and output
initialization steps outlined above.

<descrip>
<tag/InitOutput()/ Xnest takes its configuration information from
command line arguments via ddxProcessArguments().  This information
includes the real X server display to connect to, its default visual
class, the screen depth, the Xnest window's geometry, etc.  Xnest then
connects to the real X server and gathers visual, colormap, depth and
pixmap information about that server's display, creates a window on that
server, which will be used as the root window for Xnest.

<p>Next, Xnest initializes its internal data structures and uses the
data from the real X server's pixmaps to initialize its own pixmap
formats.  Finally, it calls AddScreen(xnestOpenScreen, argc, argv) to
initialize each of its screens.

<tag/ScreenInit()/ Xnest's ScreenInit() function is called
xnestOpenScreen().  This function initializes its screen's depth and
visual information, and then calls miScreenInit() to set up the default
screen functions.  It then calls miInitializeBackingStore() and
miDCInitialize() to initialize backing store and the software cursor.
Finally, it replaces many of the screen functions with its own
functions that repackage and send the requests to the real X server to
which Xnest is attached.

<tag/CloseScreen()/ This function frees its internal data structure
allocations.  Since it replaces instead of wrapping screen functions,
there are no function pointers to unwrap.  This can potentially lead to
problems during server regeneration.

<tag/GC operations/ The GC operations in Xnest are very simple since
they leave all of the drawing to the real X server to which Xnest is
attached.  Each of the GCOps takes the request and sends it to the
real X server using standard Xlib calls.  For example, the X
application issues a XDrawLines() call.  This function turns into a
protocol request to Xnest, which calls the xnestPolylines() function
through Xnest's GCOps function pointer table.  The xnestPolylines()
function is only a single line, which calls XDrawLines() using the same
arguments that were passed into it.  Other GCOps functions are very
similar.  Two exceptions to the simple GCOps functions described above
are the image functions and the BLT operations.

<p>The image functions, GetImage() and PutImage(), must use a temporary
image to hold the image to be put of the image that was just grabbed
from the screen while it is in transit to the real X server or the
client.  When the image has been transmitted, the temporary image is
destroyed.

<p>The BLT operations, CopyArea() and CopyPlane(), handle not only the
copy function, which is the same as the simple cases described above,
but also the graphics exposures that result when the GC's graphics
exposure bit is set to True.  Graphics exposures are handled in a helper
function, xnestBitBlitHelper().  This function collects the exposure
events from the real X server and, if any resulting in regions being
exposed, then those regions are passed back to the MI layer so that it
can generate exposure events for the X application.
</descrip>

<p>The Xnest server takes its input from the X server to which it is
connected.  When the mouse is in the Xnest server's window, keyboard and
mouse events are received by the Xnest server, repackaged and sent back
to any client that requests those events.

<sect2>Shadow framebuffer

<p>The most common type of framebuffer is a linear array memory that
maps to the video memory on the graphics device.  However, accessing
that video memory over an I/O bus (e.g., ISA or PCI) can be slow.  The
shadow framebuffer layer allows the developer to keep the entire
framebuffer in main memory and copy it back to video memory at regular
intervals.  It also has been extended to handle planar video memory and
rotated framebuffers.

<p>There are two main entry points to the shadow framebuffer code:

<descrip>
<tag/shadowAlloc(width, height, bpp)/ This function allocates the in
memory copy of the framebuffer of size width*height*bpp.  It returns a
pointer to that memory, which will be used by the framebuffer
ScreenInit() code during the screen's initialization.

<tag/shadowInit(pScreen, updateProc, windowProc)/ This function
initializes the shadow framebuffer layer.  It wraps several screen
drawing functions, and registers a block handler that will update the
screen.  The updateProc is a function that will copy the damaged regions
to the screen, and the windowProc is a function that is used when the
entire linear video memory range cannot be accessed simultaneously so
that only a window into that memory is available (e.g., when using the
VGA aperture).
</descrip>

<p>The shadow framebuffer code keeps track of the damaged area of each
screen by calculating the bounding box of all drawing operations that
have occurred since the last screen update.  Then, when the block handler
is next called, only the damaged portion of the screen is updated.

<p>Note that since the shadow framebuffer is kept in main memory, all
drawing operations are performed by the CPU and, thus, no accelerated
hardware drawing operations are possible.


<sect1>Xinerama

<p>Xinerama is an X extension that allows multiple physical screens
controlled by a single X server to appear as a single screen.  Although
the extension allows clients to find the physical screen layout via
extension requests, it is completely transparent to clients at the core
X11 protocol level.  The original public implementation of Xinerama came
from Digital/Compaq.  XFree86 rewrote it, filling in some missing pieces
and improving both X11 core protocol compliance and performance.  The
Xinerama extension will be passing through X.Org's standardization
process in the near future, and the sample implementation will be based
on this rewritten version.

<p>The current implementation of Xinerama is based primarily in the DIX
(device independent) and MI (machine independent) layers of the X
server.  With few exceptions the DDX layers do not need any changes to
support Xinerama.  X server extensions often do need modifications to
provide full Xinerama functionality.

<p>The following is a code-level description of how Xinerama functions.

<p>Note: Because the Xinerama extension was originally called the
PanoramiX extension, many of the Xinerama functions still have the
PanoramiX prefix.

<descrip>
    <tag/PanoramiXExtensionInit()/ PanoramiXExtensionInit() is a
    device-independent extension function that is called at the start of
    each server generation from InitExtensions(), which is called from
    the X server's main() function after all output devices have been
    initialized, but before any input devices have been initialized.

    <p>PanoramiXNumScreens is set to the number of physical screens.  If
    only one physical screen is present, the extension is disabled, and
    PanoramiXExtensionInit() returns without doing anything else.

    <p>The Xinerama extension is registered by calling AddExtension().
    
    <p>A local per-screen array of data structures
    (panoramiXdataPtr&lsqb;&rsqb;)
    is allocated for each physical screen, and GC and Screen private
    indexes are allocated, and both GC and Screen private areas are
    allocated for each physical screen.  These hold Xinerama-specific
    per-GC and per-Screen data.  Each screen's CreateGC and CloseScreen
    functions are wrapped by XineramaCreateGC() and
    XineramaCloseScreen() respectively.  Some new resource classes are
    created for Xinerama drawables and GCs, and resource types for
    Xinerama windows, pixmaps and colormaps.

    <p>A region (XineramaScreenRegions&lsqb;i&rsqb;) is initialized for each
    physical screen, and single region (PanoramiXScreenRegion) is
    initialized to be the union of the screen regions.  The
    panoramiXdataPtr&lsqb;&rsqb; array is also initialized with the size and
    origin of each screen.  The relative positioning information for the
    physical screens is taken from the array
    dixScreenOrigins&lsqb;&rsqb;, which
    the DDX layer must initialize in InitOutput().  The bounds of the
    combined screen is also calculated (PanoramiXPixWidth and
    PanoramiXPixHeight).

    <p>The DIX layer has a list of function pointers
    (ProcVector&lsqb;&rsqb;) that
    holds the entry points for the functions that process core protocol
    requests.  The requests that Xinerama must intercept and break up
    into physical screen-specific requests are wrapped.  The original
    set is copied to SavedProcVector&lsqb;&rsqb;.  The types of requests
    intercepted are Window requests, GC requests, colormap requests,
    drawing requests, and some geometry-related requests.  This wrapping
    allows the bulk of the protocol request processing to be handled
    transparently to the DIX layer.  Some operations cannot be dealt with
    in this way and are handled with Xinerama-specific code within the
    DIX layer.

    <tag/PanoramiXConsolidate()/ PanoramiXConsolidate() is a
    device-independent extension function that is called directly from
    the X server's main() function after extensions and input/output
    devices have been initialized, and before the root windows are
    defined and initialized.

    <p>This function finds the set of depths (PanoramiXDepths&lsqb;&rsqb;) and
    visuals (PanoramiXVisuals&lsqb;&rsqb;)
    common to all of the physical screens.
    PanoramiXNumDepths is set to the number of common depths, and
    PanoramiXNumVisuals is set to the number of common visuals.
    Resources are created for the single root window and the default
    colormap.  Each of these resources has per-physical screen entries.

    <tag/PanoramiXCreateConnectionBlock()/ PanoramiXConsolidate() is a
    device-independent extension function that is called directly from
    the X server's main() function after the per-physical screen root
    windows are created.  It is called instead of the standard DIX
    CreateConnectionBlock() function.  If this function returns FALSE,
    the X server exits with a fatal error.  This function will return
    FALSE if no common depths were found in PanoramiXConsolidate().
    With no common depths, Xinerama mode is not possible.

    <p>The connection block holds the information that clients get when
    they open a connection to the X server.  It includes information
    such as the supported pixmap formats, number of screens and the
    sizes, depths, visuals, default colormap information, etc, for each
    of the screens (much of information that <tt/xdpyinfo/ shows).  The
    connection block is initialized with the combined single screen
    values that were calculated in the above two functions.

    <p>The Xinerama extension allows the registration of connection
    block callback functions.  The purpose of these is to allow other
    extensions to do processing at this point.  These callbacks can be
    registered by calling XineramaRegisterConnectionBlockCallback() from
    the other extension's ExtensionInit() function.  Each registered
    connection block callback is called at the end of
    PanoramiXCreateConnectionBlock().
</descrip>

<sect2>Xinerama-specific changes to the DIX code

<p>There are a few types of Xinerama-specific changes within the DIX
code.  The main ones are described here.

<p>Functions that deal with colormap or GC -related operations outside of
the intercepted protocol requests have a test added to only do the
processing for screen numbers > 0.  This is because they are handled for
the single Xinerama screen and the processing is done once for screen 0.

<p>The handling of motion events does some coordinate translation between
the physical screen's origin and screen zero's origin.  Also, motion
events must be reported relative to the composite screen origin rather
than the physical screen origins.

<p>There is some special handling for cursor, window and event processing
that cannot (either not at all or not conveniently) be done via the
intercepted protocol requests.  A particular case is the handling of
pointers moving between physical screens.

<sect2>Xinerama-specific changes to the MI code

<p>The only Xinerama-specific change to the MI code is in miSendExposures()
to handle the coordinate (and window ID) translation for expose events.

<sect2>Intercepted DIX core requests

<p>Xinerama breaks up drawing requests for dispatch to each physical
screen.  It also breaks up windows into pieces for each physical screen.
GCs are translated into per-screen GCs.  Colormaps are replicated on
each physical screen.  The functions handling the intercepted requests
take care of breaking the requests and repackaging them so that they can
be passed to the standard request handling functions for each screen in
turn.  In addition, and to aid the repackaging, the information from
many of the intercepted requests is used to keep up to date the
necessary state information for the single composite screen.  Requests
(usually those with replies) that can be satisfied completely from this
stored state information do not call the standard request handling
functions.

<!-- ============================================================ -->

<sect>Development Results

<p>In this section the results of each phase of development are
discussed.  This development took place between approximately June 2001
and July 2003.

<sect1>Phase I

<p>The initial development phase dealt with the basic implementation
including the bootstrap code, which used the shadow framebuffer, and the
unoptimized implementation, based on an Xnest-style implementation.

<sect2>Scope

<p>The goal of Phase I is to provide fundamental functionality that can
act as a foundation for ongoing work:
<enum>
    <item>Develop the proxy X server
    <itemize>
	<item>The proxy X server will operate on the X11 protocol and
	relay requests as necessary to correctly perform the request.
	<item>Work will be based on the existing work for Xinerama and
	Xnest.
	<item>Input events and windowing operations are handled in the
	proxy server and rendering requests are repackaged and sent to
	each of the back-end servers for display.
	<item>The multiple screen layout (including support for
	overlapping screens) will be user configurable via a
	configuration file or through the configuration tool.
    </itemize>
    <item>Develop graphical configuration tool
    <itemize>
	<item>There will be potentially a large number of X servers to
	configure into a single display.  The tool will allow the user
	to specify which servers are involved in the configuration and
	how they should be laid out.
    </itemize>
    <item>Pass the X Test Suite
    <itemize>
	<item>The X Test Suite covers the basic X11 operations.  All
	tests known to succeed must correctly operate in the distributed
	X environment.
    </itemize>
</enum>

<p>For this phase, the back-end X servers are assumed to be unmodified X
servers that do not support any DMX-related protocol extensions; future
optimization pathways are considered, but are not implemented; and the
configuration tool is assumed to rely only on libraries in the X source
tree (e.g., Xt).

<sect2>Results

<p>The proxy X server, Xdmx, was developed to distribute X11 protocol
requests to the set of back-end X servers.  It opens a window on each
back-end server, which represents the part of the front-end's root
window that is visible on that screen.  It mirrors window, pixmap and
other state in each back-end server.  Drawing requests are sent to
either windows or pixmaps on each back-end server.  This code is based
on Xnest and uses the existing Xinerama extension.

<p>Input events can be taken from (1) devices attached to the back-end
server, (2) core devices attached directly to the Xdmx server, or (3)
from a ``console'' window on another X server.  Events for these devices
are gathered, processed and delivered to clients attached to the Xdmx
server.

<p>An intuitive configuration format was developed to help the user
easily configure the multiple back-end X servers.  It was defined (see
grammar in Xdmx man page) and a parser was implemented that is used by
the Xdmx server and by a standalone xdmxconfig utility.  The parsing
support was implemented such that it can be easily factored out of the X
source tree for use with other tools (e.g., vdl).  Support for
converting legacy vdl-format configuration files to the DMX format is
provided by the vdltodmx utility.

<p>Originally, the configuration file was going to be a subsection of
XFree86's XF86Config file, but that was not possible since Xdmx is a
completely separate X server.  Thus, a separate config file format was
developed.  In addition, a graphical configuration
tool, xdmxconfig, was developed to allow the user to create and arrange
the screens in the configuration file.  The <bf/-configfile/ and <bf/-config/
command-line options can be used to start Xdmx using a configuration
file.

<p>An extension that enables remote input testing is required for the X
Test Suite to function.  During this phase, this extension (XTEST) was
implemented in the Xdmx server.  The results from running the X Test
Suite are described in detail below.

<sect2>X Test Suite

          <sect3> Introduction
            <p>
              The X Test Suite contains tests that verify Xlib functions
              operate correctly.  The test suite is designed to run on a
              single X server; however, since X applications will not be
              able to tell the difference between the DMX server and a
              standard X server, the X Test Suite should also run on the
              DMX server.
            <p>
              The Xdmx server was tested with the X Test Suite, and the
              existing failures are noted in this section.  To put these
              results in perspective, we first discuss expected X Test
              failures and how errors in underlying systems can impact
              Xdmx test results.

          <sect3>Expected Failures for a Single Head
            <p>
              A correctly implemented X server with a single screen is
              expected to fail certain X Test tests.  The following
              well-known errors occur because of rounding error in the X
              server code:
              <verb>
XDrawArc: Tests 42, 63, 66, 73
XDrawArcs: Tests 45, 66, 69, 76
              </verb>
            <p>
              The following failures occur because of the high-level X
              server implementation:
              <verb>
XLoadQueryFont: Test 1
XListFontsWithInfo: Tests 3, 4
XQueryFont: Tests 1, 2
              </verb>
            <p>
              The following test fails when running the X server as root
              under Linux because of the way directory modes are
              interpreted:
              <verb>
XWriteBitmapFile: Test 3
              </verb>
            <p>
              Depending on the video card used for the back-end, other
              failures may also occur because of bugs in the low-level
              driver implementation.  Over time, failures of this kind
              are usually fixed by XFree86, but will show up in Xdmx
              testing until then.

          <sect3>Expected Failures for Xinerama
            <p>
              Xinerama fails several X Test Suite tests because of
              design decisions made for the current implementation of
              Xinerama.  Over time, many of these errors will be
              corrected by XFree86 and the group working on a new
              Xinerama implementation.  Therefore, Xdmx will also share
              X Suite Test failures with Xinerama.
            <p>
              We may be able to fix or work-around some of these
              failures at the Xdmx level, but this will require
              additional exploration that was not part of Phase I.
            <p>
              Xinerama is constantly improving, and the list of
              Xinerama-related failures depends on XFree86 version and
              the underlying graphics hardware.  We tested with a
              variety of hardware, including nVidia, S3, ATI Radeon,
              and Matrox G400 (in dual-head mode).  The list below
              includes only those failures that appear to be from the
              Xinerama layer, and does not include failures listed in
              the previous section, or failures that appear to be from
              the low-level graphics driver itself:
            <p>
              These failures were noted with multiple Xinerama
              configurations:
              <verb>
XCopyPlane: Tests 13, 22, 31 (well-known Xinerama implementation issue)
XSetFontPath: Test 4
XGetDefault: Test 5
XMatchVisualInfo: Test 1
              </verb>
            <p>
              These failures were noted only when using one dual-head
              video card with a 4.2.99.x XFree86 server:
              <verb>
XListPixmapFormats: Test 1
XDrawRectangles: Test 45
              </verb>
            <p>
              These failures were noted only when using two video cards
              from different vendors with a 4.1.99.x XFree86 server:
              <verb>
XChangeWindowAttributes: Test 32
XCreateWindow: Test 30
XDrawLine: Test 22
XFillArc: Test 22
XChangeKeyboardControl: Tests 9, 10
XRebindKeysym: Test 1
              </verb>

          <sect3>Additional Failures from Xdmx
            <p>
              When running Xdmx, no unexpected failures were noted.
              Since the Xdmx server is based on Xinerama, we expect to
              have most of the Xinerama failures present in the Xdmx
              server.  Similarly, since the Xdmx server must rely on the
              low-level device drivers on each back-end server, we also
              expect that Xdmx will exhibit most of the back-end
              failures.  Here is a summary:
              <verb>
XListPixmapFormats: Test 1 (configuration dependent)
XChangeWindowAttributes: Test 32
XCreateWindow: Test 30
XCopyPlane: Test 13, 22, 31
XSetFontPath: Test 4
XGetDefault: Test 5 (configuration dependent)
XMatchVisualInfo: Test 1
XRebindKeysym: Test 1 (configuration dependent)
              </verb>
            <p>
              Note that this list is shorter than the combined list for
              Xinerama because Xdmx uses different code paths to perform
              some Xinerama operations.  Further, some Xinerama failures
              have been fixed in the XFree86 4.2.99.x CVS repository.
            
          <sect3>Summary and Future Work
            <p>
              Running the X Test Suite on Xdmx does not produce any
              failures that cannot be accounted for by the underlying
              Xinerama subsystem used by the front-end or by the
              low-level device-driver code running on the back-end X
              servers.  The Xdmx server therefore is as ``correct'' as
              possible with respect to the standard set of X Test Suite
              tests.
            <p>
              During the following phases, we will continue to verify
              Xdmx correctness using the X Test Suite.  We may also use
              other tests suites or write additional tests that run
              under the X Test Suite that specifically verify the
              expected behavior of DMX.

<sect2>Fonts

<p>In Phase I, fonts are handled directly by both the front-end and the
back-end servers, which is required since we must treat each back-end
server during this phase as a ``black box''.  What this requires is that
<bf/the front- and back-end servers must share the exact same font
path/.  There are two ways to help make sure that all servers share the
same font path:

<enum>
    <item>First, each server can be configured to use the same font
    server.  The font server, xfs, can be configured to serve fonts to
    multiple X servers via TCP.

    <item>Second, each server can be configured to use the same font
    path and either those font paths can be copied to each back-end
    machine or they can be mounted (e.g., via NFS) on each back-end
    machine.
</enum>

<p>One additional concern is that a client program can set its own font
path, and if it does so, then that font path must be available on each
back-end machine.

<p>The -fontpath command line option was added to allow users to
initialize the font path of the front end server.  This font path is
propagated to each back-end server when the default font is loaded.  If
there are any problems, an error message is printed, which will describe
the problem and list the current font path.  For more information about
setting the font path, see the -fontpath option description in the man
page.

<sect2>Performance

<p>Phase I of development was not intended to optimize performance.  Its
focus was on completely and correctly handling the base X11 protocol in
the Xdmx server.  However, several insights were gained during Phase I,
which are listed here for reference during the next phase of
development.

<enum>
    <item>Calls to XSync() can slow down rendering since it requires a
    complete round trip to and from a back-end server.  This is
    especially problematic when communicating over long haul networks.
    <item>Sending drawing requests to only the screens that they overlap
    should improve performance.
</enum>

<sect2>Pixmaps

<p>Pixmaps were originally expected to be handled entirely in the
front-end X server; however, it was found that this overly complicated
the rendering code and would have required sending potentially large
images to each back server that required them when copying from pixmap
to screen.  Thus, pixmap state is mirrored in the back-end server just
as it is with regular window state.  With this implementation, the same
rendering code that draws to windows can be used to draw to pixmaps on
the back-end server, and no large image transfers are required to copy
from pixmap to window.

<!-- ============================================================ -->
<sect1>Phase II

<p>The second phase of development concentrates on performance
optimizations.  These optimizations are documented here, with
<tt/x11perf/ data to show how the optimizations improve performance.

<p>All benchmarks were performed by running Xdmx on a dual processor
1.4GHz AMD Athlon machine with 1GB of RAM connecting over 100baseT to
two single-processor 1GHz Pentium III machines with 256MB of RAM and ATI
Rage 128 (RF) video cards.  The front end was running Linux
2.4.20-pre1-ac1 and the back ends were running Linux 2.4.7-10 and
version 4.2.99.1 of XFree86 pulled from the XFree86 CVS repository on
August 7, 2002.  All systems were running Red Hat Linux 7.2.

<sect2>Moving from XFree86 4.1.99.1 to 4.2.0.0

<p>For phase II, the working source tree was moved to the branch tagged
with dmx-1-0-branch and was updated from version 4.1.99.1 (20 August
2001) of the XFree86 sources to version 4.2.0.0 (18 January 2002).
After this update, the following tests were noted to be more than 10%
faster:
            <verb>
1.13   Fill 300x300 opaque stippled trapezoid (161x145 stipple) 
1.16   Fill 1x1 tiled trapezoid (161x145 tile) 
1.13   Fill 10x10 tiled trapezoid (161x145 tile) 
1.17   Fill 100x100 tiled trapezoid (161x145 tile) 
1.16   Fill 1x1 tiled trapezoid (216x208 tile) 
1.20   Fill 10x10 tiled trapezoid (216x208 tile) 
1.15   Fill 100x100 tiled trapezoid (216x208 tile) 
1.37   Circulate Unmapped window (200 kids) 
            </verb>
And the following tests were noted to be more than 10% slower:
            <verb>
0.88   Unmap window via parent (25 kids) 
0.75   Circulate Unmapped window (4 kids) 
0.79   Circulate Unmapped window (16 kids) 
0.80   Circulate Unmapped window (25 kids) 
0.82   Circulate Unmapped window (50 kids) 
0.85   Circulate Unmapped window (75 kids) 
            </verb>
<p>These changes were not caused by any changes in the DMX system, and
may point to changes in the XFree86 tree or to tests that have more
"jitter" than most other <tt/x11perf/ tests.

<sect2>Global changes

<p>During the development of the Phase II DMX server, several global
changes were made.  These changes were also compared with the Phase I
server.  The following tests were noted to be more than 10% faster:
            <verb>
1.13   Fill 300x300 opaque stippled trapezoid (161x145 stipple) 
1.15   Fill 1x1 tiled trapezoid (161x145 tile) 
1.13   Fill 10x10 tiled trapezoid (161x145 tile) 
1.17   Fill 100x100 tiled trapezoid (161x145 tile) 
1.16   Fill 1x1 tiled trapezoid (216x208 tile) 
1.19   Fill 10x10 tiled trapezoid (216x208 tile) 
1.15   Fill 100x100 tiled trapezoid (216x208 tile) 
1.15   Circulate Unmapped window (4 kids) 
            </verb>

<p>The following tests were noted to be more than 10% slower:
            <verb>
0.69   Scroll 10x10 pixels 
0.68   Scroll 100x100 pixels 
0.68   Copy 10x10 from window to window 
0.68   Copy 100x100 from window to window 
0.76   Circulate Unmapped window (75 kids) 
0.83   Circulate Unmapped window (100 kids) 
            </verb>

<p>For the remainder of this analysis, the baseline of comparison will
be the Phase II deliverable with all optimizations disabled (unless
otherwise noted).  This will highlight how the optimizations in
isolation impact performance.

<sect2>XSync() Batching

<p>During the Phase I implementation, XSync() was called after every
protocol request made by the DMX server.  This provided the DMX server
with an interactive feel, but defeated X11's protocol buffering system
and introduced round-trip wire latency into every operation.  During
Phase II, DMX was changed so that protocol requests are no longer
followed by calls to XSync().  Instead, the need for an XSync() is
noted, and XSync() calls are only made every 100mS or when the DMX
server specifically needs to make a call to guarantee interactivity.
With this new system, X11 buffers protocol as much as possible during a
100mS interval, and many unnecessary XSync() calls are avoided.

<p>Out of more than 300 <tt/x11perf/ tests, 8 tests became more than 100
times faster, with 68 more than 50X faster, 114 more than 10X faster,
and 181 more than 2X faster.  See table below for summary.

<p>The following tests were noted to be more than 10% slower with
XSync() batching on:
            <verb>
0.88   500x500 tiled rectangle (161x145 tile) 
0.89   Copy 500x500 from window to window 
            </verb>

<sect2>Offscreen Optimization

<p>Windows span one or more of the back-end servers' screens; however,
during Phase I development, windows were created on every back-end
server and every rendering request was sent to every window regardless
of whether or not that window was visible.  With the offscreen
optimization, the DMX server tracks when a window is completely off of a
back-end server's screen and, in that case, it does not send rendering
requests to those back-end windows.  This optimization saves bandwidth
between the front and back-end servers, and it reduces the number of
XSync() calls.  The performance tests were run on a DMX system with only
two back-end servers.  Greater performance gains will be had as the
number of back-end servers increases.

<p>Out of more than 300 <tt/x11perf/ tests, 3 tests were at least twice as
fast, and 146 tests were at least 10% faster.  Two tests were more than
10% slower with the offscreen optimization:
            <verb>
0.88   Hide/expose window via popup (4 kids) 
0.89   Resize unmapped window (75 kids) 
            </verb>

<sect2>Lazy Window Creation Optimization

<p>As mentioned above, during Phase I, windows were created on every
back-end server even if they were not visible on that back-end.  With
the lazy window creation optimization, the DMX server does not create
windows on a back-end server until they are either visible or they
become the parents of a visible window.  This optimization builds on the
offscreen optimization (described above) and requires it to be enabled.

<p>The lazy window creation optimization works by creating the window
data structures in the front-end server when a client creates a window,
but delays creation of the window on the back-end server(s).  A private
window structure in the DMX server saves the relevant window data and
tracks changes to the window's attributes and stacking order for later
use.  The only times a window is created on a back-end server are (1)
when it is mapped and is at least partially overlapping the back-end
server's screen (tracked by the offscreen optimization), or (2) when the
window becomes the parent of a previously visible window.  The first
case occurs when a window is mapped or when a visible window is copied,
moved or resized and now overlaps the back-end server's screen.  The
second case occurs when starting a window manager after having created
windows to which the window manager needs to add decorations.

<p>When either case occurs, a window on the back-end server is created
using the data saved in the DMX server's window private data structure.
The stacking order is then adjusted to correctly place the window on the
back-end and lastly the window is mapped.  From this time forward, the
window is handled exactly as if the window had been created at the time
of the client's request.

<p>Note that when a window is no longer visible on a back-end server's
screen (e.g., it is moved offscreen), the window is not destroyed;
rather, it is kept and reused later if the window once again becomes
visible on the back-end server's screen.  Originally with this
optimization, destroying windows was implemented but was later rejected
because it increased bandwidth when windows were opaquely moved or
resized, which is common in many window managers.

<p>The performance tests were run on a DMX system with only two back-end
servers.  Greater performance gains will be had as the number of
back-end servers increases.

<p>This optimization improved the following <tt/x11perf/ tests by more
than 10%:
            <verb>
1.10   500x500 rectangle outline 
1.12   Fill 100x100 stippled trapezoid (161x145 stipple) 
1.20   Circulate Unmapped window (50 kids) 
1.19   Circulate Unmapped window (75 kids) 
            </verb>
            
<sect2>Subdividing Rendering Primitives

<p>X11 imaging requests transfer significant data between the client and
the X server.  During Phase I, the DMX server would then transfer the
image data to each back-end server.  Even with the offscreen
optimization (above), these requests still required transferring
significant data to each back-end server that contained a visible
portion of the window.  For example, if the client uses XPutImage() to
copy an image to a window that overlaps the entire DMX screen, then the
entire image is copied by the DMX server to every back-end server.

<p>To reduce the amount of data transferred between the DMX server and
the back-end servers when XPutImage() is called, the image data is
subdivided and only the data that will be visible on a back-end server's
screen is sent to that back-end server.  Xinerama already implements a
subdivision algorithm for XGetImage() and no further optimization was
needed.

<p>Other rendering primitives were analyzed, but the time required to
subdivide these primitives was a significant proportion of the time
required to send the entire rendering request to the back-end server, so
this optimization was rejected for the other rendering primitives.

<p>Again, the performance tests were run on a DMX system with only two
back-end servers.  Greater performance gains will be had as the number
of back-end servers increases.

<p>This optimization improved the following <tt/x11perf/ tests by more
than 10%:
            <verb>
1.12   Fill 100x100 stippled trapezoid (161x145 stipple) 
1.26   PutImage 10x10 square 
1.83   PutImage 100x100 square 
1.91   PutImage 500x500 square 
1.40   PutImage XY 10x10 square 
1.48   PutImage XY 100x100 square 
1.50   PutImage XY 500x500 square 
1.45   Circulate Unmapped window (75 kids) 
1.74   Circulate Unmapped window (100 kids) 
            </verb>

<p>The following test was noted to be more than 10% slower with this
optimization:
            <verb>
0.88   10-pixel fill chord partial circle 
            </verb>

<sect2>Summary of x11perf Data

<p>With all of the optimizations on, 53 <tt/x11perf/ tests are more than
100X faster than the unoptimized Phase II deliverable, with 69 more than
50X faster, 73 more than 10X faster, and 199 more than twice as fast.
No tests were more than 10% slower than the unoptimized Phase II
deliverable.  (Compared with the Phase I deliverable, only Circulate
Unmapped window (100 kids) was more than 10% slower than the Phase II
deliverable.  As noted above, this test seems to have wider variability
than other <tt/x11perf/ tests.)

<p>The following table summarizes relative <tt/x11perf/ test changes for
all optimizations individually and collectively.  Note that some of the
optimizations have a synergistic effect when used together.
            <verb>

1: XSync() batching only
2: Off screen optimizations only
3: Window optimizations only
4: Subdivprims only
5: All optimizations

    1     2    3    4      5 Operation
------ ---- ---- ---- ------ ---------
  2.14 1.85 1.00 1.00   4.13 Dot 
  1.67 1.80 1.00 1.00   3.31 1x1 rectangle 
  2.38 1.43 1.00 1.00   2.44 10x10 rectangle 
  1.00 1.00 0.92 0.98   1.00 100x100 rectangle 
  1.00 1.00 1.00 1.00   1.00 500x500 rectangle 
  1.83 1.85 1.05 1.06   3.54 1x1 stippled rectangle (8x8 stipple) 
  2.43 1.43 1.00 1.00   2.41 10x10 stippled rectangle (8x8 stipple) 
  0.98 1.00 1.00 1.00   1.00 100x100 stippled rectangle (8x8 stipple) 
  1.00 1.00 1.00 1.00   0.98 500x500 stippled rectangle (8x8 stipple) 
  1.75 1.75 1.00 1.00   3.40 1x1 opaque stippled rectangle (8x8 stipple) 
  2.38 1.42 1.00 1.00   2.34 10x10 opaque stippled rectangle (8x8 stipple) 
  1.00 1.00 0.97 0.97   1.00 100x100 opaque stippled rectangle (8x8 stipple) 
  1.00 1.00 1.00 1.00   0.99 500x500 opaque stippled rectangle (8x8 stipple) 
  1.82 1.82 1.04 1.04   3.56 1x1 tiled rectangle (4x4 tile) 
  2.33 1.42 1.00 1.00   2.37 10x10 tiled rectangle (4x4 tile) 
  1.00 0.92 1.00 1.00   1.00 100x100 tiled rectangle (4x4 tile) 
  1.00 1.00 1.00 1.00   1.00 500x500 tiled rectangle (4x4 tile) 
  1.94 1.62 1.00 1.00   3.66 1x1 stippled rectangle (17x15 stipple) 
  1.74 1.28 1.00 1.00   1.73 10x10 stippled rectangle (17x15 stipple) 
  1.00 1.00 1.00 0.89   0.98 100x100 stippled rectangle (17x15 stipple) 
  1.00 1.00 1.00 1.00   0.98 500x500 stippled rectangle (17x15 stipple) 
  1.94 1.62 1.00 1.00   3.67 1x1 opaque stippled rectangle (17x15 stipple) 
  1.69 1.26 1.00 1.00   1.66 10x10 opaque stippled rectangle (17x15 stipple) 
  1.00 0.95 1.00 1.00   1.00 100x100 opaque stippled rectangle (17x15 stipple) 
  1.00 1.00 1.00 1.00   0.97 500x500 opaque stippled rectangle (17x15 stipple) 
  1.93 1.61 0.99 0.99   3.69 1x1 tiled rectangle (17x15 tile) 
  1.73 1.27 1.00 1.00   1.72 10x10 tiled rectangle (17x15 tile) 
  1.00 1.00 1.00 1.00   0.98 100x100 tiled rectangle (17x15 tile) 
  1.00 1.00 0.97 0.97   1.00 500x500 tiled rectangle (17x15 tile) 
  1.95 1.63 1.00 1.00   3.83 1x1 stippled rectangle (161x145 stipple) 
  1.80 1.30 1.00 1.00   1.83 10x10 stippled rectangle (161x145 stipple) 
  0.97 1.00 1.00 1.00   1.01 100x100 stippled rectangle (161x145 stipple) 
  1.00 1.00 1.00 1.00   0.98 500x500 stippled rectangle (161x145 stipple) 
  1.95 1.63 1.00 1.00   3.56 1x1 opaque stippled rectangle (161x145 stipple) 
  1.65 1.25 1.00 1.00   1.68 10x10 opaque stippled rectangle (161x145 stipple) 
  1.00 1.00 1.00 1.00   1.01 100x100 opaque stippled rectangle (161x145...
  1.00 1.00 1.00 1.00   0.97 500x500 opaque stippled rectangle (161x145...
  1.95 1.63 0.98 0.99   3.80 1x1 tiled rectangle (161x145 tile) 
  1.67 1.26 1.00 1.00   1.67 10x10 tiled rectangle (161x145 tile) 
  1.13 1.14 1.14 1.14   1.14 100x100 tiled rectangle (161x145 tile) 
  0.88 1.00 1.00 1.00   0.99 500x500 tiled rectangle (161x145 tile) 
  1.93 1.63 1.00 1.00   3.53 1x1 tiled rectangle (216x208 tile) 
  1.69 1.26 1.00 1.00   1.66 10x10 tiled rectangle (216x208 tile) 
  1.00 1.00 1.00 1.00   1.00 100x100 tiled rectangle (216x208 tile) 
  1.00 1.00 1.00 1.00   1.00 500x500 tiled rectangle (216x208 tile) 
  1.82 1.70 1.00 1.00   3.38 1-pixel line segment 
  2.07 1.56 0.90 1.00   3.31 10-pixel line segment 
  1.29 1.10 1.00 1.00   1.27 100-pixel line segment 
  1.05 1.06 1.03 1.03   1.09 500-pixel line segment 
  1.30 1.13 1.00 1.00   1.29 100-pixel line segment (1 kid) 
  1.32 1.15 1.00 1.00   1.32 100-pixel line segment (2 kids) 
  1.33 1.16 1.00 1.00   1.33 100-pixel line segment (3 kids) 
  1.92 1.64 1.00 1.00   3.73 10-pixel dashed segment 
  1.34 1.16 1.00 1.00   1.34 100-pixel dashed segment 
  1.24 1.11 0.99 0.97   1.23 100-pixel double-dashed segment 
  1.72 1.77 1.00 1.00   3.25 10-pixel horizontal line segment 
  1.83 1.66 1.01 1.00   3.54 100-pixel horizontal line segment 
  1.86 1.30 1.00 1.00   1.84 500-pixel horizontal line segment 
  2.11 1.52 1.00 0.99   3.02 10-pixel vertical line segment 
  1.21 1.10 1.00 1.00   1.20 100-pixel vertical line segment 
  1.03 1.03 1.00 1.00   1.02 500-pixel vertical line segment 
  4.42 1.68 1.00 1.01   4.64 10x1 wide horizontal line segment 
  1.83 1.31 1.00 1.00   1.83 100x10 wide horizontal line segment 
  1.07 1.00 0.96 1.00   1.07 500x50 wide horizontal line segment 
  4.10 1.67 1.00 1.00   4.62 10x1 wide vertical line segment 
  1.50 1.24 1.06 1.06   1.48 100x10 wide vertical line segment 
  1.06 1.03 1.00 1.00   1.05 500x50 wide vertical line segment 
  2.54 1.61 1.00 1.00   3.61 1-pixel line 
  2.71 1.48 1.00 1.00   2.67 10-pixel line 
  1.19 1.09 1.00 1.00   1.19 100-pixel line 
  1.04 1.02 1.00 1.00   1.03 500-pixel line 
  2.68 1.51 0.98 1.00   3.17 10-pixel dashed line 
  1.23 1.11 0.99 0.99   1.23 100-pixel dashed line 
  1.15 1.08 1.00 1.00   1.15 100-pixel double-dashed line 
  2.27 1.39 1.00 1.00   2.23 10x1 wide line 
  1.20 1.09 1.00 1.00   1.20 100x10 wide line 
  1.04 1.02 1.00 1.00   1.04 500x50 wide line 
  1.52 1.45 1.00 1.00   1.52 100x10 wide dashed line 
  1.54 1.47 1.00 1.00   1.54 100x10 wide double-dashed line 
  1.97 1.30 0.96 0.95   1.95 10x10 rectangle outline 
  1.44 1.27 1.00 1.00   1.43 100x100 rectangle outline 
  3.22 2.16 1.10 1.09   3.61 500x500 rectangle outline 
  1.95 1.34 1.00 1.00   1.90 10x10 wide rectangle outline 
  1.14 1.14 1.00 1.00   1.13 100x100 wide rectangle outline 
  1.00 1.00 1.00 1.00   1.00 500x500 wide rectangle outline 
  1.57 1.72 1.00 1.00   3.03 1-pixel circle 
  1.96 1.35 1.00 1.00   1.92 10-pixel circle 
  1.21 1.07 0.86 0.97   1.20 100-pixel circle 
  1.08 1.04 1.00 1.00   1.08 500-pixel circle 
  1.39 1.19 1.03 1.03   1.38 100-pixel dashed circle 
  1.21 1.11 1.00 1.00   1.23 100-pixel double-dashed circle 
  1.59 1.28 1.00 1.00   1.58 10-pixel wide circle 
  1.22 1.12 0.99 1.00   1.22 100-pixel wide circle 
  1.06 1.04 1.00 1.00   1.05 500-pixel wide circle 
  1.87 1.84 1.00 1.00   1.85 100-pixel wide dashed circle 
  1.90 1.93 1.01 1.01   1.90 100-pixel wide double-dashed circle 
  2.13 1.43 1.00 1.00   2.32 10-pixel partial circle 
  1.42 1.18 1.00 1.00   1.42 100-pixel partial circle 
  1.92 1.85 1.01 1.01   1.89 10-pixel wide partial circle 
  1.73 1.67 1.00 1.00   1.73 100-pixel wide partial circle 
  1.36 1.95 1.00 1.00   2.64 1-pixel solid circle 
  2.02 1.37 1.00 1.00   2.03 10-pixel solid circle 
  1.19 1.09 1.00 1.00   1.19 100-pixel solid circle 
  1.02 0.99 1.00 1.00   1.01 500-pixel solid circle 
  1.74 1.28 1.00 0.88   1.73 10-pixel fill chord partial circle 
  1.31 1.13 1.00 1.00   1.31 100-pixel fill chord partial circle 
  1.67 1.31 1.03 1.03   1.72 10-pixel fill slice partial circle 
  1.30 1.13 1.00 1.00   1.28 100-pixel fill slice partial circle 
  2.45 1.49 1.01 1.00   2.71 10-pixel ellipse 
  1.22 1.10 1.00 1.00   1.22 100-pixel ellipse 
  1.09 1.04 1.00 1.00   1.09 500-pixel ellipse 
  1.90 1.28 1.00 1.00   1.89 100-pixel dashed ellipse 
  1.62 1.24 0.96 0.97   1.61 100-pixel double-dashed ellipse 
  2.43 1.50 1.00 1.00   2.42 10-pixel wide ellipse 
  1.61 1.28 1.03 1.03   1.60 100-pixel wide ellipse 
  1.08 1.05 1.00 1.00   1.08 500-pixel wide ellipse 
  1.93 1.88 1.00 1.00   1.88 100-pixel wide dashed ellipse 
  1.94 1.89 1.01 1.00   1.94 100-pixel wide double-dashed ellipse 
  2.31 1.48 1.00 1.00   2.67 10-pixel partial ellipse 
  1.38 1.17 1.00 1.00   1.38 100-pixel partial ellipse 
  2.00 1.85 0.98 0.97   1.98 10-pixel wide partial ellipse 
  1.89 1.86 1.00 1.00   1.89 100-pixel wide partial ellipse 
  3.49 1.60 1.00 1.00   3.65 10-pixel filled ellipse 
  1.67 1.26 1.00 1.00   1.67 100-pixel filled ellipse 
  1.06 1.04 1.00 1.00   1.06 500-pixel filled ellipse 
  2.38 1.43 1.01 1.00   2.32 10-pixel fill chord partial ellipse 
  2.06 1.30 1.00 1.00   2.05 100-pixel fill chord partial ellipse 
  2.27 1.41 1.00 1.00   2.27 10-pixel fill slice partial ellipse 
  1.98 1.33 1.00 0.97   1.97 100-pixel fill slice partial ellipse 
 57.46 1.99 1.01 1.00 114.92 Fill 1x1 equivalent triangle 
 56.94 1.98 1.01 1.00  73.89 Fill 10x10 equivalent triangle 
  6.07 1.75 1.00 1.00   6.07 Fill 100x100 equivalent triangle 
 51.12 1.98 1.00 1.00 102.81 Fill 1x1 trapezoid 
 51.42 1.82 1.01 1.00  94.89 Fill 10x10 trapezoid 
  6.47 1.80 1.00 1.00   6.44 Fill 100x100 trapezoid 
  1.56 1.28 1.00 0.99   1.56 Fill 300x300 trapezoid 
 51.27 1.97 0.96 0.97 102.54 Fill 1x1 stippled trapezoid (8x8 stipple) 
 51.73 2.00 1.02 1.02  67.92 Fill 10x10 stippled trapezoid (8x8 stipple) 
  5.36 1.72 1.00 1.00   5.36 Fill 100x100 stippled trapezoid (8x8 stipple) 
  1.54 1.26 1.00 1.00   1.59 Fill 300x300 stippled trapezoid (8x8 stipple) 
 51.41 1.94 1.01 1.00 102.82 Fill 1x1 opaque stippled trapezoid (8x8 stipple) 
 50.71 1.95 0.99 1.00  65.44 Fill 10x10 opaque stippled trapezoid (8x8...
  5.33 1.73 1.00 1.00   5.36 Fill 100x100 opaque stippled trapezoid (8x8...
  1.58 1.25 1.00 1.00   1.58 Fill 300x300 opaque stippled trapezoid (8x8...
 51.56 1.96 0.99 0.90 103.68 Fill 1x1 tiled trapezoid (4x4 tile) 
 51.59 1.99 1.01 1.01  62.25 Fill 10x10 tiled trapezoid (4x4 tile) 
  5.38 1.72 1.00 1.00   5.38 Fill 100x100 tiled trapezoid (4x4 tile) 
  1.54 1.25 1.00 0.99   1.58 Fill 300x300 tiled trapezoid (4x4 tile) 
 51.70 1.98 1.01 1.01 103.98 Fill 1x1 stippled trapezoid (17x15 stipple) 
 44.86 1.97 1.00 1.00  44.86 Fill 10x10 stippled trapezoid (17x15 stipple) 
  2.74 1.56 1.00 1.00   2.73 Fill 100x100 stippled trapezoid (17x15 stipple) 
  1.29 1.14 1.00 1.00   1.27 Fill 300x300 stippled trapezoid (17x15 stipple) 
 51.41 1.96 0.96 0.95 103.39 Fill 1x1 opaque stippled trapezoid (17x15...
 45.14 1.96 1.01 1.00  45.14 Fill 10x10 opaque stippled trapezoid (17x15...
  2.68 1.56 1.00 1.00   2.68 Fill 100x100 opaque stippled trapezoid (17x15...
  1.26 1.10 1.00 1.00   1.28 Fill 300x300 opaque stippled trapezoid (17x15...
 51.13 1.97 1.00 0.99 103.39 Fill 1x1 tiled trapezoid (17x15 tile) 
 47.58 1.96 1.00 1.00  47.86 Fill 10x10 tiled trapezoid (17x15 tile) 
  2.74 1.56 1.00 1.00   2.74 Fill 100x100 tiled trapezoid (17x15 tile) 
  1.29 1.14 1.00 1.00   1.28 Fill 300x300 tiled trapezoid (17x15 tile) 
 51.13 1.97 0.99 0.97 103.39 Fill 1x1 stippled trapezoid (161x145 stipple) 
 45.14 1.97 1.00 1.00  44.29 Fill 10x10 stippled trapezoid (161x145 stipple) 
  3.02 1.77 1.12 1.12   3.38 Fill 100x100 stippled trapezoid (161x145 stipple) 
  1.31 1.13 1.00 1.00   1.30 Fill 300x300 stippled trapezoid (161x145 stipple) 
 51.27 1.97 1.00 1.00 103.10 Fill 1x1 opaque stippled trapezoid (161x145...
 45.01 1.97 1.00 1.00  45.01 Fill 10x10 opaque stippled trapezoid (161x145...
  2.67 1.56 1.00 1.00   2.69 Fill 100x100 opaque stippled trapezoid (161x145..
  1.29 1.13 1.00 1.01   1.27 Fill 300x300 opaque stippled trapezoid (161x145..
 51.41 1.96 1.00 0.99 103.39 Fill 1x1 tiled trapezoid (161x145 tile) 
 45.01 1.96 0.98 1.00  45.01 Fill 10x10 tiled trapezoid (161x145 tile) 
  2.62 1.36 1.00 1.00   2.69 Fill 100x100 tiled trapezoid (161x145 tile) 
  1.27 1.13 1.00 1.00   1.22 Fill 300x300 tiled trapezoid (161x145 tile) 
 51.13 1.98 1.00 1.00 103.39 Fill 1x1 tiled trapezoid (216x208 tile) 
 45.14 1.97 1.01 0.99  45.14 Fill 10x10 tiled trapezoid (216x208 tile) 
  2.62 1.55 1.00 1.00   2.71 Fill 100x100 tiled trapezoid (216x208 tile) 
  1.28 1.13 1.00 1.00   1.20 Fill 300x300 tiled trapezoid (216x208 tile) 
 50.71 1.95 1.00 1.00  54.70 Fill 10x10 equivalent complex polygon 
  5.51 1.71 0.96 0.98   5.47 Fill 100x100 equivalent complex polygons 
  8.39 1.97 1.00 1.00  16.75 Fill 10x10 64-gon (Convex) 
  8.38 1.83 1.00 1.00   8.43 Fill 100x100 64-gon (Convex) 
  8.50 1.96 1.00 1.00  16.64 Fill 10x10 64-gon (Complex) 
  8.26 1.83 1.00 1.00   8.35 Fill 100x100 64-gon (Complex) 
 14.09 1.87 1.00 1.00  14.05 Char in 80-char line (6x13) 
 11.91 1.87 1.00 1.00  11.95 Char in 70-char line (8x13) 
 11.16 1.85 1.01 1.00  11.10 Char in 60-char line (9x15) 
 10.09 1.78 1.00 1.00  10.09 Char16 in 40-char line (k14) 
  6.15 1.75 1.00 1.00   6.31 Char16 in 23-char line (k24) 
 11.92 1.90 1.03 1.03  11.88 Char in 80-char line (TR 10) 
  8.18 1.78 1.00 0.99   8.17 Char in 30-char line (TR 24) 
 42.83 1.44 1.01 1.00  42.11 Char in 20/40/20 line (6x13, TR 10) 
 27.45 1.43 1.01 1.01  27.45 Char16 in 7/14/7 line (k14, k24) 
 12.13 1.85 1.00 1.00  12.05 Char in 80-char image line (6x13) 
 10.00 1.84 1.00 1.00  10.00 Char in 70-char image line (8x13) 
  9.18 1.83 1.00 1.00   9.12 Char in 60-char image line (9x15) 
  9.66 1.82 0.98 0.95   9.66 Char16 in 40-char image line (k14) 
  5.82 1.72 1.00 1.00   5.99 Char16 in 23-char image line (k24) 
  8.70 1.80 1.00 1.00   8.65 Char in 80-char image line (TR 10) 
  4.67 1.66 1.00 1.00   4.67 Char in 30-char image line (TR 24) 
 84.43 1.47 1.00 1.00 124.18 Scroll 10x10 pixels 
  3.73 1.50 1.00 0.98   3.73 Scroll 100x100 pixels 
  1.00 1.00 1.00 1.00   1.00 Scroll 500x500 pixels 
 84.43 1.51 1.00 1.00 134.02 Copy 10x10 from window to window 
  3.62 1.51 0.98 0.98   3.62 Copy 100x100 from window to window 
  0.89 1.00 1.00 1.00   1.00 Copy 500x500 from window to window 
 57.06 1.99 1.00 1.00  88.64 Copy 10x10 from pixmap to window 
  2.49 2.00 1.00 1.00   2.48 Copy 100x100 from pixmap to window 
  1.00 0.91 1.00 1.00   0.98 Copy 500x500 from pixmap to window 
  2.04 1.01 1.00 1.00   2.03 Copy 10x10 from window to pixmap 
  1.05 1.00 1.00 1.00   1.05 Copy 100x100 from window to pixmap 
  1.00 1.00 0.93 1.00   1.04 Copy 500x500 from window to pixmap 
 58.52 1.03 1.03 1.02  57.95 Copy 10x10 from pixmap to pixmap 
  2.40 1.00 1.00 1.00   2.45 Copy 100x100 from pixmap to pixmap 
  1.00 1.00 1.00 1.00   1.00 Copy 500x500 from pixmap to pixmap 
 51.57 1.92 1.00 1.00  85.75 Copy 10x10 1-bit deep plane 
  6.37 1.75 1.01 1.01   6.37 Copy 100x100 1-bit deep plane 
  1.26 1.11 1.00 1.00   1.24 Copy 500x500 1-bit deep plane 
  4.23 1.63 0.98 0.97   4.38 Copy 10x10 n-bit deep plane 
  1.04 1.02 1.00 1.00   1.04 Copy 100x100 n-bit deep plane 
  1.00 1.00 1.00 1.00   1.00 Copy 500x500 n-bit deep plane 
  6.45 1.98 1.00 1.26  12.80 PutImage 10x10 square 
  1.10 1.87 1.00 1.83   2.11 PutImage 100x100 square 
  1.02 1.93 1.00 1.91   1.91 PutImage 500x500 square 
  4.17 1.78 1.00 1.40   7.18 PutImage XY 10x10 square 
  1.27 1.49 0.97 1.48   2.10 PutImage XY 100x100 square 
  1.00 1.50 1.00 1.50   1.52 PutImage XY 500x500 square 
  1.07 1.01 1.00 1.00   1.06 GetImage 10x10 square 
  1.01 1.00 1.00 1.00   1.01 GetImage 100x100 square 
  1.00 1.00 1.00 1.00   1.00 GetImage 500x500 square 
  1.56 1.00 0.99 0.97   1.56 GetImage XY 10x10 square 
  1.02 1.00 1.00 1.00   1.02 GetImage XY 100x100 square 
  1.00 1.00 1.00 1.00   1.00 GetImage XY 500x500 square 
  1.00 1.00 1.01 0.98   0.95 X protocol NoOperation 
  1.02 1.03 1.04 1.03   1.00 QueryPointer 
  1.03 1.02 1.04 1.03   1.00 GetProperty 
100.41 1.51 1.00 1.00 198.76 Change graphics context 
 45.81 1.00 0.99 0.97  57.10 Create and map subwindows (4 kids) 
 78.45 1.01 1.02 1.02  63.07 Create and map subwindows (16 kids) 
 73.91 1.01 1.00 1.00  56.37 Create and map subwindows (25 kids) 
 73.22 1.00 1.00 1.00  49.07 Create and map subwindows (50 kids) 
 72.36 1.01 0.99 1.00  32.14 Create and map subwindows (75 kids) 
 70.34 1.00 1.00 1.00  30.12 Create and map subwindows (100 kids) 
 55.00 1.00 1.00 0.99  23.75 Create and map subwindows (200 kids) 
 55.30 1.01 1.00 1.00 141.03 Create unmapped window (4 kids) 
 55.38 1.01 1.01 1.00 163.25 Create unmapped window (16 kids) 
 54.75 0.96 1.00 0.99 166.95 Create unmapped window (25 kids) 
 54.83 1.00 1.00 0.99 178.81 Create unmapped window (50 kids) 
 55.38 1.01 1.01 1.00 181.20 Create unmapped window (75 kids) 
 55.38 1.01 1.01 1.00 181.20 Create unmapped window (100 kids) 
 54.87 1.01 1.01 1.00 182.05 Create unmapped window (200 kids) 
 28.13 1.00 1.00 1.00  30.75 Map window via parent (4 kids) 
 36.14 1.01 1.01 1.01  32.58 Map window via parent (16 kids) 
 26.13 1.00 0.98 0.95  29.85 Map window via parent (25 kids) 
 40.07 1.00 1.01 1.00  27.57 Map window via parent (50 kids) 
 23.26 0.99 1.00 1.00  18.23 Map window via parent (75 kids) 
 22.91 0.99 1.00 0.99  16.52 Map window via parent (100 kids) 
 27.79 1.00 1.00 0.99  12.50 Map window via parent (200 kids) 
 22.35 1.00 1.00 1.00  56.19 Unmap window via parent (4 kids) 
  9.57 1.00 0.99 1.00  89.78 Unmap window via parent (16 kids) 
 80.77 1.01 1.00 1.00 103.85 Unmap window via parent (25 kids) 
 96.34 1.00 1.00 1.00 116.06 Unmap window via parent (50 kids) 
 99.72 1.00 1.00 1.00 124.93 Unmap window via parent (75 kids) 
112.36 1.00 1.00 1.00 125.27 Unmap window via parent (100 kids) 
105.41 1.00 1.00 0.99 120.00 Unmap window via parent (200 kids) 
 51.29 1.03 1.02 1.02  74.19 Destroy window via parent (4 kids) 
 86.75 0.99 0.99 0.99 116.87 Destroy window via parent (16 kids) 
106.43 1.01 1.01 1.01 127.49 Destroy window via parent (25 kids) 
120.34 1.01 1.01 1.00 140.11 Destroy window via parent (50 kids) 
126.67 1.00 0.99 0.99 145.00 Destroy window via parent (75 kids) 
126.11 1.01 1.01 1.00 140.56 Destroy window via parent (100 kids) 
128.57 1.01 1.00 1.00 137.91 Destroy window via parent (200 kids) 
 16.04 0.88 1.00 1.00  20.36 Hide/expose window via popup (4 kids) 
 19.04 1.01 1.00 1.00  23.48 Hide/expose window via popup (16 kids) 
 19.22 1.00 1.00 1.00  20.44 Hide/expose window via popup (25 kids) 
 17.41 1.00 0.91 0.97  17.68 Hide/expose window via popup (50 kids) 
 17.29 1.01 1.00 1.01  17.07 Hide/expose window via popup (75 kids) 
 16.74 1.00 1.00 1.00  16.17 Hide/expose window via popup (100 kids) 
 10.30 1.00 1.00 1.00  10.51 Hide/expose window via popup (200 kids) 
 16.48 1.01 1.00 1.00  26.05 Move window (4 kids) 
 17.01 0.95 1.00 1.00  23.97 Move window (16 kids) 
 16.95 1.00 1.00 1.00  22.90 Move window (25 kids) 
 16.05 1.01 1.00 1.00  21.32 Move window (50 kids) 
 15.58 1.00 0.98 0.98  19.44 Move window (75 kids) 
 14.98 1.02 1.03 1.03  18.17 Move window (100 kids) 
 10.90 1.01 1.01 1.00  12.68 Move window (200 kids) 
 49.42 1.00 1.00 1.00 198.27 Moved unmapped window (4 kids) 
 50.72 0.97 1.00 1.00 193.66 Moved unmapped window (16 kids) 
 50.87 1.00 0.99 1.00 195.09 Moved unmapped window (25 kids) 
 50.72 1.00 1.00 1.00 189.34 Moved unmapped window (50 kids) 
 50.87 1.00 1.00 1.00 191.33 Moved unmapped window (75 kids) 
 50.87 1.00 1.00 0.90 186.71 Moved unmapped window (100 kids) 
 50.87 1.00 1.00 1.00 179.19 Moved unmapped window (200 kids) 
 41.04 1.00 1.00 1.00  56.61 Move window via parent (4 kids) 
 69.81 1.00 1.00 1.00 130.82 Move window via parent (16 kids) 
 95.81 1.00 1.00 1.00 141.92 Move window via parent (25 kids) 
 95.98 1.00 1.00 1.00 149.43 Move window via parent (50 kids) 
 96.59 1.01 1.01 1.00 153.98 Move window via parent (75 kids) 
 97.19 1.00 1.00 1.00 157.30 Move window via parent (100 kids) 
 96.67 1.00 0.99 0.96 159.44 Move window via parent (200 kids) 
 17.75 1.01 1.00 1.00  27.61 Resize window (4 kids) 
 17.94 1.00 1.00 0.99  25.42 Resize window (16 kids) 
 17.92 1.01 1.00 1.00  24.47 Resize window (25 kids) 
 17.24 0.97 1.00 1.00  24.14 Resize window (50 kids) 
 16.81 1.00 1.00 0.99  22.75 Resize window (75 kids) 
 16.08 1.00 1.00 1.00  21.20 Resize window (100 kids) 
 12.92 1.00 0.99 1.00  16.26 Resize window (200 kids) 
 52.94 1.01 1.00 1.00 327.12 Resize unmapped window (4 kids) 
 53.60 1.01 1.01 1.01 333.71 Resize unmapped window (16 kids) 
 52.99 1.00 1.00 1.00 337.29 Resize unmapped window (25 kids) 
 51.98 1.00 1.00 1.00 329.38 Resize unmapped window (50 kids) 
 53.05 0.89 1.00 1.00 322.60 Resize unmapped window (75 kids) 
 53.05 1.00 1.00 1.00 318.08 Resize unmapped window (100 kids) 
 53.11 1.00 1.00 0.99 306.21 Resize unmapped window (200 kids) 
 16.76 1.00 0.96 1.00  19.46 Circulate window (4 kids) 
 17.24 1.00 1.00 0.97  16.24 Circulate window (16 kids) 
 16.30 1.03 1.03 1.03  15.85 Circulate window (25 kids) 
 13.45 1.00 1.00 1.00  14.90 Circulate window (50 kids) 
 12.91 1.00 1.00 1.00  13.06 Circulate window (75 kids) 
 11.30 0.98 1.00 1.00  11.03 Circulate window (100 kids) 
  7.58 1.01 1.01 0.99   7.47 Circulate window (200 kids) 
  1.01 1.01 0.98 1.00   0.95 Circulate Unmapped window (4 kids) 
  1.07 1.07 1.01 1.07   1.02 Circulate Unmapped window (16 kids) 
  1.04 1.09 1.06 1.05   0.97 Circulate Unmapped window (25 kids) 
  1.04 1.23 1.20 1.18   1.05 Circulate Unmapped window (50 kids) 
  1.18 1.53 1.19 1.45   1.24 Circulate Unmapped window (75 kids) 
  1.08 1.02 1.01 1.74   1.01 Circulate Unmapped window (100 kids) 
  1.01 1.12 0.98 0.91   0.97 Circulate Unmapped window (200 kids) 
            </verb>

<sect2>Profiling with OProfile

<p>OProfile (available from http://oprofile.sourceforge.net/) is a
system-wide profiler for Linux systems that uses processor-level
counters to collect sampling data.  OProfile can provide information
that is similar to that provided by <tt/gprof/, but without the
necessity of recompiling the program with special instrumentation (i.e.,
OProfile can collect statistical profiling information about optimized
programs).  A test harness was developed to collect OProfile data for
each <tt/x11perf/ test individually.

<p>Test runs were performed using the RETIRED_INSNS counter on the AMD
Athlon and the CPU_CLK_HALTED counter on the Intel Pentium III (with a
test configuration different from the one described above).  We have
examined OProfile output and have compared it with <tt/gprof/ output.
This investigation has not produced results that yield performance
increases in <tt/x11perf/ numbers.

<!--
<sect3>Retired Instructions

<p>The initial tests using OProfile were done using the RETIRED_INSNS
counter with DMX running on the dual-processor AMD Athlon machine - the
same test configuration that was described above and that was used for
other tests.  The RETIRED_INSNS counter counts retired instructions and
showed drawing, text, copying, and image tests to be dominated (&gt;
30%) by calls to Hash(), SecurityLookupIDByClass(),
SecurityLookupIDByType(), and StandardReadRequestFromClient().  Some of
these tests also executed significant instructions in
WaitForSomething().

<p>In contrast, the window tests executed significant
instructions in SecurityLookupIDByType(), Hash(),
StandardReadRequestFromClient(), but also executed significant
instructions in other routines, such as ConfigureWindow().  Some time
was spent looking at Hash() function, but optimizations in this routine
did not lead to a dramatic increase in <tt/x11perf/ performance.
-->

<!-- 
<sect3>Clock Cycles

<p>Retired instructions can be misleading because Intel/AMD instructions
execute in variable amounts of time.  The OProfile tests were repeated
using the Intel CPU_CLK_HALTED counter with DMX running on the second
back-end machine.  Note that this is a different test configuration that
the one described above.  However, these tests show the amount of time
(as measured in CPU cycles) that are spent in each routine.  Because
<tt/x11perf/ was running on the first back-end machine and because
window optimizations were on, the load on the second back-end machine
was not significant.

<p>Using CPU_CLK_HALTED, DMX showed simple drawing
tests spending more than 10% of their time in
StandardReadRequestFromClient(), with significant time (&gt; 20% total)
spent in SecurityLookupIDByClass(), WaitForSomething(), and Dispatch().
For these tests, &lt; 5% of the time was spent in Hash(), which explains
why optimizing the Hash() routine did not impact <tt/x11perf/ results.

<p>The trapezoid, text, scrolling, copying, and image tests were
dominated by time in ProcFillPoly(), PanoramiXFillPoly(), dmxFillPolygon(),
SecurityLookupIDByClass(), SecurityLookupIDByType(), and
StandardReadRequestFromClient().  Hash() time was generally above 5% but
less than 10% of total time.
-->

<sect2>X Test Suite

<p>The X Test Suite was run on the fully optimized DMX server using the
configuration described above.  The following failures were noted:
            <verb>
XListPixmapFormats: Test 1              [1]
XChangeWindowAttributes: Test 32        [1]
XCreateWindow: Test 30                  [1]
XFreeColors: Test 4                     [3]
XCopyArea: Test 13, 17, 21, 25, 30      [2]
XCopyPlane: Test 11, 15, 27, 31         [2]
XSetFontPath: Test 4                    [1]
XChangeKeyboardControl: Test 9, 10      [1]

[1] Previously documented errors expected from the Xinerama
    implementation (see Phase I discussion).
[2] Newly noted errors that have been verified as expected
    behavior of the Xinerama implementation.
[3] Newly noted error that has been verified as a Xinerama
    implementation bug.
            </verb>
          
<!-- ============================================================ -->
<sect1>Phase III

<p>During the third phase of development, support was provided for the
following extensions: SHAPE, RENDER, XKEYBOARD, XInput.

<sect2>SHAPE

<p>The SHAPE extension is supported.  Test applications (e.g., xeyes and
oclock) and window managers that make use of the SHAPE extension will
work as expected.

<sect2>RENDER

<p>The RENDER extension is supported.  The version included in the DMX
CVS tree is version 0.2, and this version is fully supported by Xdmx.
Applications using only version 0.2 functions will work correctly;
however, some apps that make use of functions from later versions do not
properly check the extension's major/minor version numbers.  These apps
will fail with a Bad Implementation error when using post-version 0.2
functions.  This is expected behavior.  When the DMX CVS tree is updated
to include newer versions of RENDER, support for these newer functions
will be added to the DMX X server.

<sect2>XKEYBOARD

<p>The XKEYBOARD extension is supported.  If present on the back-end X
servers, the XKEYBOARD extension will be used to obtain information
about the type of the keyboard for initialization.  Otherwise, the
keyboard will be initialized using defaults.  Note that this departs
from older behavior: when Xdmx is compiled without XKEYBOARD support,
the map from the back-end X server will be preserved.  With XKEYBOARD
support, the map is not preserved because better information and control
of the keyboard is available.

<sect2>XInput

<p>The XInput extension is supported.  Any device can be used as a core
device and be used as an XInput extension device, with the exception of
core devices on the back-end servers.  This limitation is present
because cursor handling on the back-end requires that the back-end
cursor sometimes track the Xdmx core cursor -- behavior that is
incompatible with using the back-end pointer as a non-core device.

<p>Currently, back-end extension devices are not available as Xdmx
extension devices, but this limitation should be removed in the future.

<p>To demonstrate the XInput extension, and to provide more examples for
low-level input device driver writers, USB device drivers have been
written for mice (usb-mou), keyboards (usb-kbd), and
non-mouse/non-keyboard USB devices (usb-oth).  Please see the man page
for information on Linux kernel drivers that are required for using
these Xdmx drivers.

<sect2>DPMS

<p>The DPMS extension is exported but does not do anything at this time.

<sect2>Other Extensions

<p>The LBX,
       SECURITY,
       XC-APPGROUP, and
       XFree86-Bigfont
extensions do not require any special Xdmx support and have been exported.

<p>The
    BIG-REQUESTS,
    DEC-XTRAP,
    DOUBLE-BUFFER,
    Extended-Visual-Information,
    FontCache,
    GLX,
    MIT-SCREEN-SAVER,
    MIT-SHM,
    MIT-SUNDRY-NONSTANDARD,
    RECORD,
    SECURITY,
    SGI-GLX,
    SYNC,
    TOG-CUP,
    X-Resource,
    XC-MISC,
    XFree86-DGA,
    XFree86-DRI,
    XFree86-Misc,
    XFree86-VidModeExtension, and
    XVideo
extensions are <it/not/ supported at this time, but will be evaluated
for inclusion in future DMX releases.  <bf>See below for additional work
on extensions after Phase III.</bf>

<sect1>Phase IV

<sect2>Moving to XFree86 4.3.0

<p>For Phase IV, the recent release of XFree86 4.3.0 (27 February 2003)
was merged onto the dmx.sourceforge.net CVS trunk and all work is
proceeding using this tree.

<sect2>Extensions 

<sect3>XC-MISC (supported)

<p>XC-MISC is used internally by the X library to recycle XIDs from the
X server.  This is important for long-running X server sessions.  Xdmx
supports this extension.  The X Test Suite passed and failed the exact
same tests before and after this extension was enabled.
<!-- Tested February/March 2003 -->

<sect3>Extended-Visual-Information (supported)

<p>The Extended-Visual-Information extension provides a method for an X
client to obtain detailed visual information.  Xdmx supports this
extension.  It was tested using the <tt>hw/dmx/examples/evi</tt> example
program.  <bf/Note that this extension is not Xinerama-aware/ -- it will
return visual information for each screen even though Xinerama is
causing the X server to export a single logical screen.
<!-- Tested March 2003 -->

<sect3>RES (supported)

<p>The X-Resource extension provides a mechanism for a client to obtain
detailed information about the resources used by other clients.  This
extension was tested with the <tt>hw/dmx/examples/res</tt> program.  The
X Test Suite passed and failed the exact same tests before and after
this extension was enabled.
<!-- Tested March 2003 -->

<sect3>BIG-REQUESTS (supported)

<p>This extension enables the X11 protocol to handle requests longer
than 262140 bytes.  The X Test Suite passed and failed the exact same
tests before and after this extension was enabled.
<!-- Tested March 2003 -->

<sect3>XSYNC (supported)

<p>This extension provides facilities for two different X clients to
synchronize their requests.  This extension was minimally tested with
<tt/xdpyinfo/ and the X Test Suite passed and failed the exact same
tests before and after this extension was enabled.
<!-- Tested March 2003 -->

<sect3>XTEST, RECORD, DEC-XTRAP (supported) and XTestExtension1 (not supported)

<p>The XTEST and RECORD extension were developed by the X Consortium for
use in the X Test Suite and are supported as a standard in the X11R6
tree.  They are also supported in Xdmx.  When X Test Suite tests that
make use of the XTEST extension are run, Xdmx passes and fails exactly
the same tests as does a standard XFree86 X server.  When the
<tt/rcrdtest/ test (a part of the X Test Suite that verifies the RECORD
extension) is run, Xdmx passes and fails exactly the same tests as does
a standard XFree86 X server. <!-- Tested February/March 2003 -->

<p>There are two older XTEST-like extensions: DEC-XTRAP and
XTestExtension1.  The XTestExtension1 extension was developed for use by
the X Testing Consortium for use with a test suite that eventually
became (part of?) the X Test Suite.  Unlike XTEST, which only allows
events to be sent to the server, the XTestExtension1 extension also
allowed events to be recorded (similar to the RECORD extension).  The
second is the DEC-XTRAP extension that was developed by the Digital
Equipment Corporation.

<p>The DEC-XTRAP extension is available from Xdmx and has been tested
with the <tt/xtrap*/ tools which are distributed as standard X11R6
clients. <!-- Tested March 2003 -->

<p>The XTestExtension1 is <em/not/ supported because it does not appear
to be used by any modern X clients (the few that support it also support
XTEST) and because there are no good methods available for testing that
it functions correctly (unlike XTEST and DEC-XTRAP, the code for
XTestExtension1 is not part of the standard X server source tree, so
additional testing is important). <!-- Tested March 2003 -->

<p>Most of these extensions are documented in the X11R6 source tree.
Further, several original papers exist that this author was unable to
locate -- for completeness and historical interest, citations are
provide:
<descrip>
<tag/XRECORD/ Martha Zimet. Extending X For Recording.  8th Annual X
Technical Conference Boston, MA January 24-26, 1994.
<tag/DEC-XTRAP/ Dick Annicchiarico, Robert Chesler, Alan Jamison. XTrap
Architecture. Digital Equipment Corporation, July 1991.
<tag/XTestExtension1/ Larry Woestman. X11 Input Synthesis Extension
Proposal. Hewlett Packard, November 1991.
</descrip>

<sect3>MIT-MISC (not supported)

<p>The MIT-MISC extension is used to control a bug-compatibility flag
that provides compatibility with xterm programs from X11R1 and X11R2.
There does not appear to be a single client available that makes use of
this extension and there is not way to verify that it works correctly.
The Xdmx server does <em/not/ support MIT-MISC.

<sect3>SCREENSAVER (not supported)

<p>This extension provides special support for the X screen saver.  It
was tested with beforelight, which appears to be the only client that
works with it.  When Xinerama was not active, <tt/beforelight/ behaved
as expected.  However, when Xinerama was active, <tt/beforelight/ did
not behave as expected.  Further, when this extension is not active,
<tt/xscreensaver/ (a widely-used X screen saver program) did not behave
as expected.  Since this extension is not Xinerama-aware and is not
commonly used with expected results by clients, we have left this
extension disabled at this time.

<sect3>GLX (supported)

<p>The GLX extension provides OpenGL and GLX windowing support.  In
Xdmx, the extension is called glxProxy, and it is Xinerama aware.  It
works by either feeding requests forward through Xdmx to each of the
back-end servers or handling them locally.  All rendering requests are
handled on the back-end X servers.  This code was donated to the DMX
project by SGI.  For the X Test Suite results comparison, see below.

<sect3>RENDER (supported)

<p>The X Rendering Extension (RENDER) provides support for digital image
composition.  Geometric and text rendering are supported.  RENDER is
partially Xinerama-aware, with text and the most basic compositing
operator; however, its higher level primitives (triangles, triangle
strips, and triangle fans) are not yet Xinerama-aware.  The RENDER
extension is still under development, and is currently at version 0.8.
Additional support will be required in DMX as more primitives and/or
requests are added to the extension.

<p>There is currently no test suite for the X Rendering Extension;
however, there has been discussion of developing a test suite as the
extension matures.  When that test suite becomes available, additional
testing can be performed with Xdmx.  The X Test Suite passed and failed
the exact same tests before and after this extension was enabled.

<sect3>Summary

<!-- WARNING: this list is duplicated in the "Common X extension
support" section -->
<p>To summarize, the following extensions are currently supported:
    BIG-REQUESTS,
    DEC-XTRAP,
    DMX,
    DPMS,
    Extended-Visual-Information,
    GLX,
    LBX,
    RECORD,
    RENDER,
    SECURITY,
    SHAPE,
    SYNC,
    X-Resource,
    XC-APPGROUP,
    XC-MISC,
    XFree86-Bigfont,
    XINERAMA,
    XInputExtension,
    XKEYBOARD, and
    XTEST.

<p>The following extensions are <em/not/ supported at this time:
    DOUBLE-BUFFER,
    FontCache,
    MIT-SCREEN-SAVER,
    MIT-SHM,
    MIT-SUNDRY-NONSTANDARD,
    TOG-CUP,
    XFree86-DGA,
    XFree86-Misc,
    XFree86-VidModeExtension,
    XTestExtensionExt1, and
    XVideo.

<sect2>Additional Testing with the X Test Suite

<sect3>XFree86 without XTEST

<p>After the release of XFree86 4.3.0, we retested the XFree86 X server
with and without using the XTEST extension.  When the XTEST extension
was <em/not/ used for testing, the XFree86 4.3.0 server running on our
usual test system with a Radeon VE card reported unexpected failures in
the following tests:
<verb>
XListPixmapFormats: Test 1
XChangeKeyboardControl: Tests 9, 10
XGetDefault: Test 5
XRebindKeysym: Test 1
</verb>

<sect3>XFree86 with XTEST

<p>When using the XTEST extension, the XFree86 4.3.0 server reported the
following errors:
<verb>
XListPixmapFormats: Test 1
XChangeKeyboardControl: Tests 9, 10
XGetDefault: Test 5
XRebindKeysym: Test 1

XAllowEvents: Tests 20, 21, 24
XGrabButton: Tests 5, 9-12, 14, 16, 19, 21-25
XGrabKey: Test 8
XSetPointerMapping: Test 3
XUngrabButton: Test 4
</verb>

<p>While these errors may be important, they will probably be fixed
eventually in the XFree86 source tree.  We are particularly interested
in demonstrating that the Xdmx server does not introduce additional
failures that are not known Xinerama failures.

<sect3>Xdmx with XTEST, without Xinerama, without GLX

<p>Without Xinerama, but using the XTEST extension, the following errors
were reported from Xdmx (note that these are the same as for the XFree86
4.3.0, except that XGetDefault no longer fails):
<verb>
XListPixmapFormats: Test 1
XChangeKeyboardControl: Tests 9, 10
XRebindKeysym: Test 1

XAllowEvents: Tests  20, 21, 24
XGrabButton: Tests 5, 9-12, 14, 16, 19, 21-25
XGrabKey: Test 8
XSetPointerMapping: Test 3
XUngrabButton: Test 4
</verb>

<sect3>Xdmx with XTEST, with Xinerama, without GLX

<p>With Xinerama, using the XTEST extension, the following errors
were reported from Xdmx:
<verb>
XListPixmapFormats: Test 1
XChangeKeyboardControl: Tests 9, 10
XRebindKeysym: Test 1

XAllowEvents: Tests 20, 21, 24
XGrabButton: Tests 5, 9-12, 14, 16, 19, 21-25
XGrabKey: Test 8
XSetPointerMapping: Test 3
XUngrabButton: Test 4

XCopyPlane: Tests 13, 22, 31 (well-known XTEST/Xinerama interaction issue)
XDrawLine: Test 67
XDrawLines: Test 91
XDrawSegments: Test 68
</verb>
Note that the first two sets of errors are the same as for the XFree86
4.3.0 server, and that the XCopyPlane error is a well-known error
resulting from an XTEST/Xinerama interaction when the request crosses a
screen boundary.  The XDraw* errors are resolved when the tests are run
individually and they do not cross a screen boundary.  We will
investigate these errors further to determine their cause.

<sect3>Xdmx with XTEST, with Xinerama, with GLX

<p>With GLX enabled, using the XTEST extension, the following errors
were reported from Xdmx (these results are from early during the Phase
IV development, but were confirmed with a late Phase IV snapshot):
<verb>
XListPixmapFormats: Test 1
XChangeKeyboardControl: Tests 9, 10
XRebindKeysym: Test 1

XAllowEvents: Tests 20, 21, 24
XGrabButton: Tests 5, 9-12, 14, 16, 19, 21-25
XGrabKey: Test 8
XSetPointerMapping: Test 3
XUngrabButton: Test 4

XClearArea: Test 8
XCopyArea: Tests 4, 5, 11, 14, 17, 23, 25, 27, 30
XCopyPlane: Tests 6, 7, 10, 19, 22, 31
XDrawArcs: Tests 89, 100, 102
XDrawLine: Test 67
XDrawSegments: Test 68
</verb>
Note that the first two sets of errors are the same as for the XFree86
4.3.0 server, and that the third set has different failures than when
Xdmx does not include GLX support.  Since the GLX extension adds new
visuals to support GLX's visual configs and the X Test Suite runs tests
over the entire set of visuals, additional rendering tests were run and
presumably more of them crossed a screen boundary.  This conclusion is
supported by the fact that nearly all of the rendering errors reported
are resolved when the tests are run individually and they do no cross a
screen boundary.

<p>Further, when hardware rendering is disabled on the back-end displays,
many of the errors in the third set are eliminated, leaving only:
<verb>
XClearArea: Test 8
XCopyArea: Test 4, 5, 11, 14, 17, 23, 25, 27, 30
XCopyPlane: Test 6, 7, 10, 19, 22, 31
</verb>

<sect3>Conclusion

<p>We conclude that all of the X Test Suite errors reported for Xdmx are
the result of errors in the back-end X server or the Xinerama
implementation.  Further, all of these errors that can be reasonably
fixed at the Xdmx layer have been.  (Where appropriate, we have
submitted patches to the XFree86 and Xinerama upstream maintainers.)

<sect2>Dynamic Reconfiguration

<p>During this development phase, dynamic reconfiguration support was
added to DMX.  This support allows an application to change the position
and offset of a back-end server's screen.  For example, if the
application would like to shift a screen slightly to the left, it could
query Xdmx for the screen's &lt;x,y&gt; position and then dynamically
reconfigure that screen to be at position &lt;x+10,y&gt;.  When a screen
is dynamically reconfigured, input handling and a screen's root window
dimensions are adjusted as needed.  These adjustments are transparent to
the user.

<sect3>Dynamic reconfiguration extension

<p>The application interface to DMX's dynamic reconfiguration is through
a function in the DMX extension library:
<verb>
Bool DMXReconfigureScreen(Display *dpy, int screen, int x, int y)
</verb>
where <it/dpy/ is DMX server's display, <it/screen/ is the number of the
screen to be reconfigured, and <it/x/ and <it/y/ are the new upper,
left-hand coordinates of the screen to be reconfigured.

<p>The coordinates are not limited other than as required by the X
protocol, which limits all coordinates to a signed 16 bit number.  In
addition, all coordinates within a screen must also be legal values.
Therefore, setting a screen's upper, left-hand coordinates such that the
right or bottom edges of the screen is greater than 32,767 is illegal.

<sect3>Bounding box

<p>When the Xdmx server is started, a bounding box is calculated from
the screens' layout given either on the command line or in the
configuration file.  This bounding box is currently fixed for the
lifetime of the Xdmx server.

<p>While it is possible to move a screen outside of the bounding box, it
is currently not possible to change the dimensions of the bounding box.
For example, it is possible to specify coordinates of &lt;-100,-100&gt;
for the upper, left-hand corner of the bounding box, which was
previously at coordinates &lt;0,0&gt;.  As expected, the screen is moved
down and to the right; however, since the bounding box is fixed, the
left side and upper portions of the screen exposed by the
reconfiguration are no longer accessible on that screen.  Those
inaccessible regions are filled with black.

<p>This fixed bounding box limitation will be addressed in a future
development phase.

<sect3>Sample applications

<p>An example of where this extension is useful is in setting up a video
wall.  It is not always possible to get everything perfectly aligned,
and sometimes the positions are changed (e.g., someone might bump into a
projector).  Instead of physically moving projectors or monitors, it is
now possible to adjust the positions of the back-end server's screens
using the dynamic reconfiguration support in DMX.

<p>Other applications, such as automatic setup and calibration tools,
can make use of dynamic reconfiguration to correct for projector
alignment problems, as long as the projectors are still arranged
rectilinearly.  Horizontal and vertical keystone correction could be
applied to projectors to correct for non-rectilinear alignment problems;
however, this must be done external to Xdmx.

<p>A sample test program is included in the DMX server's examples
directory to demonstrate the interface and how an application might use
dynamic reconfiguration.  See <tt/dmxreconfig.c/ for details.

<sect3>Additional notes

<p>In the original development plan, Phase IV was primarily devoted to
adding OpenGL support to DMX; however, SGI became interested in the DMX
project and developed code to support OpenGL/GLX.  This code was later
donated to the DMX project and integrated into the DMX code base, which
freed the DMX developers to concentrate on dynamic reconfiguration (as
described above).

<sect2>Doxygen documentation

<p>Doxygen is an open-source (GPL) documentation system for generating
browseable documentation from stylized comments in the source code.  We
have placed all of the Xdmx server and DMX protocol source code files
under Doxygen so that comprehensive documentation for the Xdmx source
code is available in an easily browseable format.

<sect2>Valgrind

<p>Valgrind, an open-source (GPL) memory debugger for Linux, was used to
search for memory management errors.  Several memory leaks were detected
and repaired.  The following errors were not addressed:
<enum>
    <item>
        When the X11 transport layer sends a reply to the client, only
        those fields that are required by the protocol are filled in --
        unused fields are left as uninitialized memory and are therefore
        noted by valgrind.  These instances are not errors and were not
        repaired.
    <item>
        At each server generation, glxInitVisuals allocates memory that
        is never freed.  The amount of memory lost each generation
        approximately equal to 128 bytes for each back-end visual.
        Because the code involved is automatically generated, this bug
        has not been fixed and will be referred to SGI.
    <item>
        At each server generation, dmxRealizeFont calls XLoadQueryFont,
        which allocates a font structure that is not freed.
        dmxUnrealizeFont can free the font structure for the first
        screen, but cannot free it for the other screens since they are
        already closed by the time dmxUnrealizeFont could free them.
        The amount of memory lost each generation is approximately equal
        to 80 bytes per font per back-end.  When this bug is fixed in
        the the X server's device-independent (dix) code, DMX will be
        able to properly free the memory allocated by XLoadQueryFont.
</enum>

<sect2>RATS

<p>RATS (Rough Auditing Tool for Security) is an open-source (GPL)
security analysis tool that scans source code for common
security-related programming errors (e.g., buffer overflows and TOCTOU
races).  RATS was used to audit all of the code in the hw/dmx directory
and all "High" notations were checked manually.  The code was either
re-written to eliminate the warning, or a comment containing "RATS" was
inserted on the line to indicate that a human had checked the code.
Unrepaired warnings are as follows:
<enum>
    <item>
        Fixed-size buffers are used in many areas, but code has been
        added to protect against buffer overflows (e.g., XmuSnprint).
        The only instances that have not yet been fixed are in
        config/xdmxconfig.c (which is not part of the Xdmx server) and
        input/usb-common.c.
    <item>
        vprintf and vfprintf are used in the logging routines.  In
        general, all uses of these functions (e.g., dmxLog) provide a
        constant format string from a trusted source, so the use is
        relatively benign.
    <item>
        glxProxy/glxscreens.c uses getenv and strcat.  The use of these
        functions is safe and will remain safe as long as
        ExtensionsString is longer then GLXServerExtensions (ensuring
        this may not be ovious to the casual programmer, but this is in
        automatically generated code, so we hope that the generator
        enforces this constraint).
</enum>

  </article>
  
  <!-- Local Variables: -->
  <!-- fill-column: 72  -->
  <!-- End:             -->