DfMirage SDK v1.2 Developer’s Guide
Introduction
DfMirage is “cutting edge" video driver mirroring technology for the Windows NT OS family. It is a driver for a virtual video device managed at the DDML level of the graphics system that exactly mirrors the drawing operations of one or more physical display devices. A detailed explanation of how a mirroring video driver works may be found in the Windows DDK. Display mirroring technology is widely employed by remote desktop applications such as: NetMeeting, PC Anywhere, VNC, Webex, etc. Mirroring is a technically superior method when compared to primitive screen grabbing, because it allows the capture of only the minimally updated regions and retrieves the data directly, bypassing the intermediate copy. Using the DfMirage driver solves the problem of reliably and efficiently detecting modified areas on the screen. This driver may be used transparently with office, CAD and other types of business and utility applications. An example is the open-source TightVNC application which uses the DfMirage driver with great success.
Operating environment and setup
The DfMirage driver is targeted to the Microsoft Windows NT OS family, which includes Windows NT4, Windows 2000, Windows XP/2003 and future versions of Windows as well. It has been tested on Windows 2000 service packs 0-4, and Windows XP service packs 1 and 2. Windows NT4 service pack 6 is also supported (driver package available upon request).
Installation
To setup the DfMirage driver, run MirageSetup.exe.
The underlying virtual display device uses plug-n-play technology, so the installation requires no reboot. When the driver is installed, it appears in device manager as shown in the following picture:
To ensure that the installation was successfull run dfstudio-mirage.exe (this is a special build of DemoForge Studio 2 which uses DfMirage driver internally) and the press the Record button. You may then locate the file just recorded in your %TEMP% folder. This file may be played back. It is rather large because it contains only raw, uncompressed data. If you compress it using any modern archival software, you will see a considerable difference in size. You may also save the recorded file to DML format to see the image data which has been captured by the mirage driver.
Uninstallation
The uninstallation of DfMirage is fully supported. The uninstaller is registered with the “Add/Remove Programs” applet of Control Panel.
Removing the device is a two-phase process. This is due to shortcomings within the Windows system video port driver. Microsoft has promised to resolve this issue with the release of Windows Vista. When the first phase is finished, the system must be rebooted. The 2nd phase will automatically begin following the reboot, when any member of the local Administrators group logs in. For a more detailed explanation of DfMirage setup considerations see the file “Mirage\Setup\Installer\Guide.en.doc”.
Operating principles
The DfMirage driver tracks the minimal areas of screen update and enables the client software to retrieve those updates directly (by means of the screen memory shared between the driver and application). This method provides excellent results in terms of traffic and CPU usage, while still maintaining its ease of use features. The Mirror driver follows the standard bi-component model of NT video drivers. That is, it uses miniport and display driver modules. Miniport is a low-level component. It represents a virtual video device. It is loaded by the OS IO manager and remains in memory until the OS is terminated. The miniport device is inaccessible to user-mode code. The display driver is loaded and unloaded on demand by changes in video mode. Basically, when it is loaded and running, video mirroring takes place. The DfMirage driver maps its screen surface into the user-mode application’s virtual memory space. Normally, the format and size of the buffer corresponds exactly to the format and size of the primary screen surface. This is not the case when its color format is overridden – (there is available, a registry-based switch for the driver to enforce the fixed color depth of the mirrored screen surface). The mapping of memory is performed via a file mapping object (or a “Section” using NT kernel slang). In this way an application “sees” everything that’s being drawn on the mirrored screen surface, resulting in the application’s ability to perform a direct copy of the modified screen area. It has but read-only access to this buffer.
The driver provides an application with the ability to retrieve ONLY those areas, which are modified, in spite of this, the entire screen buffer is always available to the application for read access.
The user-mode application communicates with the DfMirage driver via the ExtEscape() Windows API function. ExtEscape serves the role of an extensible escape hatch for the Windows GDI as it allows the passing of custom or non-standard requests to the video driver.
There are a number of private escape codes defined by DfMirage. Escape function codes and input/output structures are declared in display-esc.h (This is an interface header shared among the driver and application code modules).
The API
Following is a brief description of this API:
ESCAPE CODE: dmf_esc_usm_pipe_map
INPUT:None
OUTPUT:struct GETCHANGESBUF
FUNCTION:Creates a mirror screen and updates queue mappings for the calling process.
RETURN VALUES:If the function succeeds, the return value is greater than zero.
The buffers are mapped upon receiving this escape request. The output buffer format for the request is GETCHANGESBUF. GETCHANGESBUF::Userbuffer is actually a pointer to the mirror screen surface view. GETCHANGESBUF::buffer points to the queue of modified rectangles. The screen surface is “top to bottom, so the 1st line of pixels in the screen surface commences with the 1st byte of screen memory. For example, you can address the 2nd line of pixels by adding a fixed positive value (stride, also known as a pitch or delta) to the address of the 1st line. The userbuffer value actually points to the line 0. Surface dimensions correspond to the dimensions of the primary display. The stride of the surface is DWORD-aligned upwards. (The stride is calculated as follows: (screen_width*(screen_bits_per_pel>3)+3)&-4.)
ESCAPE CODE: dmf_esc_usm_pipe_unmap
INPUT:struct GETCHANGESBUF
OUTPUT:None
FUNCTION:Terminates the mappings of the shared memory structures.
RETURN VALUES:If the function succeeds, the return value is greater than zero.
There is a reciprocal unmap API - dmf_esc_usm_pipe_unmap. It is recommended that this API be used to terminate the capture under normal conditions. Under abnormal conditions it is best to unmap the shared memory buffers by calling UnmapViewOfFile(). The dmf_esc_usm_pipe_unmap/UnmapViewOfFile() call should always be accompanied by the DeleteDC() call for a previously-obtained mirror-device DC to release an instance of the driver.
(NOTE: this function is new, available in driver version 1.1+)
ESCAPE CODE: dmf_esc_qry_ver_info
INPUT:struct Esc_dmf_Qvi_IN
OUTPUT:struct Esc_dmf_Qvi_OUT
FUNCTION:Queries for driver’s version and determines if driver and application versions are compatible.
RETURN VALUES:If the function succeeds, the return value is greater than zero.
If the function succeeds, this means that driver and application are version compatible.
In Esc_dmf_Qvi_IN application passes it’s version information. See dmf-proto-version.h: set Esc_dmf_Qvi_IN::app_actual_version to DMF_PROTO_VER_CURRENT, and Esc_dmf_Qvi_IN::display_minreq_version to DMF_PROTO_VER_MINCOMPAT version number constants from SDK version it was built with. In Esc_dmf_Qvi_OUT driver returns information about its version. Esc_dmf_Qvi_OUT::display_actual_version is the driver’s version number. Esc_dmf_Qvi_OUT::app_minreq_version is a minimum application version number which is supported by the driver.
CHANGES_BUF is a circular queue of modified rect records (CHANGES_RECORD) (the current capacity setting is 20,000 records, which is normally far more than enough). The structure is defined as follows:
structCHANGES_RECORD
{
ULONGtype;
RECTrect;
RECTorigrect;
POINTpoint;
ULONGcolor;
ULONGrefcolor;
};
This definition is governed by historical reasons (legacy software compatibility). Not all fields in this structure are actually used with Mirage driver.
CHANGES_RECORD::type takes the following values:
typedef enum
{
dmf_dfo_SCREEN_SCREEN= 11,
dmf_dfo_BLIT= 12,
dmf_dfo_TEXTOUT= 18,
dmf_dfo_Ptr_Engage= 48,
dmf_dfo_Ptr_Avert= 49,
dmf_dfn_assert_on= 64,
dmf_dfn_assert_off= 65,
}
dmf_UpdEvent;
The actual screen update events are dmf_dfo_SCREEN_SCREEN, dmf_dfo_BLIT, dmf_dfo_TEXTOUT. Records dmf_dfo_Ptr_Engage and dmf_dfo_Ptr_Avert are mouse pointer status events. Engage shows (“checks”) the pointer in a position specified, whereas Avert hides the pointer from screen. Pointer coordinates go through CHANGES_RECORD::point field.
Records dmf_dfn_assert_on and dmf_dfn_assert_off notify the application about the driver’s status.
dmf_dfo_BLIT event is the most common update event type. dmf_dfo_TEXTOUT is results from text output. dmf_dfo_SCREEN_SCREEN is a screen-to-screen copy BitBlt.
Applications should disregard origrect, color, refcolorfields.
All update types (range dmf_dfo_SCREEN_SCREEN..dmf_dfo_TEXTOUT) may ultimately be treated as BitBlt. The only relevant and valid field of CHANGES_RECORD is then CHANGES_RECORD::rect. It is the screen destination update rectangle (a rectangular boundary of an actual update region) of the above-mentioned graphic operations.
dmf_dfo_SCREEN_SCREENevent type is supported in Mirage driver version 1.2+. To allow this new event type it must be enabled (see below). dmf_dfo_SCREEN_SCREEN may be used as an additional optimization option (it may be useful in cases of window drag, scroll, etc.) In addition to CHANGES_RECORD::rect it has valid CHANGES_RECORD::point field – the source point of screen copy operation.
The screen of Mirage driver is not divided into a fixed set of predefined rectangles. The update rectangles are placed, sized and ordered exactly in the same manner that the underlying graphic operations were.
The update rect coordinates are relative to the driver’s attach point on the desktop. If the driver is attached at (0,0) (as usual), rects are in a desktop coordinate space. The head of the buffer is moved cyclically forward when a newly-modified rect arrives (the queue constantly wraps through its end to the beginning). There is an ever-increasing counter of records: The ULONGCHANGES_BUF::counter. Using this counter, one may determine whether there are any newly-modified rects available. When the application detects the newly-modified rectangles, it has an opportunity to retrieve those rectangles directly from the mapped screen buffer memory. The application then makes copies of these modified rects and performs the required process on them (for example, sending them over wire).
If an application detects that it has lost one or more modified rects, (possible when the system is sorely over-taxed) nothing is really broken. The application should simply skip any unprocessed rects and then perform an update of the entire screen . The entire contents of the modified rects buffer is available for read from the application at any time.
It is important to understand that shared memory structures remain available to the application despite any change to the video mode when the driver is unloaded completely or just disabled temporarily. The shared memory objects are managed by 2 references: from the driver and from the application. These are unavailable when, and only when, both the driver and application release them. This is by design, to ensure reliability in the case of an unexpected disconnection with the client application. This is why it is important to perform the cleanup before reconnection.
To start working with the driver, the application must first activate it.
The driver is set to enabled (“Attach.ToDesktop” value in system registry). The video mode change is then performed and the new activation setting for the driver takes effect.
When the application is finished with the driver, it is deactivated in a similar manner. The driver will remain active, consuming CPU and memory resources until it is unloaded.
An error will not result from an attempt to load or unload the driver more than once. Only one instance of the driver will be loaded globally during a given interactive logon session.
While there is no technical limit on the number of applications which may simultaneously read the mirror screen, only one of them should attempt to manage the driver.
The Mirage driver has a boot-time checker that ensures that the driver is deactivated even when the managing application has failed to deactivate it (for example, when the application has crashed).
The driver is never loaded (and can’t be activated) in the Safe (or VGA) boot mode of Windows.
Options
There are some registry-based switches that affect the driver’s behavior.
All switches are derived from the values located in one of the following registry keys:
[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\dfmirage\Device0] or
[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Hardware Profiles\Current\
System\CurrentControlSet\SERVICES\dfmirage\DEVICE0]
These supported switches are:
“Pointer.Enabled”=dword: 0 (default) or 1.
First (simple) mode (Pointer.Enabled=1) just renders the mouse pointer movements and animation and produces only simple update events from these. Second mode (Pointer.Enabled=0) excludes the mouse from the screen buffer, posting only position and status events.
There are 2 types of events that bypass the “update buffer” regarding the mouse pointer capture:
- 48(dmf_dfo_Ptr_Engage): show the pointer (cursor) [in a position specified]
- 49(dmf_dfo_Ptr_Avert): hide the pointer.
The current pointer shape is queried via a special call.
An application can disable mouse pointer rendering if it wants to draw it on its own.
“Screen.ForcedBpp”=dword: 0 (default), 16, 24 or 32.
This switch forces the mirror screen color depth. 0 enables the native color depth of a primary display device to be used for a mirror screen. There are 3 valid values of color depth that may be enforced – 16, 24 and 32 bits per pixel. No values other than 0, 16, 24 and 32 are allowed.
NOTE that 16 bits per pixel forced color depth option is available only in driver version 1.2+.
Color depth enforcement may sometimes be useful. Through the use of this feature, an application may implement only true color encoding, resulting in its not having to deal with palettes or worry about color depth switching. Nevertheless, it should still be noted that native 8-bit or 16-bit screen images are more compact.
“Cap.DfbBackingMode”=dword: 0, 2, or 3 (default).
This is the powerful switch of screen memory backing mode.
Backing mode 3 (the default and recommended option) results in an “opaque” mirror screen. That is -- a screen that cannot be accessed by GDI directly. This means nothing drawn on the mirror screen escapes the notice of the driver. This option guarantees the maximum mirroring consistency possible in the Windows GDI.
Backing mode 0 disables memory backing of the mirroring screen altogether. The mirroring screen cannot be accessed by the GDI and applications either. Beware, GETCHANGESBUF::Userbuffer is NULL in this mode. Memory is not allocated and no rendering is performed. The driver in this mode only collects update regions. This mode may be useful for applications that bypass the driver when retrieving the screen image.
Backing mode 2 is somewhat experimental. The mirror screen is transparent to the GDI in this mode. Rendering is on and screen image is accessible. When the screen is transparent, the GDI is not always accurate when using the driver’s hooks to perform the rendering, and screen mirroring consistency may become compromised under certain circumstances.
Still, some customers claim that certain graphic-intensive applications such as video players perform significantly faster in this mode. This mode is not recommended unless you know exactly what you’re doing.
The following options are new to Mirage driver version 1.2+:
“Order.BltCopyBits.Enabled” = dword: 0 (default) or 1.
Enables or disables generation of dmf_dfo_SCREEN_SCREEN update events in case of screen-2-screen BitBlt.
“Cap.Bootup” = dword: 0 (default), 1 or 2.
Controls “boot-time checker” which normally checks if the driver is unintentionally attached to screen during the system boot and disables the attachment. Value 0 corresponds to usual behavior. Value 1 enables for 1-time (this boot only) driver attach from system boot. Value 2 (and higher) suppresses the check at all, thus the driver will always be attached to screen right from the system startup.
This switch is generally useless with modern (Windows 2000+) OSes because applications should load and unload the driver dynamically. If it is anyway needed (for instance, on NT4), use it with caution.
Performance considerations
Screen grabbing with the DfMirage driver is as simple as performing a direct memory copy of the modifications made to the framebuffer. The driver provides direct read access to the screen framebuffer for the client application. The DfMirage driver enables a far more efficient screen capture method when compared to the usual way of blitting from the screen DC. Using BitBlt from the screen requires at least one syscall (BitBlt), a global screen update lock performed internally by BitBlt plus an extra memory copy operation. The DfMirage driver requires no syscalls or screen locks.
In principle, you can grab the screen, or any part of it when you need or want to, resulting in flexible performance and instant on-the-fly tuning, for example, to accommodate high-stress conditions).
The application receives prompt notifications of regions which have been modified, but they may be allowed to accumulate if the updates are occurring faster than desired. They also may be accumulated for short periods simply to minimize the traffic. You are not compelled to keep up with the real screen video driver, but you can. It is very flexible.
Updates are often small, such as those resulting from a single mouse pointer motion. The application only needs to grab and update the small area behind the pointer. In this case there is very little CPU consumption. But from time to time of course, there will be major updates which will then need to be handled.
The present driver’s API requires a poll of the CHANGES_BUF::counter. As a matter of fact, the poll in itself doesn’t consume much in the way of CPU resources. Yes, an Event could be used to check for new modifications, but as strange as it may seem, the poll method is much faster and provides much better performance. To illustrate:
VNC applications typically prefer to accumulate the modifications every, 100 ms or so or even for a longer period. It may simply use a 100ms timer and process all of the modifications discovered (which may overlap each other) at once. The poll scenario therefore, requires no direct interaction with the driver at all. Now imagine the extremely taxing demands of synchronous notification when an event object is used. The application is interrupted for every single update! For example, a mere mouse pointer motion may produce tens or even hundreds of events per second. More importantly, each update event must be individually confirmed to reset the raised notification event object. This is rarely considered desirable behavior! Note that we now have an extra confirmation “syscall” from the application to the driver. This may introduce notable latency under high load conditions. This also makes it more difficult to accumulate the updates. According to our own observations, we have concluded that the poll method is typically more CPU efficient than event-based synchronization whether under low or high load conditions.