A VNC remote desktop server for vintage Macintosh computers, including the Mac Plus.
MiniVNC is a remote desktop server that has been written from the ground up for best performance on 68k Macintosh computers.
It was originally an experiment to see whether a Macintosh Plus could be controlled remotely, but has since been expanded to support the Apple Lisa and also all vintage color Macs! :rainbow:
MiniVNC v0.9 used TRLE encoding (Tiled Run-Length Encoding) and paletted colors. While this encoding provides the best performance on B&W compact Macs, client support is very limited. After v1.1, MiniVNC was overhauled to supports ZRLE and Hextile encodings for maximum client compatibility, but these encodings are limited to color-capable Macs.
Here are the server capabilities:
Server | OS | TRLE | ZRLE | Hextile | Uploads | Notes |
---|---|---|---|---|---|---|
MiniVNC v1.4 | macOS 7 | Yes | Yes | Yes | Yes | Compatible with all tested clients |
MiniVNC v1.3 | macOS 7 | Yes | Yes | Yes | No | Compatible with all tested clients |
MiniVNC v1.2 | macOS 7 | Yes | Yes | Yes | No | Crashes TurboVNC |
MiniVNC v0.9 | macOS 7 | Yes | No | No | No | Only works with RealVNC Viewer |
Here are the client capabilities:
Client | OS | TRLE | ZRLE | Hextile | Palette | Uploads | Notes |
---|---|---|---|---|---|---|---|
RealVNC Viewer | Windows | Yes | Yes | Yes | Yes | No | Only viewer for B&W Macs |
TightVNC | Windows | No | Yes | Yes | No | Yes | Requires MiniVNC 1.2+ |
TurboVNC | Windows | No | Yes | Yes | No | No | Requires MiniVNC 1.3+ |
TigerVNC | Windows | No | Yes | Yes | No | No | Tested w/ MiniVNC 1.3+ |
MacVNC | macOS 7 | No | No | Yes | No | No | Requires MiniVNC 1.2+ |
VNCThing 2.2 | macOS 8 | No | No | Yes | No | No | Requires MiniVNC 1.2+ |
Screen Sharing | macOS X | No | Yes | No | No | No | Requires MiniVNC 1.2+ |
MiniVNC is built on MacTCP and requires System 7, but it will operate on later Macs using Open Transport. MiniVNC has been developed and tested using a RaSCSI device operating as an Ethernet bridge, but should also work using a Mac with a built-in Ethernet port.
As a special reward to GitHub sponsor at the "Sustaining Supporter" tier or above, I am providing a 🎁 bonus archive in the releases page with additional perks. These perks are:
Please refer to your sponsorship welcome letter or GitHub sponsor dashboard for the archive password.
You can help this project in one of the following ways:
The goal of MiniVNC is to provide better performance and compatibility with older Macs than ChromiVNC
It accomplishes these goals by:
Development of MiniVNC required months of experimentation and hacking. The following write-up provides details for developers who have an interest.
MacTCP is Apple's first TCP/IP networking stack and is the only
networking API available on the Macintosh Plus. The
MacTCP Programmer's Guide is a good resource, but lacks code
samples and makes no mention of high-level languages. The
MacTCP Cookbook article by Steve Falkenburg provides more meat
to chew on but actual source code is worth a thousand words. I
finally it on the Apple Developer Group CD Volume VII in the
form of a "finger" protocol example in the directory
Dev.CD Vol. VII:develop:develop 6 code:TCP:finger
This is a good starting point as the "TCPRoutines.c" file provides an example of using MacTCP parameter blocks from a high-level language. Steve's helper routines, while easy-to-use for simple tasks, are very slow. As I later learned, the most efficient way to do MacTCP programming is via asynchronous callback routines.
Since these callback routines execute in interrupt time, writing them in a high-level language is challenging. I used the technique from the article "Asynchronous Routines on the Macintosh" in Develop magazine, March 1993 which involves an assembly language glue routine. Later on, I used a similar routine for making a Vertical Retrace task for gathering screen updates.
Today we take for granted threads which make it easy to implement network applications. The use of callback routines is a huge step backwards.
For one, it precludes using temporary stack-based storage and instead all state must be stored in global variables. Doing things like loops are trivial in a thread but very difficult using callback routines.
The VNC protocol is fairly simple, but the code is far more complicated than it would have been had I modern techniques at my disposal.
As an aside, Ari Halberstadt wrote a very promissing thread library for the Macintosh. Getting it to work with MacTCP might have simplified the programing model, but at the time it was too much of a heavy lift for me to get it to work while I was also learning MacTCP. I ended up using some of his basic OS utilities code in MiniVNC, but not the thread library itself.
On the Macintosh, the mouse can be progmattically changed by writing
the new position to the low-memory globals MouseTemp
and
RawMouseLocation
and then copying the value of CrsrCouple
to
CursorNew
to signal the change. This technique is borrowed from
ChromiVNC.
Mouse button control presented a challenge. The technique used by
ChromiVNC is to write the new mouse button state to the low-memory
global MBState
while simultaneously posting a mouseUp
or mouseDown
event to the System event queue. This works on modern Macs, but on the
Macintosh Plus it only works for clicking. Mouse dragging — and
crucially, menu selection — does not work. On that machine the
VIA interrupt in ROM constantly overwrites MBState
with the button
state from the physical mouse, so the trick of writing a value to
MBState
does not work.
I attempted patching the Button
trap and implemented a journaling driver,
but the former was ineffective while the latter was found to be broken and
unusable under System 7's multi-tasking model.
At last, an analysis of the disassembled code for the VIA interrupt
revealed it deglitched the mouse button by waiting three ticks prior to
updating MBState
. The low-memory variable MBTicks
indicates the start
of the wait period and by periodically setting it to the future &emdash;
in advance of TickCount
— I found I could keep the VIA routine
waiting indefinitely so that I could alter MBState
at will. This
allowed me control of the mouse button on all Macs, including the Macintosh
Plus.
On the vintage Macintosh, the address of the framebuffer is stored in the
low-memory global ScrnBase
. A crucial part of a VNC server is detecting
changes to the screen. While this could be done by maintaining a separate
copy of the framebuffer and comparing every pixel, this requires a lot of
memory and a lot of memory accesses.
In MiniVNC, I decided to compute a 32-bit sum across each row of pixels and a 32-bit sum up and down the screen. For both the horizontal and vertical sums, the new and old sums are compared to detect screen changes. It turns out this is also an quick way to determine the bounds of the change rectangle as the location of the first and last changed sum along an axis can be taken as the rectangle bounds along the axis.
This is an inexact approach which makes the server blind to many types of screen changes. Using a [position-dependent checksum] like a Fletcher's checksum would improve accuracy, at the expense of more computation and storage.
An ordinary VNC server would attempt to minimize transmission by sending the smallest possible change rectangle. MiniVNC keeps the horizonal bounds of change rectangles aligned to byte boundaries. On the Macintosh Plus, which uses one bit per pixel, this is crucial as it allows screen data to be sent without any bit-shift or bit masking operations, which are particularly slow on the 68000.
When using a 32-bit column sum (as opposed to the XOR operation used in my first implementation), a change in one pixel column can cause a carry to the next, meaning that the horizontal bounds of change rectangles are further aligned to 32-bit boundaries.
On a B&W display, this causes change rectangle bounds to fall on 32 pixel increments, which might cause quite a lot of extra data to be transmitted.
The effect is minimized on color displays, which pack fewer pixels per byte. For example, on a 256 color display, the change rectangles can fall on 4 pixel boundaries, mimimizing the size of change rectangles.
The VNC protocol is meant for color computers and most encodings assume a color depth of at least eight bits. However, the TRLE encoding is unique in that it supports a paletted tile type that allows for a 1-bit, 2-bit and 4-bit encodings.
The ability to transmit 1-bit data is what caused me to require MiniVNC to use TRLE encoding exclusively. This does not meet the VNC specifications, but is compatible with modern clients. To support raw encoding, as required by the specifications, would necessitate expanding each 1-bit pixel in a byte into a full color byte, an operation which would be prohibitive on older Macs.
On the Macintosh Plus, I transmit all tiles using the 1-bit paletted tile type. This, together with byte alignment, allows the encoding process to be a straight copy full bytes without any bit shifting or masking.
As an optimization, during the copy I will detect if a tile consists of all zeros (i.e. white). If this is the case, I emit a solid white tile. There is no corresponding case for black pixels, so a white region will compress better than a black region.
On color Macintosh computers, where I have the luxury of a 68020 with an instruction and data cache, I perform additional computation in order to choose from among the following TRLE tile types:
On a color Mac, the full encoding process would work like this:
In practice, I found that doing this whole process actually hurt performance on an Macintosh LC II, so MiniVNC supports different packing levels which omit many of the steps listed above.
For example, packing level 2 looks like this:
The key different between packing 2 and the full process is that I do not emit a tile with fewer colors than the color depth of the screen itself. Doing so requires finding the unique colors in a tile and mapping those to a smaller palette, which is expensive. Emitting a tile with the same number of colors as the screen, however, is trivial as it only involves a straight copy with no changes to the color values themselves.
There is not, unfortunately, a perfect development environment for the Macintosh when it comes to writing code that mixes C++ and assembly language. My go to language used to be Symantec C++ 7 because it allows you to write C++ routines and add assembly language bits and pieces just where you need it, minimizing the learning curve. It is also very good for things like code resources or device drivers.
CodeWarrior 8, on the other hand, works great on Basillisk II and has a better IDE. I've lately begun using it over Symantec C++ 7 for these reasons, but one frustrating aspect is that it does not allow you to mix C and assembly language in one function. Instead, you must chose one or the other. This meant that in MiniVNC I had to write whole functions in assembly language, which meant doing things like pulling arguments off the stack myself, something Symantec C++ 7 would do for me.
The TRLE encoder was written in 68x assembly for best performance.
For the color encoders, a 68020 is assumed and the code makes special
use of 68020 instructions such as bfextu
. Most functions take
advantage of as many of the available eight data and eight address
registers as possible, as to minimize memory accesses.
One example is the color palette gathering code. In the RLE encoding routine, use two register halves of two registers to keep track of colors I have seen before, which allows me to generate a palette of up to four colors without having to write to memory. I implemented another routine which uses eight 32-bit registers as bitfields to tally up to 256 unique colors—again, with no memory access. This routine would be needed for generating 16 bit color tiles when the screen is in 256 color mode, although I have not actually implemented this at this point (and probably never will, as it appears that even generating 2 or 4 color tiles is too slow to be worth the effort).