Tuesday, October 11, 2011

SCSI support and a big surprise

Last week I added SCSI disk support for the CD-i 60x extension board to CD-i Emulator. It took somewhat longer then I expected, though. This was mostly because the DP5380 SCSI controller chip exposes most low-level details of the SCSI protocol to the driver which means that all of these details have to be emulated.

The emulation ended up to be a more-or-less complete software implementation of the parallel SCSI-2 protocol, including most of the low-level signaling on the BSY, SEL, ATN, MSG, C/D-, I/O-, REQ and ACK lines. This is all implemented by the new CScsiBus class representing the SCSI bus that connects up to 16 instances of the CScsiPort class that each represent a single SCSI-2 bus interface. I was able to mostly avoid per-byte signaling of REQ and ACK if the target device implementation supports block transfers, a big performance win.

The new CCdiScsiDevice class emulates the DP5380 controller chip, working in conjunction with the CCdiScsiRamDevice and CCdiScsiDmaDevice classes that emulate the 32 KB of local extension SRAM and the discrete DMA logic around it that are included on the CD-i 60x extension board.

The CD-i 182 extension uses a compatible SCSI controller chip but a different DMA controller and has no local extension SRAM. I have not yet emulated these because I have almost no software to test it.

The new CScsiDevice class implements a generic SCSI device emulating minimal versions of the four SCSI commands that are mandatory for all SCSI device types: TEST UNIT READY, REQUEST SENSE, INQUIRY and SEND DIAGNOSTIC. It implements most of the boiler-plate of low-level SCSI signaling for target devices and the full command and status phases of SCSI command processing, allowing subclasses to focus on implementing the content aspects of the data transfer phase.

The CScsiFile class emulates a SCSI device backed by a file on the host PC; it includes facilities for managing the SCSI block size and the transfer of block-sized data to and from the backing file.

The CScsiDisk and CScsiTape classes emulate a SCSI disk and tape device, respectively, currently supporting a block size of 512 bytes only. Instances of these classes are connected to the SCSI bus by using the new
-s[csi]d[isk][0-7] FILE and -s[csi]t[ape][0-7] FILE options of CD-i Emulator.

The CD-i 60x extension board normally uses SCSI id 5; the built-in ROM device descriptors for SCSI disks use SCSI ids starting at zero (/h0 /h1 /h2) while the built-in device descriptor for a SCSI tape uses SCSI id 4 (/mt0). This means that the useful options with the 60x are -scsidisk0, -scsidisk1, -scsidisk2 and -scsitape 4.

I've added the new dsk subdirectory to contain disk images; tape images have no standard location as they are mostly intended for bulk-transfer purposes (see below).

Inside the CD-i player this leads to the following response to the built-in inquire command:
$ inquire -i=0
vendor identification:"CDIFAN CDIEMU SCSIDISK "

$ inquire -i=4
vendor identification:"CDIFAN CDIEMU SCSITAPE "
where the "CDIFAN " part is the vendor name and the "CDIEMU SCSIXXXX " part is the product name.

In the previous post I described a 450 MB OS-9 hard disk image that I found on the Internet. After mounting it with
-scsidisk0 mw.dsk I got the following output:
$ free /h0
"MediaWorkshop" created on: Feb 17, 1994
Capacity: 1015812 sectors (512-byte sectors, 32-sector clusters)
674144 free sectors, largest block 655552 sectors
345161728 of 520095744 bytes (329.17 of 496.00 Mb) free on media (66%)
335642624 bytes (320.09 Mb) in largest free block

$ dir -d /h0

Directory of /h0 23:49:36
ASU/ AUDIO/ CDI_BASECASE/ CINERGY/ CMDS/
COPY/ CURSORS/ DEFS/ DEMOS/ ENET/
ETC/ FDRAW/ FONTS/ FontExample/ ISP/
LIB/ MAUI/ MAUIDEMO/ MENU/ MWOS/
NFS/ README_CIN README_MWS SCRIPT/ SHARE/
SHIP/ SYS/ T2D_RUNTIME/ TEMP/ TEMPMARK/
TEST/ USR/ VIDEO/ abstract.txt bibliographic.txt
bkgd.c8 bkgd.d cdb cdb1 cdb2
cdi_opt_install chris_test cin copyright.mws copyright.txt
csd_605 custominits_cin delme dos/ file
font8x8 get globs.mod go go.mkfont
inetdb ipstat kick1a_f.c8 kick2a_f.c8 mtitle
mws net new_shell new_shell.stb scratch
screen startup_cin thelist
You can see why thought it was a MediaWorkshop disc, but on closer inspection this turned out to something quite different. Some basic scrutiny lead to the hypothesis that this is probably a disk backup of someone from Microware working on early development of the DAVID (Digital Audio Video Interactive Decoder) platform. There are various surprises on the disk which I will describe below.

Anyway, I wanted to transfer the contents to the PC as a tar archive, similar to the procedure I used for my CD-i floppy collection. After starting CD-i Emulator with a -scsitape4 mw.tar option this was simply a matter of typing the following into the terminal window:
tar cb 1/h0
This command runs the "tape archiver" program to create a tape with the contents of the /h0 directory, using a tape blocking size of 1 (necessary because my SCSI tape emulation doesn't yet support larger block sizes). The resulting mw.tar file on the PC is only 130 MB, not 450 MB which indicates that the disk is mostly empty. At some point I might use an OS-9 "undelete" program to find out if there are additional surprises.

Extracting the mw.tar file was now a simple matter of running the PC command
tar xvf mv.tar
This produced an exact copy of the OS-9 directory structure and files on the PC.

Many of the directories on the hard disk are clearly copies of various distribution media (e.g. CDI_BASECASE, CINERGY, CURSORS, ENET, FONTS, ISP, MWOS, NFS). The contents of the ENET, ISP and NFS directories at first appear to match some of my floppies, including version numbers, but on closer inspection the binaries are different. Running some of them produces "Illegal instruction" errors so I suspect that these are 68020 binaries.

The SHIP directory contains some prerelease RTNFM software; the readme talks about PES which is a type of MPEG-2 stream (Packetized Elementary Stream). Various asset directories contain versions of a "DAVID" logo.

The CMDS directory contains working versions of the Microware C compiler, identical to the ones I already had and also many other programs. It also contains some "cdb" files (configuration database?) that mention the 68340 processor.

The contents of the CMDS/BOOTOBJS directory produced a first surprise: it contains a subdirectory JNMS containing among others files named "rb1793" and "scsijnms". Could this be floppy and SCSI drivers for the CD-i 182 extension, as it contains with a 1793 floppy drive controller (the CD-i 60x uses a different one) and the player has a "JNMS" serial number?

Well, yes and no. Disassembly of the scsijnms file proved it to be compiled C code using an interface different from OS-9 2.4 drivers, so I suspect this is an OS-9 3.x driver. In any case, I cannot use it with the stock CD-i 180 player ROMs. Bummer...

And now for the big surprise: deeply hidden in a directory structure inside the innocently named COPY directory is the complete assembly source for the VMPEG video driver module "fmvdrv". At first glance it looked very familiar from my disassembly exercises on the identically-named Gate Array 2 MPEG driver module "fmvdrv", which is as expected because I had already noticed the large similarity between these two hardware generations.

The source calls the VMPEG hardware the "IC3" implementation, which matches CD-i digital video history as I know it. The Gate Array MPEG hardware would be "IC2" and the original prototype hardware would be "IC1". Furthermore, the sources contain three source files named fmvbugs1.a to fmvbugs3.a whose source file titles are "FMV first silicon bugs routines" to "FMV third silicon bugs routines". The supplied makefile currently uses only fmvbugs3.a as is to be expected for a VMPEG driver.

The fmvbugs1.a source contains some of the picture buffer manipulation logic that I've so far carefully avoided triggering because I couldn't understand it from my disassemblies, and this is now perfectly understandable: they are workarounds for hardware bugs!

As of two hours ago, I have verified that with a little tweaking and reconstruction of a single missing constants library file these sources produce the exact "fmvdrv" driver module contained in the vmpega.rom file directly obtained from my VMPEG cartridge.

In general these sources are very heavily commented, including numerous change management comments. They also include a full set of hardware register and bit names, although no comments directly describing the hardware. This should be of great help in finally getting the digital video emulation completely working.

All of the comments are English, although a few stray words and developer initials lead me to believe that the programmers were either Dutch or Belgian.

Disassembly comparisons lead me to the conclusion that careful undoing of numerous changes should result in exact sources for the GMPEGA2 driver module "fmvdrv" as well. I might even do it at some point, although this is not high priority for me.

The disk image containing all of these surprises is publicly available on the Internet since at least 2009, which is probably someone's mistake but one for which I'm very grateful at this point!

No comments:

Post a Comment