Gavare's eXperimental Emulator: Miscellaneous

Miscellaneous

Writing operating system code, or developing firmware, using GXemul
Using GXemul in compiler contruction courses
How to start the emulator with a disk image
How to start the emulator with tape images
How to use disk image overlays
Transfering files to/from the guest OS
How to extract large gzipped disk images
Running userland binaries
Using a PROM dump from a real machine

Writing operating system code, or developing firmware, using GXemul:

Is this a good idea? The answer is yes and no, depending on the level of detail you need in your simulations. If you are developing an operating system or operating system kernel of your own, then the emulator can be a complement to testing on real hardware.

Important things to keep in mind:

Porting code to a specific machine mode, e.g. a Silicon Graphics machine, using GXemul, will not "magically" cause the code to work on a real machine. Sometimes code works in GXemul which doesn't work on real hardware, sometimes it's the other way around.
GXemul contains bugs, and many things are not yet implemented.
Very important! I have only implemented devices in GXemul to the degree that NetBSD, OpenBSD, Linux, etc don't complain too much.
If you are developing a driver for a device which is emulated by GXemul, and your driver does not seem to be working, then the probability of a bug in GXemul's implementation of the device is very much higher than that of a bug in your driver.
The device implementations in GXemul are based on the assumption that the emulated OS is already developed and bug-free. They are not primarily intended to be used for development of new device driver code in operating systems, so if you do that, then be prepared for bugs and inconsitencies.
CPU details in GXemul are usually wrong. If your code depends on, say, R10000 or MIPS64 specifics, chances are that GXemul will not be sufficient. One example is different revisions of ISAs; some instructions which should trigger an exception on a real MIPS processor usually execute anyway in GXemul. Another example is if userland code tries to access kernel memory; in some cases there is protection against this, but not in all cases (to get higher performance).
Caches. There is no cache emulation in GXemul right now. Caches for R2000/R3000 are faked well enough to run NetBSD, Ultrix, etc in the DECstation emulation mode, but other than that, cache operations are treated as nops.

The bottom line is that GXemul can be useful as yet another way to test your code during development, but it should not be fully relied on.

Using GXemul in compiler contruction courses:

If you are learning how to write a compiler, and wish to target a realistic target platform, then MIPS or ARM (as emulated by GXemul) might be suitable choices.

(+) Your compiler needs to output real assembly language code, which the assembler (e.g. gas, the GNU assembler) can then compile into object format, and then you need to link this into an executable image. This is much closer to how things work in real life than running assembly language listings in a simulator (e.g. SPIM).
(-) GXemul does not simulate out-of-order execution, penalties related to instruction scheduling, or load-delays, so it cannot be used to create optimizing compilers that take advantage of such processor features. GXemul keeps track of the number of instructions executed, and that's it.

How to start the emulator with a disk image:

Add -d [prefixes:]diskimagefilename to the command line, where prefixes are one or more single-character options. Run gxemul -h to get a list of possible options.

Here are some examples. If you want to run a NetBSD/pmax kernel on an emulated DECstation machine, you would use a command line such as this:

	$ gxemul -e 3max -d pmax_diskimage.fs netbsd-pmax-INSTALL

NOTE: For some emulation modes, such as the DECstation mode, you do not actually have to specify the name of the kernel, if the disk image is bootable!

It is possible to have more than one disk. For each -d argument, a disk image is added; the first will be SCSI target 0, the second will be target 1, and so on, unless you specify explicitly which ID number the devices should have.

	$ gxemul -e 3max -d disk0.raw -d disk1.raw -d 5:disk2.raw netbsd-pmax-INSTALL

Note: In the example above, disk2.raw will get scsi id 5.

If a filename has a 'c' prefix, or ends with ".iso", then it is assumed to be a CDROM device (this can be overridden with a 'd' prefix, to force a read/write disk). For example, the following command would start the emulator with two CDROM images, and one harddisk image:

	$ gxemul -e 3max -d image.iso -d disk0.img -d c:second_cdrom.img netbsd-pmax-INSTALL

Usually, the device with the lowest id becomes the boot device. To override this, add a 'b' prefix to one of the devices:

	$ gxemul -e 3max -d rootdisk.img -d bc:install-cd.iso name_of_kernel

If you have a physical CD-ROM drive on the host machine, say /dev/cd0c, you can use it as a CD-ROM directly accessible from within the emulator:

	$ gxemul -e 3max -d rootdisk.img -d bc:/dev/cd0c name_of_kernel

It is probably possible to use harddisks as well this way, but I would not recommend it.

How to start the emulator with tape images:

Using emulated tape drives is a bit more complicated than disks, because a tape can be made up of several "files" with space in between. The solution I have choosen is to have one file in the host's file system space for each tape file. The prefix for using tapes is 't', and the filename given is for the first file on that tape (number zero, implicitly). For files following file nr 0, a dot and the filenumber is appended to the filename.

As an example, starting the emulator with

	-d t4:mytape.img

will cause SCSI id 4 to be a tape device, using the following file number to name translation scheme:

File number:	File name in the host's filesystem:
0	mytape.img
1	mytape.img.1
2	mytape.img.2
..	..

If you already have a number of tape files, which should be placed on the same emulated tape, then you might not want to rename all those files. Use symbolic links instead (ln -s).

There is another advantage to using symbolic links for tape filenames: every time a tape is rewound, it is reopened using the filename given on the command line. By changing what the symbolic name points to, you can "switch tapes" without quiting and restarting the emulator.

Note: Tape support is most likely very buggy, because it has not been tested much, and has probably also suffered from bit-rot by now.

How to use disk image overlays:

This is most likely best understood by an example:

Install e.g. NetBSD/cats. You will end up with a disk image called nbsd_cats.img.
Running the following command will boot straight from the disk image, with no overlay images:
```
	gxemul -XEcats -d nbsd_cats.img netbsd.aout-GENERIC.gz
```
You may now create an overlay file, a corresponding map file, and start the emulator with the overlay image connected to the same (explicit) ID as the base disk image:
```
	touch overlay.img overlay.img.map
	gxemul -XEcats -d 0:nbsd_cats.img -d V0:overlay.img netbsd.aout-GENERIC.gz
```
Any changes to the filesystem you perform when using the overlay will only be written to that overlay. For example, to perform a "roll back", you can do the following:
```
	rm -f overlay.img overlay.img.map
	touch overlay.img overlay.img.map
```
and then simply start the emulator again, with the newly created overlay image.

It is also possible to add multiple overlays. In that case, writes always go the the last added overlay.

GXemul uses Unix' way of supporting files with "holes", so even if ls -l overlay.img says that the overlay is several gigabytes large, du overlay.img should reveal that only the blocks that have actually been written to have been stored in the overlay, e.g.:

	$ ls -l
	..
	-rw-r--r--  1 debug  wheel  3072319488 Mar 24 11:59 nbsd_cats.img
	-rw-r--r--  1 debug  wheel     2465354 Mar 24 11:44 netbsd.aout-GENERIC.gz
	-rw-r--r--  1 debug  wheel  2930841600 Mar 24 14:02 overlay.img
	-rw-r--r--  1 debug  wheel      715538 Mar 24 14:02 overlay.img.map
	$ du overlay.img
	864     overlay.img

The .map file is simply a raw bitmap telling which blocks of the overlay file that are in use.

Transfering files to/from the guest OS:

If the emulated machine supports networking (see this section for more info), then the easiest way to transfer files is probably to use FTP or similar methods.

There is another way of transfering files which works for any kind of emulated machine which supports disks (either SCSI or IDE). Any file can be supplied as a disk image. For example, consider the following:

	$ gxemul -XEcats -d nbsd_cats.img -d archive.tar.gz netbsd-GENERIC

This will start NetBSD/cats with nbsd_cats.img as IDE master on controller 0 (wd0), and archive.tar.gz as IDE slave on controller 0 (wd1). From inside NetBSD, it is now possible to extract the files using the following command:

	(inside emulated NetBSD/cats)
	# tar zxvf /dev/wd1c

Don't worry if NetBSD complains about lack of disklabel; it doesn't matter. On some machines, NetBSD uses wd1d instead of wd1c for the entire disk. There is also a minor problem: reading the end of the disk image. If you experience problems untaring archives like this, then pad out the archive first with some zeroes.

Transfering files out from the emulated operating system to the host can be done the same way. First, prepare an empty archive file:

	$ dd if=/dev/zero of=newarchive.tar bs=1024 count=1 seek=10000

This example created a 10 MB empty file. Then, start the emulator like this:

	$ gxemul -XEcats -d nbsd_cats.img -d archive.tar netbsd-GENERIC

and transfer files by creating an archive directly onto the disk image:

	(inside emulated NetBSD/cats)
	# tar cvf /dev/wd1c filenames

where filenames are the files or directories to transfer.

How to extract large gzipped disk images:

Unix filesystems usually support large files with "holes". Holes are zero-filled blocks that don't actually exist on disk. This is very practical for emulated disk images, as it is possible to create a very large disk image without using up much space at all.

Using gzip and gunzip on disk images can be very slow, as these files can be multiple gigabytes large, but this is usually necessary for transfering disk images over the internet. If you receive a gzipped disk image, say disk.img.gz, and run a naive

	$ gunzip disk.img.gz

on it, you will not end up with an optimized file unless gunzip supports that. (In my experiments, it doesn't.) In plain English, if you type ls -l and the filesize is 9 GB, it will actually occupy 9 GB of disk space! This is often unacceptable.

Using a simple tool which only writes blocks that are non-zero, a lot of space can be saved. Compile the program cp_removeblocks in the experiments/ directory, and type:

	$ gunzip -c disk.img.gz | cp_removeblocks /dev/stdin disk.img

This will give you a disk.img which looks like it is 9 GB, and works like the real file, but the holes are not written out to the disk. (You can see this by running for example du disk.img to see the physical block count.)

Running userland binaries:

NOTE 2007-06-15: I just removed most of the userland (syscall) emulation support, and started a rewrite from scratch. The rest of this section in the documentation is not currently valid.

There is some skeleton code for running userland programs as well. This will not emulate any particular machine, but instead try to translate syscalls from e.g. NetBSD/pmax into the host's OS' syscalls. Right now, this is just a proof-of-concept, to show that it could work; there's lots of work left to do to make it actually run useful programs.

NetBSD/pmax:
Running /bin/hostname or /bin/date and similarly trivial programs from the NetBSD/pmax distribution works:

	$ gxemul -q -u netbsd/pmax pmax_bin_hostname
	tab.csbnet.se
	$ gxemul -q -u netbsd/pmax pmax_bin_date
	Sun Jan 25 02:26:14 GMT 2004
	$ gxemul -q -u netbsd/pmax pmax_bin_sleep
	usage: pmax_bin_sleep seconds
	$ gxemul -q -u netbsd/pmax pmax_bin_sleep 5
	$ gxemul -q -u netbsd/pmax pmax_bin_sync

Ultrix:
At least /bin/date and /bin/hostname work:

	$ gxemul -q -u ultrix ultrix4_bin_date
	UNIMPLEMENTED ultrix syscall 54
	UNIMPLEMENTED ultrix syscall 62
	Mon Feb  9 12:50:59 WET 2004
	$ gxemul -q -u ultrix ultrix4_bin_hostname
	tab.csbnet.se

Using a PROM dump from a real machine:

Raw PROM images from real machines can, in a few cases, be used in the emulator. ROM code is usually much more sensitive to correctness of the emulator than operating system kernels or userland programs are, so don't expect any PROM image to just magically work.

Dumping the PROM on a DECstation 5000/125:

The image first needs to be extracted from the machine. There are several ways to do this.

Use hardware to read the PROM chip(s) directly. Not easy if you don't have such a hardware reader.
Copy the PROM memory range into a file, from a running operating system. You need a running OS, and it must have access to the PROM memory range. NetBSD, for example, doesn't allow that from userland.
Hook up a serial console and dump using the PROM's own dump command.

The easiest way is to hook up a serial console. The terminal must be able to capture output to a file.

These are approximately the commands that I used:

        >>cnfg                             Show machine configuration

        >>printenv                         Show environment variables

        >>setenv more 0                    This turns off the More messages

        >>e -x 0xbfc00000:0xbfffffff       Dump the PROM data

Remember that DECstations are little endian, so if the dump data looks like this:

        bfc00000:  0x0bf0007e

then the bytes in memory are actually 0x7e, 0x00, 0xf0, and 0x0b.

At 9600 bps, about 10KB can be dumped per minute, so it takes a while. Once enough of the PROM has been dumped, you can press CTRL-C to break out. Then, restore the more environment variable:

        >>setenv more 24

Now, convert the data you just saved (little-endian words -> bytes), and store in a file. Let's call this file DECstation5000_125_promdump.bin.

        $ decprom_dump_txt_to_bin DECstation5000_125_promdump.txt DECstation5000_125_promdump.bin

This binary image can now be used in the emulator:

	$ gxemul -e 3min -Q -M128 -q 0xbfc00000:DECstation5000_125_promdump.bin

	KN02-BA V5.7e   
	?TFL:  3/scc/access (1:Ln1 reg-12: actual=0x00 xpctd=0x01) [KN02-BA]
	?TFL:  3/scc/io (1:Ln0 tx bfr not empty. status=0X 0) [KN02-BA]
	...
	--More--?TFL: 3/scsi/cntl (CUX, cause= 1000002C)
	>>?
	 ? [cmd]
	 boot [[-z #] [-n] #/path [ARG...]]
	 cat SCRPT
	 cnfg [#]
	 d [-bhw] [-S #] RNG VAL
	 e [-bhwcdoux] [-S #] RNG
	 erl [-c]
	 go [ADR]
	 init [#] [-m] [ARG...]
	 ls [#]
	 passwd [-c] [-s]
	 printenv [EVN]
	 restart
	 script SCRPT
	 setenv EVN STR
	 sh [-belvS] [SCRPT] [ARG..]
	 t [-l] #/STR [ARG..]
	 unsetenv EVN
	>>cnfg
	 3: KN02-BA  DEC      V5.7e    TCF0  (128 MB)
	                                     (enet: 00-00-00-00-00-00)
	                                     (SCSI = 7)
	 0: PMAG-BA  DEC      V5.3a    TCF0
	>>printenv
	 boot=
	 testaction=q
	 haltaction=h
	 more=24
	 #=3
	 console=*
	 osconsole=3
	>>

(Note: at the moment, this doesn't work. I must have broken something when fixing something else, but this is what it looked like at the time.)

During bootup, the PROM complains a lot about hardware failures. That's because the emulator doesn't emulate the hardware well enough yet.

The command line options used are: -e 3min for "DECstation 3min" (5000/1xx), -Q to supress the emulator's own PROM call emulation, -M128 for 128MB RAM (because GXemul doesn't correctly emulate memory detection well enough for the PROM to accept, so it will always believe there is 128MB ram anyway), and -q to supress debug messages. The 0xbfc00000 in front of the filename tells GXemul that it is a raw binary file which should be loaded at a specific virtual address.

Dumping the PROM on a SGI O2:

The general ideas in this section applies to using ROM images from other machines as well. I have also tried this on an SGI IP32 ("O2"), in addition to the DECstation.

For the O2, a suitable command to dump the prom memory range is

	>> dump -b 0xBFC00000:0xBFC80000

Make sure you capture all the output (via serial console) into a file, and then run experiments/sgiprom_to_bin on the captured file.

The photo on the left is from the real machine. The other two are screenshots of the PROM running experimentally in GXemul, using -Y2 framebuffer scaledown.

Note 2007-10-18: Running the O2 PROM image broke somewhere between release 0.3.8 and 0.4.0. (Running with graphics broke between 0.3.7 and 0.3.8.)

Normally during bootup, the IP32 PROM does a Power-On test which makes sure that the caches and other things are working properly. GXemul doesn't emulate all those things well enough for the tests to pass. The experimental screenshots above were taken with cache detection skipped manually.

In other words: don't expect this to work out-of-the-box with GXemul right now. It might work once I've added correct cache emulation.

The command line used to start the emulator, once correct cache emulation has been implemented, would be something like gxemul -XQeo2 0xbfc00000:prom.bin.

The same caution applies when dealing with SGI PROMs as with DECstation PROMs: GXemul doesn't really emulate the hardware, it only "fakes" devices well enough to fool some things, primarily NetBSD, that it is emulating a real machine. ROM code is usually a lot more picky about the details.

The graphics used in the O2 is (as far as I know) undocumented. Combining some traces of info from how Linux/O2 draws to the screen with some reverse-engineering of my own, I've implemented enough of the controller to let the PROM draw rectangles and bitmaps, but not much more. The SCSI controller is not implemented yet either.