In the beginning was CAOS

by

Andy Finkel.

Copyright © 1988 by Andy Finkel,

Commodore-Amiga, Inc.

The very first issue of Amiga Transactor - spun off from the original Transactor - featured an article by Andy Finkel, then the Software Manager at Commodore-Amiga. Fortunately, the article survived, although Amiga Transactor didn't. It's interesting to contrast the state of AmigaDOS then and now, especially in comparison to the CAOS specs. Thanks to Andy - and an intrepid Amiga owner who typed in the text and sent it to Andy - ViewPort is able to reprint that article in its entirety. So set the WayBack machine for 1988, and read on...

As most of you know, AmigaDOS was not the first choice for the top-level OS/DOS on the Amiga computer. What we now call AmigaDOS was really the backup DOS, based on an already existing OS known as Tripos (developed at the University of Cambridge Computer Laboratory by the TRIPOS Research Group, and converted with amazing speed by Metacomco's Dr. Tim King and his band of programmers). When the original, intended, designed-for-the-Amiga DOS failed to materialize (in what would no doubt be an interesting story in itself), the Amiga was launched with AmigaDOS, and the rest is history, so to speak.

Once in a while, someone suggests that the original DOS be written according to the original specs. This was even proposed by some members of the Amiga team, but Amiga management decided that it wouldn't be possible to complete the DOS and still launch the Amiga on time, especially since the software guys had already given up weekends at home. And going home. And sleeping.

We're left with a big question .. what would the Amiga have been like if we could have used the original DOS? What would the character of the Amiga have been? Would life have then been perfect?

There's really no way of knowing, of course. But if I stopped here, this would be a really short article, right? So why not take a quick look at the specs, and maybe do a little comparison with AmigaDOS...

Because these questions greatly interest developers for some reason, a small mention of CAOS on USENET brought a flurry of electronic mail, all asking for more detail (all right, they wanted the specifications). So I decided to write this - it seemed like more fun than the article I was working on ("A Guide to DOSBase", or "Why is this library different from other libraries?") so I'll save that for another time. Ideally this article should have been written by one of the Amiga people who spec'd CAOS (primarily Carl Sassenrath who also wrote Exec) way back in the dim reaches of time (1984). However, as they've been quiet about it all this time, and I have played with the prototype (do you know what a pain a Sage can be to set up?) it's time. Anyway, the specs are from the CAOS draft document.

CAOS stands for the Commodore Amiga Operating System. Its purpose was (good place to quote from the draft specifications here):

"It is a small, full-bodied operating system, functioning on a small (but great) machine bound for the consumer market-place. Its primary purpose is not to provide a sophisticated software development environment - rather is provides a foundation for pleasant and useful application programs. This is not to say that it neglects the programmer's needs. Externally it is meant for end users; internally, however, it is designed to support the complex demands of most applications. It provides extra rich functionality for an operating system of its class.

The goals of CAOS are:

CAOS was always meant to be built on top of Exec (in much the same way that AmigaDOS is built on top of Exec). Exec would provide the low level services. It would initialize the hardware, bring the system software up, handle interrupts and exceptions, do the low-level device control (like the trackdisk.device), provide system error-handling, supply the low-level timing routines and semaphores, provide some useful utility functions, and handle the basic multitasking.

This sounds familiar, right? Yes, it's true. AmigaDOS depends on Exec for the same functions.

What would CAOS itself have given? How about multiprocessing, a file system, memory management, and commands? Basically, the same things that AmigaDOS gives us - not surprisingly, as AmigaDOS had to step into the shoes of CAOS. Both CAOS and AmigaDOS have their own special strengths, though. For instance, CAOS also did memory management to a greater extent than AmigaDOS, and would have had resource tracking.

There are three (well, at least three) general interesting areas to look at about CAOS: Multiprocessing, the File System and Memory Management. And then there's also the projected commands (for fooling around with files and similar operations.)

We might as well start with the easy ones ... multiprocessing and memory management. You may wonder why multiprocessing is one of the simple ones. It's because multiprocessing (at least in general outline) is much the same as multiprocessing under AmigaDOS.

CAOS processes are built around Exec tasks. CAOS depends heavily on the Exec data structures and control primitives for multiprocessing (sound familiar?) A CAOS Process structure is built on top of an Exec Task structure. Exec handles the context switching, inter-task communication, synchronization, queuing, and mutual exclusion. As with AmigaDOS, only Processes can talk to the file system.

The CAOS Process structure contains (along with the Task structure) information about its stack, the program data, resource tracking, and exception code. (Once again, except for the `resource tracking' area, it sounds much like the AmigaDOS process information, right?)

The resource tracking is a key difference. CAOS was intended to keep a linked list containing blocks of resources used by the process - file control blocks, I/O blocks, message ports, libraries, memory usage, shared data, overlays and so on. (This part of CAOS lagged behind the rest; which may explain why we don't have even a semblance of it now.)

Processes run in user mode, using a single (user mode style) stack. Except for context switches, which run in supervisor mode (using the supervisor stack) and of course interrupts, everything was going to stick to user mode. OS system routines are well behaved (they don't mess with the stack) and use well defined amounts of stack space (they don't allocate large areas for local variables or recurse). Now, that's a big difference. AmigaDOS makes extensive use of the stack for its operations. Oh well.

There are three ways to create a process: create it from scratch, allocating new data structures and code; clone it from the parent process, making new data structures, but sharing the parent's code; or chain it from the parent, and replace it.

The general description of process creation is:

1) Allocate the process structure
2) Allocate the stack
3) Open standard in, out, and error
4) Load the code and make it ready to execute (relocate it)
5) Load static data
6) Translate the package file (if one exists), and create a package control block
7) Pre-load the package file
8) Initialize exception vectors
9) Add the task to Exec's task list (with AddTask), which puts it on the queue to run

I mentioned the package files in the above list. What exactly are they? Simply this: a package file is a method used to describe the load and run time memory environment for an application by giving the user a way of binding application program and data files in an easy to manage form. This is something we lost when we went to AmigaDOS, and it's something I think we should look at putting in.

Another important task (sorry) that CAOS was intended to perform was memory management. Looking at the current OS one may note that Exec has a rich set of primitives for dealing with memory. CAOS added some additional functionality that AmigaDOS doesn't have (and some that it does).

CAOS basically provided well-managed regions of memory. Within this region of memory, the CAOS memory manager would reign supreme (somewhat like grabbing a portion of memory with the Exec AllocEntry call). CAOS would then manage the memory within that sub-region, allocating sections for code, data, program stack, and so on. Obviously, fragmentation could set in. Since scatter-loading was not (at the time of the spec I'm using) part of the CAOS design, the CAOS memory manager needed a way to cope with fragmentation. If a sub-region became fragmented, CAOS would attempt to perform a compaction and garbage collection of regions. First, it would delete sub-regions as volunteer overlay candidates. If this failed to regain enough memory, sub-regions with no users (libraries and devices with accessor counts of zero, for instance) would be deleted. If this failed, a special entry point would be called for relocatable and unloadable subregions. If this failed, the user would be treated to a Guru (or an out of memory requester)

When a program needed to manage memory within the CAOS-allocated subregions, the program would think of memory in terms of segments. Segments could be marked as relocatable or fixed, permanent or reusable.

The CAOS memory manager would attempt to compact relocatable segments and delete reusable segments when memory got tight. Non-relocatable segments would be kept together at either the top or bottom of a region, to avoid catastrophic failure of the relocation/compaction algorithm of the memory manager.

Segments also could undergo a crude form of swapping: they could be read in from or written out to a preallocated disk file.

Believe it or not, the AmigaDOS SegList gives almost as much functionality. Except for the relocation type of segment and the associated memory manager, you can pretty much play these same games with SegLists.

Now we get to the hard stuff - the file system. This is the area that should differ most from AmigaDOS. Providing a file system is the most important goal of a disk operating system.

CAOS used a file concept just like AmigaDOS - a file being an abstract data object, with a set of access methods, that appears to the user to be a consistent data stream regardless of its underlying storage form. Many, many computer systems have happened on this concept.

CAOS supported four types of files: ordinary files, directory files, image files, and special files.

An ordinary file is used for data storage. The data can be anything - ASCII, binary, whatever. There is no system-imposed structure on the contents of an ordinary file.

Ordinary files are typed as well, to aid CAOS and the user in understanding the files' content. An ordinary file might be typed as containing special graphic information, text, program code or program data, for instance. In the case of an ordinary file of text type, the only limit imposed is that no text line shall be longer than 512 bytes.

A directory is a file used to map file names to actual data. They are treated differently from ordinary files in that there is a system-imposed structure for their contents, and they are protected against overwriting. A directory may contain the name of another directory file, and so on - a tree file system is supported.

Image files provide a symbolic method of binding to system internal data objects. CAOS would use these files to give file system-style access to internal system objects like libraries, devices, interrupt chains, etc. This would have allowed a standard high-level method of dealing with these low-level objects. The file system would have acted as a symbolic address space.

Image files were considered an `option' - special pseudo-device files might have been used instead. I really like the concept, though. It's one of the clever things I wish we had with the current DOS. (Fortunately, I think I've got a way to retro-fit this onto AmigaDOS.)

Special files are treated just like ordinary files, but don't generally represent an interface to a storage device. They are used to provide standard access to I/O devices.

CAOS allows file names of up to 30 characters. A path can be specified using the `/' character. Both absolute (from the volume root) and relative (to the current directory) paths are supported. The reserved file name `:' refers to the current directory. The reserved file name `..' refers to the parent of the specified or current directory. A path (including file name) couldn't be longer than 255 characters.

CAOS supports both hard and soft links. A hard link is just a pointer to the internal data structure of a file (actually, a directory under CAOS would be an array of hard links). A soft link references a file by using the file name and doing text substitution. (For example, if the file system was asked by a user program to Open a file called "library", which was actually a soft link to "/exec/lib/library", the file system should substitute the name "/exec/lib/library" before doing the Open.)

CAOS maintained useful file information, like a description field (up to 255 characters of information for reference purposes), creation time, update time, link count, permissions (read, write, execute, locked), type (directory, printable, non-printable, and so on), user ID, size, and blocks allocated to the file. Image and special files might not maintain all of the information; defaults - you know, lies - are returned when an application program interrogates the file system about one of these files.

Files may be opened for read, write, or append operations. Block and character operation may be supported. You can also seek to any particular byte in an ordinary file. Image and special files disallow this, though.

System directories, like / (the root), /dev, and /Exec are special directories where the system generally keeps its special files, image files and so on.

Let's go right to the proposed limits on the CAOS file system (always a popular means of comparison). Maximum file size is one megabyte; maximum name size is 30 characters. Maximum path size (including name) is 255 characters (but may have been increased to 512 with enough demand). Maximum number of files would have been one quarter of the number of blocks (about 440 on a floppy); maximum number of open files would have been 16 per process, and the maximum length of softlink text substitution would have been the same length as a path.

There are three types of file operations: high level (safe, simple) file access, low-level file operations, which work directly on the file control block (the higher level functions devolve into calls on these low level routines), and support/maintenance operations (i.e. utilities). Naturally, all file functions are called through the CAOS library interface.

Descriptor-based file operations are the easiest means to access the file system. You Open a file using a name (and path), and in return get a descriptor, which you use for all subsequent file operations.

The functions are Open, Close, Read, Write, Seek, EOF (test for EOF), Size (return the size of the file in bytes), Truncate (I miss this one), Update (write out all dirty buffers to the device associated with the file), Name (return a pointer to the file name), Options, Lock (prevent simultaneous access), Unlock (unlock a previously Locked file), LastError, and Control (miscellaneous file access control).

The high-level file access methods described above are based on a lower level file control block access method. Functions using the file control block are: MapFile (go from a descriptor to a control block), OpenFile (given a name, return a file control block), CloseFile, ReadFile, WriteFile, and Update.

By the way, it was considered possible that the low-level access to files would be removed if all operations were possible using high level calls.

Essentially the programs would be quite similar to the ones we have now (except, of course, Unix-like syntax would have been used). We'd have Link, SoftLink, Remove, Rename, FileInfo, SearchDir, CurrentDir, ChangeDir, MakeDir, and so on.

As with AmigaDOS, CAOS supports volumes to logically map disks, partitions, devices, file networks, and structured memory. Volumes can be accessed when they are online. Disk and network volumes go online once their volume ID structure has been read and verified. (For floppy disks, this happens automatically when the disk is inserted.) When access to a volume is ended, the volume becomes either absent, offline, or critical (as in, You Must Replace Volume...)

CAOS maintains volume identification information, like name (up to 40 characters), description, time stamp and write protection status. By using the "special" type of file, it is possible to get direct access to a volume. In the case of a floppy, the special file would give the same type of access (but in a file-system standard manner) as if the trackdisk.device were used directly.

This then is a quick overview of CAOS. Let's say it was available now ... What would we have gained? Personally, I think the main thing would have been a much more integrated system. From these specs you can see that AmigaDOS is really as powerful as CAOS was intended to be. The problem is that AmigaDOS is different from the rest of the Operating System. CAOS would have used the same type of data structures as the OS, the same type of stack, the same languages (C and assembler), and would have made understanding the system easier.


This article was found in the online magazine ViewPort, Oct 1993 issue.

Carl Sassenrath made the following comment about CAOS at the AmiWest IRC conference:

CAOS was contracted out, for the most part, to a company that felt Unix was a better choice and didn't buy into my design. They became history when they started using their Sun development systems for other projects, not the Amiga higher level OS functions.