Our great sponsors
-
-
-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
> ITS is also available as open source (https://github.com/PDP-10/its)
Sometimes I wonder, if you wanted to create an operating system like ITS for more modern platforms, what it would look like?
Random thought: ITS is mostly written in PDP-10 assembly, what if someone wrote a compiler which read in PDP-10 assembly language and spit out C code? Could that be a first step to porting it?
It is probably a lot more complicated than that (I would not be surprised if it contained some self-modifying code, for instance). Obviously a lot of hardware-specific code would have to be just rewritten. And 6 character filenames without nested directories might have been acceptable in the 1970s, but few could endure it today.
ITS does have some interesting features contemporary systems don't:
Running process can submit commands to be run by the shell that spawned it – actually MS-DOS COMMAND.COM also had that feature (INT 2E), but I haven't seen anything else with it. A Unix shell could implement this by providing a Unix domain socket, and pass the path to that to subprocesses in an environment variable–but I've never seen that done.
Another was that a program being debugged could call an API to run commands in its own debugger – I've never seen that in any other debugger, although I suppose you could write a GDB plugin to implement it (have a magic do-nothing function, then a GDB Python script sets a breakpoint on that function, and interprets its argument as commands for GDB.) Actually, in ITS these two features were the exact same feature, since the debugger was used as the command shell.
Another was that a program had named subprocesses (by default the subprocess name was the same as the executable name, but it didn't have to be.) Compare that to most Unix shells, where it is easy to forget what you are running as background jobs 1, 2 or 3.
> For people who don't know DOS: DOS didn't really have multiple processes.
Well, compared to CP/M, it did. CP/M only supported a single program being loaded into memory at a time [0]–because every program was loaded into memory at the same absolute memory address (0x100). By contrast, DOS could have multiple processes in memory simultaneously, even though only one of them at a time could be the active foreground process. DOS processes can launch child processes; the parent will be suspended (while remaining in memory) while the child process executes, but will then resume execution once the child process finishes. Under CP/M, that was impossible, there was no API to do that, the architecture didn't support the existence of one.
Indeed, while the main execution of the parent process would be suspended while the child process ran, any interrupt handlers installed by the parent would still be invoked, so some of the parent's code could still run while the child was executing. A parent process can even expose an API to its children – which is exactly what COMMAND.COM does (such as INT 2E which I mentioned)
That's what makes TSRs possible. A TSR is essentially just an ordinary program, the only difference is that when it exits, it tells the OS to leave the process in memory rather than unloading it. By contrast, the CP/M equivalent to TSRs, RSXs (Resident System Extensions), are completely different from ordinary CP/M programs. They are loaded at a variable address near the top of memory, instead of a fixed address at the bottom.
> COMMAND.COM made part of it resident (TSR), which is how it "survived" invoking other applications.
COMMAND.COM isn't a TSR. COMMAND.COM's "residence" is essentially the same as any other program which spawns a child program under DOS, the parent always remains in memory while the child runs. Unlike a TSR, when a COMMAND.COM instance terminates (the initial COMMAND.COM instance won't, but inferior instances will), it doesn't stay in memory, it is unloaded.
What makes COMMAND.COM somewhat unusual, is that it splits itself into two portions, a "resident" portion and a "transient" portion. The resident porition is loaded at the bottom of memory, the transient portion is loaded at the top of free memory, without being officially allocated. Since DOS normally allocates memory in a bottom-up manner, if the child program doesn't use much memory, the transient portion will not be overwritten. When the chlid program returns, COMMAND.COM tests the integrity of the transient portion – if it finds it has been overwritten (because the child program needed that much memory), it reloads it from disk before continuing. That's quite unusual behaviour, but still rather different from how TSRs behave.
> At any given point, there was exactly one process running, and all the state you'd typically associate with a process was "system-wide".
That's not true. When you spawn a child process, you can either let it inherit your environment variables, or you can provide it with a new environment segment – thus each process in the system can potentially have different environment variables. Similarly, file handles are per-process. Similar to Unix – indeed, the DOS 2.x file handle design was copied from Unix – every handle has an "inherit" flag, which determines whether child processses inherit that handle from their parent. DOS keeps track of which processes have which files open, so when the last process with that file open terminates, the file is closed. Likewise, allocated memory blocks are per-process, so when a process is unloaded, DOS frees its memory blocks, but not those of ancestor processes or TSRs.
If DOS has per-process environment variables, file handles and memory allocations – why not current directories as well? I don't know, but I can speculate: DOS 2.0 copied the idea of a current directory from Unix, so like Unix they probably would have initially planned to make it per-process. However, requirements for backward compatiblity with DOS 1.x programs pushed them towards having a separate current directory per each drive. Keeping a separate current directory per-process might have been feasible if it was just a single current directory, but having to do so separately for each drive would have made it a lot more complex and wasteful of memory. I think that is why they made it system-wide instead of per-process.
Microsoft always planned to make MS-DOS a multi-tasking operating system, and much of its internal architecture was designed to support gradual evolution towards multi-tasking [1] – which makes its architecture somewhat closer to a multi-tasking operating system than that of a true single-tasking OS such as CP/M. In fact, they even developed a multi-tasking version of MS-DOS (the so-called "European MS-DOS 4.0" [2]) but it was largely unsuccessful. Eventually Microsoft gave up on the idea of making MS-DOS multitasking, but much of the ideas and experience they developed in trying to do so ended up going into OS/2 and Windows instead.
[0] CP/M later evolved into MP/M, which was a true multi-tasking operating system, but I'm not talking about that here
[1] See for example this design document in the MS-DOS 2.11 source code: https://github.com/microsoft/MS-DOS/blob/master/v2.0/bin/DEV...
[2] https://www.pcjs.org/software/pcx86/sys/dos/microsoft/4.0M/