Sunday, 10 May 2009

More muck'n about

In the quest to find something to keep me busy I've switched tacks again. I've been playing about with writing an operating system, of all things. I guess it will only be a diversion as it would take years of work to get anywhere, but it is nice to play with some low level stuff for once.

I think the idea came after I found some link to Minix 3 and had a look at that. It looked quite interesting, and I always had a soft-spot for micro-kernels. It's surprisingly functional and very small - boots in an instant. Although one problem is you really need to buy the book to understand the code as the documentation is lacking, and even then the code has moved on quite a bit since the book was written, so it isn't terribly useful anyway. The way the various servers inter-communicate isn't very simple either; for a microkernel it seems to have a tighter coupling between servers than I'd have imagined, although there is probably a reason for that. I think the original goal of nice clean readable code has also been corrupted by the bolting on of new features like virtual memory and a vfs layer. And then there are some serious performance issues with it's synchronous message passing design; well not so much the message passing bit (although the overhead is quite large) as the way the services have been implemented. I'm sure it makes the code simpler but it just doesn't scale. And no threads. I do like my threads.

But anyway, the idea of a tiny micro-kernel intrigued me, so I read up on Minix 3, HURD, Mach, L4, ECos. Grand ideas flashed before my eyes. Then all the work involved pulled me back to earth. It took me days just to get a context switch working because of the shitty PC hardware (more on that below). So I definitely want to keep things simple and have pretty low expectations of where things might end up, if not just another long-forgotten directory on one of my PC's (which is pretty much a dead-set certainty no matter what else happens).

The PC isn't nice to play with at all. How disappointing that such an utterly bullshit architecture ever got so common and cheap. First there's the processor, which starts with an ancient and obtuse instruction set and then just goes down-hill from there. Segments and TSS? Enough said. Even the manuals aren't very good; after having read other processor and hardware documentation, they seem to make things far more difficult to find than necessary and leave out too many important details. Then there's the rubbish hardware beyond the rubbish processor. Only one real timing source? And even that isn't very good. Jesus, even the bloody C=64 had far more advanced peripheral hardware than that. Even one of it's timer chips had 2 timers (plus a TOD/clock - which also had an interrupting alarm function), each of which could create interrupts, and they could be cascaded for very long intervals, they even had safe easy programming interfaces - and it had two of these chips! I don't even have to mention the SID chip vs's the PC sound - oh that's right, lets just use another timer channel to make offensive beeps instead.

Sigh. So I should really be targeting the PS3 I guess; maybe if I get the system simulator up I'll look into it. The docs on the hypervisor are a bit slim though.

Well, despite all the rubbish you have to deal with I'm making slow progress; thanks to Grub and some of the free OS's out there and making lots of mistakes in Qemu and Bochs. I have a pretty simple context switch and system call mechanism sorted out, and basic interrupt handlers for timer and keyboard. I got bogged down just writing a very simple memory allocator, which was quite disheartening (I blame lack of sleep; I made it in the end). And instead of Minix or HURD I'm looking to AmigaOS for inspiration instead. It's such a simple and functional design; although I'd like to play with memory protection and similar ideas which don't translate directly and will make things a bit trickier and slower (than they might be otherwise). So that probably means the next thing to look at is signals - which are quite trivial (the only mechanism in AmigaOS used to wait on events or signal them), and then message ports - which aren't much more complex (although separate address spaces makes the whole notion quite different). Once they're done I can probably look at implementing simple devices like keyboard and timing. Although I might also have to look at how I might handle process isolation and virtual memory at some point fairly early in the piece too.

Well, unless I get bored with it all before then.

No comments: