ArcanOS PRE-MAGUS: frame allocator rudiments

It’s been entirely too long coming, but as of ArcanOS commit e1a75edd4b0a, the rudiments of the frame allocator are available.  Those who’ve followed the project will note that the memory manager was disabled and still hasn’t come back, but this will be coming along very soon.  Really, “memory manager” was a misnomer anyway, as the real goal was to have a basic malloc( )function.  Being able to summon up a pointer to an arbitrary chunk of memory could be considered “handy” for early development.  So, the original allocator is what I’ve been calling a “scratch pad allocator.”  Basically, it just assumes there’s a spare megabyte of RAM located just beyond the kernel’s load location and it uses this area for serving memory requests.  This actually was a pretty neat thing to have on hand in the early days of ArcanOS, before we were using paging and when it was more important to me to have a robust kernel debug log than it was to do anything else.  Of course, this had to be retired when ArcanOS switched over to paging, because the 1MB “scratch pad” shouldn’t be assumed to be mapped in.  No…the malloc( )function should necessarily rely on the paging system.

There are some distinct phases to go through in making a mechanism for allocating out the available RAM:

  1. Identify the physical addresses of all the RAM in system.
  2. Create a tracking system to know what RAM is free and which is used.
  3. Create a mechanism for mapping unused RAM into the address space of the process requesting it.

Step (1) was nicely done for us already by GRUB and the map is passed in as part of the multiboot info struct when GRUB dispatches to the kernel’s entry point.  To know where all the free RAM is in the system, you need only parse the struct.  Of course, to make use of the information in the struct, it’s handy to also tackle step (2), so that you have somewhere to record this information about available memory.  Since I think it’s probably the most minimal data structure possible (and, therefore, the easiest to read, learn, and explain…and thus in line with a goal of ArcanOS), I have opted for the ever-humble bitmap.  ArcanOS has an intentional design decision to target a very generic 32-bit x86 architecture and thus limits its address space to 4GB.  This is broken down into page-size chunks and each one gets a bit in the bitmap.  A 1 in the bitmap denotes free RAM, otherwise it should not be touched.

I set up the bitmap in memmgr.c in multiple passes.  In the first pass, I set the bitmap so that all the page frames are invalidated:

void frame_allocator_init(multiboot_info_t* mbi, uint32_t kernel_base, uint32_t kernel_end) {
	
	int i;
	int frames_allocated = 0;
	for(i=0; i<(PHYS_MEM_MAP_SIZE); i++) {
		phys_mem_map[i] = 0;
	}

Next, I iterate through the memory map given by GRUB and mark all the full and available 4KB frames as being available:

   multiboot_memory_map_t* mmap;
   for (mmap = (multiboot_memory_map_t *) (mbi->mmap_addr);
		(unsigned long) mmap < (mbi->mmap_addr) + mbi->mmap_length;
		mmap = (multiboot_memory_map_t *) ((unsigned long) mmap
								 + mmap->size + sizeof (mmap->size)))
	{
	 //_kern_print(" size = 0x%x, base_addr = 0x%x%x,"
		//	 " length = 0x%x%x, type = 0x%x\n",
			// (unsigned) mmap->size,
		//	 mmap->addr_high,
	//		 mmap->addr_low,
		//	 mmap->len_high,
	//		 mmap->len_low,
		//	 (unsigned) mmap->type);
			 
	if (mmap->type == MULTIBOOT_MEMORY_AVAILABLE) {
		 //Kindly note here that we will *NOT* be making use of the "high"
		 //fields here.  The memory manager is current capped at 4GB, so
		 //we will deal only with the lower 32-bits of addresses.
		 
		 //The goal here is to flip on the bits for available RAM.  After
		 //that, we will flip OFF the bits for the areas where the kernel
		 //was loaded and where the initial page tables sit.
		 
		 //The frame allocator is very happy to fail to mark partial
		 //frames as "not available."  This is a good, safe starting point
		 //as it means that the only physical memory allocated out will
		 //"really be there."
		 
		 uint32_t base = mmap->addr_low;
		 uint32_t len = mmap->len_low;
	
	     //First, advance up to the next page frame boundary.
		 if ((base % PAGE_SIZE) != 0) {
			 base = base + (PAGE_SIZE - (base % PAGE_SIZE));
			 len = len - (PAGE_SIZE - (base % PAGE_SIZE));
		 }
	
	     //As long as there is another frame to allocate, allocate it.
		 while (!(len<PAGE_SIZE)) {
			 mark_frame(base, FRAME_STATUS_FREE);
			 ++frames_allocated;
			 base += PAGE_SIZE;
			 len -= PAGE_SIZE;
		 }
	}
   }

And then finally I make sure that the space taken up by the kernel and the initial page directory and page tables is marked as unavailable.  This is important to remember to do…the space you’ve used before turning on your memory manager is, by definition, not available to use.  Otherwise, you’ll accidentally write over your page tables and other stuff you’ve so laboriously set up:

   //Now, time to make sure the page directory and any existing page
   //tables are never again re-allocated.
   mark_frame(INITIAL_PDE, FRAME_STATUS_USED);
   
   uint32_t* page_entry = (uint32_t*)INITIAL_PDE;
   for (i=0; i<1024; i++) {
		uint32_t pde_entry = page_entry[i];
		pde_entry &= 0xFFFFF000; //Page table address is in the upper 20 bits
		if (pde_entry != 0) {
			_kern_print("Located existing page at 0x%x and will mark it as used\n", pde_entry);
			mark_frame(pde_entry, FRAME_STATUS_USED);
		}
   }
   
   //Finally, mark off the area where the kernel is loaded
   kernel_base &= 0xFFFFF000; //Find the nearest page (rounding down)
   while(kernel_base < kernel_end) {
	   mark_frame(kernel_base, FRAME_STATUS_USED);
	   kernel_base += PAGE_SIZE;
   }
}

And, presto.  It’s a bitmap of memory.  The process of feeding free frames of memory back into the allocator is now a process of finding a 1 in the bitmap, calculating its physical address, and then updating the page directory and tables appropriately.  When I write that code, I’ll break it down on here.  I want to add that there are some obvious drawbacks to the use of a bitmap in this fashion.  The first is that the bitmap is going to be pretty sparse, meaning that space is dedicating to bits that will never actually be used.  The bitmap is also very minimal in what it can do– it’s a map of available RAM, but if a frame is marked as 0, there’s no way to tell why.  I also suspect that the process of returning frames to available RAM is not going to be as simple as it could be, and I further suspect that requesting large data extents (such as 4KB for making a new page table) may not go in a perfectly optimal fashion.  All of these, however, are hits I’m willing to take in exchange for the fact that it’s a simple, easy to code, and easy to explain data structure.  This shows how to take the information in a GRUB multiboot info struct and use it to build a map of the available page frames in memory, and being able to show this in a single blog post isn’t so bad.

So, the next step is to get the allocator allocating, which is to say the next step is to take frames out of the bitmap and map their physical addresses to available addresses in the kernel’s address space.

The Long, Hard Road out of Segmentation

As of last night’s Pre-MAGUS check-in, ArcanOS now uses paging rather than segmentation.  Of course, this is a pretty critical step in the MAGUS goals.  Segmentation is really just a legacy aspect of x86 these days, and there are really no meaningful examples of its use in modern OS development.  So, why did I ever use segmentation in the first place?  Well, the biggest reason is that I inherited it.  Remember that ArcanOS began life as an early coursework lab for a teaching operating system called JOS, and very specifically an early introductory form of it that focused on the boot loader.  ArcanOS, like JOS, and like a lot of other operating systems, is designed to be a “higher half” kernel, which means that it is linked to 0xC00000, but it’s actually loaded at a much lower address (right at 1MB, for those playing along at home).  If you really want to split hairs, ArcanOS is linked at 0xC0100000 (which will change soon, possibly in the next check-in).  In order to really “run” ArcanOS, the x86 processor has to be set so that address 0xC0100000 is mapped to 0x00100000.

This was originally done with something the hobby OS community generally calls “the GDT trick“.  In order to run in 32-bit mode, an x86 requires you set up a GDT, and this will set the values of the segment registers.  These registers will be added to every address to get a final result that goes out onto the memory bus.  It turns out that, in this addition process, overflowing 32 bits will just cause a wrap-around.  So, you can exploit this to get your kernel’s link addresses to refer to its actual load addresses (0xC0000000 + 0x4000000 = 0x00000000 after wrap-around).  It’s good for getting things up and running, but you can’t really stake much on it after that.  Segmentation is inflexible and you can’t do any swapping with it.  On top of that, there are all sorts of memory addresses that correspond to physical hardware, and you have to remember to adjust those addresses to compensate for your segmentation trick.  Even more so, though, it seems that the GDT trick isn’t universally supported.  ArcanOS uses Bochs as its reference platform, but it would appear that VirtualBox doesn’t support the GDT trick.  Long story short– getting onto paging is pretty important.

So, this left me with some questions about how I wanted to get onto paging.  My first thought was to employ the GDT trick to jump into my proper kernel code, then set up paging and reload the GDT with segment descriptors that used a base of 0x00000000 (effectively turning off segmentation).  The more I considered this, though, the less I liked it.  Part of the reason ArcanOS started out using the GDT trick was because it also started out with its own boot loader and you have to set up a GDT to get into 32-bit mode in the first place.  ArcanOS now uses GRUB for its boot loader, though, and GRUB already sets up 32-bit mode and installs a GDT to support a flat memory model.  Why was I going to change the GDT only end up setting it back the way it was before?  No, this was a good opportunity to separate ArcanOS from its roots by just going straight to paging.

Setting up paging, it turns out, is a little bit easier than I thought.  You need about 12KB of memory to establish some initial page tables, and if you’re loading the kernel at 1MB, there’s usually plenty of room just under the kernel to use for those purposes.  After that, it’s really just zeroing-out the memory and stuffing some values in the tables.  Info on the layout of the tables, with reference for making a higher half kernel, is in an article on the OSDev Wiki.  The sticky wicket, in my mind, was that I would really just rather not write it in assembly.  I don’t mind writing in assembly, but for anything non-trivial (especially in a hobby project where my free time is at a premium), I trust a compiler more.  But…if I wanted to do this in C, wouldn’t I have to write it in the kernel proper, thus requiring I rely on the GDT trick?  As it turns out, the answer is “no.”  What I realized is that the code would be location-independent as long as all I did was write some simple loops and not refer to any memory except the page tables which I was trying to fill.  All of the control flow would be PC-relative (based on adding offsets to the current code location), so it would never care where it was really located as long as I called into it correctly.  The result is the init_paging( ) function in kernel/memmgr.c.  All I have to do to call the function is to store its address in a register, adjust its link address to its load address, and then call the function from the register.  I call this function, let it do its thing, and then I finish up setting paging in assembly.

Originally, I paged the first megabyte to itself and paged 0x00100000 to 0xC0100000.  This makes the hardware exposed in low memory easy to access, ensures my first page tables are mapped in, etc, and it ensures the symbols in the kernel were all mapped in.  And, then I set up paging and….crash.  Why?  Well, because I’m already running my code at 0x00100000…which is in the second megabyte.  The address of the current instruction was not mapped to anything!  So, once paging was turned on, my own code became inaccessible.  In the end, I ended up deciding to map the first four megabytes to themselves.  I’ll want to clean that out eventually, but for initialization, it’s something I can live with.  So, this was enough to get me into the ArcanOS kernel proper, but then…crash.  What was it this time?  Well…GRUB put the multiboot info struct below 0xC0100000, so that wasn’t mapped in.  Again, just to keep the bases covered, I mapped the first four megabytes to the address space starting at 0xC0000000.  This does mean there are currently two ways to access the same 4MB of memory, but it gets ArcanOS up and booting, and there should be some easy ways to clean this up.

Technically, with this view of memory, ArcanOS would have run just like the GDT trick was still being used, but I still did due diligence and I went through and stopped referring to addresses as being relative to KERNBASE.  My word of advice to any OS hobbyist who’s starting out– I really encourage you to consider doing rudimentary paging right from the start.  If you don’t, you’re going to get married to an address management scheme you’re likely to throw away later, and if you’ve been using it for a while, then it’ll take a harder portage effort to get off of it.  Just page your memory in from the get go.  If you’re using GRUB, even more to the point, because GRUB already set up the GDT for you to use paging with a flat memory model.  The biggest concern I had with setting up paging was that it looked hard at a time when I just wanted to get a kernel booting and doing some interesting things.  In reality, I was able to sling together a C function which would set up the page tables, and this made things very easy indeed.  Don’t get into using segmentation.  If you do, it’ll be a long, hard road getting back out.

So, from here, what I really need to do is clean up the paging a little bit.  Once I’m jumped into the kernel, I can stop identity-paging the first 4MB.  I also want to map only enough pages to hold the kernel.  Also, I don’t see a good reason to link the kernel at 0xC01000000 any more, since doing that was mostly to make the GDT trick a little bit simpler.  I have a good map of physical memory thanks to GRUB and I can make and manipulate pages, so this is everything the memory manager needs to be pretty full-featured.

ArcanOS PRE-MAGUS Update

With most of the “fun” of buying a house and moving to a new part of the Bay behind me, I’ve finally had some leisure time to turn my attention to something I’ve really missed doing…namely, developing ArcanOS.  What have I been up to on that front?

Well, after having been off the project for several months due to the whole “becoming a homeowner” thing, I took a not-so-long look at the project and concluded that I needed to be more realistic about the boot loader situation.  When we last left our hero, he was intending to continue development on his boot loader until such time as it could produce a working memory map, at which point ArcanOS would be ported to the Multiboot Specification so that it could be booted by GRUB.  While I still think it would have been a lot of fun to do that, I ultimately had to be realistic about what I was getting myself into.  I’d be building multiple boot loader stages, making a new build process to support them, and enjoying the hair-pulling process of chain loading and making sure I hold onto the memory map in the process.  That would have been a good learning experience, but it would take a lot of time to do and I already lost a lot of momentum over the past few months.  Doing a lot of work only to throw it away and do a portage effort just didn’t look as appetizing.

So, as of a recent commit, ArcanOS now boots via GRUB.  The current commit will generate a disk image file for Bochs which contains the GRUB stage1 and stage2 boot loaders already.  As of tonight’s commit, ArcanOS now receives an accurate memory map from GRUB and this will give me the material I need to make a frame allocator and then work on enabling paging.

Part of me is tempted to take a side trip and include some bells and whistles to make the boot process easier.  I’m getting tired of typing in the block list of the kernel on every test boot.  I’m fairly sure that much of that sort of thing, though, requires a file system GRUB understands, and if I leap onto a canonical file system, I might never get out.  I should also note that my build process needs some tweaking…right now I have to regenerate the file in boot/post_pad every time the kernel size changes!

Since I’m back to proper feature development for the MAGUS goals, I’ve changed the version name to PRE-MAGUS.

Introducing the HP TouchPad

I’d like to take a moment to introduce to you the project that has been taking up a considerable amount of my time and attention lately.  Here it is:

HP TouchPad

Hi. I'm the TouchPad.

This, my dear friends and readers, is the HP TouchPad, a cornerstone of HP’s new line of WebOS products.  Since HP’s acquisition of Palm last summer, I have been working on the internals of WebOS, the operating system which has powered Palm’s current line of devices, and I’ve also been hard at work bringing new features to help make the transition from smart phones to a tablet a pleasant one for both developers and users.  Officially, I’ve been working on the windowing system (known as “cards” and “card stacks”), but the reach of my efforts has extended into the native applications and gaming.  I’ve also been a journeyman bug fixer during my transition into the Palm organization, before finally settling into feature sets I can call my own.

I’m very, very excited about the release of the TouchPad.  I have to admit that I don’t really think people have seen the potential of WebOS until now.  Phones are just not the ideal place to demonstrate the features of a fully multitasking operating system.  The limited amount of screen space and limited amount of attention span devoted to one’s phone just don’t add up to the kind of rich, long-lived interactions which benefit the WebOS model of concurrently running applications.  I do, however, think that this model can seriously up the game in the tablets market, which is currently dominated only by iOS and which will only become a truly competitive market this year as the TouchPad enters the fray with both iPad and Android products.

The TouchPad was announced today, along with the HP Pre3 and HP Veer (both of which I have also been an active contributor to), at a major demo in San Francisco.  I really think the demo went quite well, with only a small hiccup.  It looks like the hands-on demos afterward likewise went well.  I haven’t been able to find video from the formal on-stage demos yet, but here are some great videos of our official features and a hands-on:

Of course, this tablet is “my kid,” so I’m going to have opinions about how great I think it is.  Either way, though, this is a very big day for me personally.  You see, I’ve always had a dream, ever since the age of five when I first put my hands on a keyboard, that I would one day “make the computers go.”  I always dreamed not of being an applications developer but of a designer of the machines which people would use.  I used to lay awake on summer nights in my home in central Florida, laying on top of the blankets, sweating lightly in the thick air, and think about how, one day, I’d know enough about what was going on in that multitude of silicon to the point that I would make it come to life.  That dream is what sent me off to college to learn computer science, and it’s what sent me to graduate school when my undergrad years failed to teach me enough to fulfill my dream.  Since leaving school, I’ve been fortunate enough to work in the mobile industry where I’ve had the chance to work on phones and netbooks, but every project of mine has had a feeling of ill fate to it.  The products never gained the level of attention that made me feel like I’d put something out there in the world.  Yes, I can boast that someone has probably been able to dial 911 because I did my job, but I’ve never met anyone who’s used a Motorola Q9h…and I sincerely doubt anyone outside of me and my old HP team have touched the Compaq Airlife.

So, there’s always been something a little bit anticlimactic to much of my career.  Today has been something to change that narrative, though.  In those demos, you actually saw my work.  You don’t know that you saw it, but you did.  Things in those demos worked because I took care of them.  That itself is enough for me.  In this moment, I feel that I’ve arrived, and I have a great deal of excitement about the potential future of this product.  It’s taken several years of paying my dues, but I can finally point to something in the tech world and say that it’s me in there.  Deep down inside that device, there’s a whisper from the hopeful mind of a little boy in Garfield pajamas staccato-typing in BASIC and dreaming of Silicon Valley.

Just keep that in mind if you happen to be shopping for a tablet this summer.

ArcanOS FOOL released

It may have taken a little longer than I would have initially liked, but I’m pleased to announce that, as of this evening, ArcanOS has had its first milestone release — FOOL.  As I have mentioned in the ArcanOS wiki on Google Code, I think version numbers are pretty arbitrary, so I have instead opted to make milestone releases and name them after the trump cards in a tarot deck.  The goals for FOOL are pretty apt for its name.  In a tarot deck, “the fool” is often depicted as a young man blissfully stepping off a cliff.  The goals for this release were to basically lay the groundwork for getting to the “real work.”  This meant having a coherent build process and test/debugging platform and developing a bootloader, rudimentary kernel, basic debugging tools (like logging), exception handlers, and at least one interrupt handler.  Having all these basic pieces in place means that I have something stable I can use for future development.

Currently, ArcanOS is little more than some initialization code and a spinlock.  There is a basic handler for the keyboard interrupt, but that’s it.  There’s very little to try out, as there’s no user-space and thus no programs written for it.  Still, I’m pretty pleased with it.  It’s the first personal project I’ve taken on since 2001, and most personal open-source projects don’t even make it to their first milestone release.

The next release will be ArcanOS MAGUS.  I’m currently scoping out my goals for it.  I know I want to move the memory model to a flat, paged model.  This will affect multiple parts of the code base because it’ll mean that I will be able to stop adjusting for the segmentation offset whenever I need a linear or physical address.  I suspect this alone will make paging the major effort, but I want to do some other things that will keep moving things forward.  After all, having an environment for user processes is the point at which things get really fun.

As I mention in the ArcanOS notes, I owe a debt of gratitude to Dr. Lorenzo Alvisi as well as the xv6 project over at MIT.  xv6, in an earlier iteration, was known as JOS, which Dr. Alvisi used to produce a series of operating systems lab exercises for his operating system courses.  ArcanOS started out as one of those lab exercises, and though it has changed very significantly already, it still bears much of the original bootloader code.

ArcanOS FOOL Feature Complete

I am pleased to announce that the current push of ArcanOS represents a completion of features for release FOOL.  The feature list won’t look very big, but remember that this is a labor of love being done in my (not very ample) spare time, and that most homebrew OS projects never even make it this far.  The list of features slated for the FOOL release are:

  • Bootloader which reads the kernel from disk
  • Protected mode operation
  • Flat memory model achieved through segmentation only
  • Identification of the major kernel’s binary sections in memory
  • Scratch space memory allocator
  • Debugging output via a console print
  • Hardware exception handlers
  • Keyboard interrupt handler

I’m opting to call feature complete at this point because ArcanOS FOOL represents a small laundry list of features necessary to build the next set of kernel features.  That is…FOOL is a kernel that doesn’t just fall completely on its face.  Issues of booting have been worked out, the memory model is understood (if primitive), the exception handlers are in place so errors don’t cause a spontaneous reset, and the keyboard interrupt handler demonstrates that the PIC is remapped and that code for creating and registering interrupt service routines works.  I don’t want to go any further on feature development at this point because I want to enable paging and use it to enforce the flat memory model, and switching to paging will mean having to deal with memory-mapped hardware (like the console) differently.  A smaller codebase is easier to convert than a larger one.

ArcanOS is going to go through a cleanup and documentation phase now.  I want to thoroughly scrub all unnecessary code from the codebase and to fully get ArcanOS off its JOS heritage so that it’ll be a consistent project moving forward.  I’ll then carve off a release and post a build so people can play with it using Bochs, and then I’ll write Wiki pages to help document my major headaches.  Then, it’ll be time to start looking forward to release MAGUS.

ArcanOS pre-FOOL: Boot Summary and Heap Allocator

Last night, I did a push to the code repository for ArcanOS which represents the last of the work needed for its first heap allocator.  I inherited the majority of my bootloader code (please see the project website for my acknowledgements), so I decided that my first milestone for ArcanOS wouldn’t be “it boots and can run C code.”  That’s an amazing luxury, I might add.  The x86 boot process is rather arcane, as is learning how to read from a hard drive.  In short, the bootloader has to do a few things.  First, it has to set up a stack.  Then, it has to enable the A20 line so you can access some real memory.  Then, it has to set up some sort of segmentation as a prelude to switching to protected mode (so that 32-bit code can be used, among other things).  After all of that, being in 32-bit mode and armed with a stack, it can run the rest of itself in fairly standard gcc C, and you can thus more easily do the final task– loading and dispatching a bigger binary to do more meaningful work.

ArcanOS currently has no file system, so the kernel is just stuffed on the sector after the boot sector.  The kernel is an ELF binary, and the headers for ELF give the necessary information about where to load the kernel and how to jump to it.  The entry point to the kernel is actually in assembly because the first thing the kernel does is re-do memory segmentation so that the kernel loads into the higher memory addresses.  The boot loader actually loads the kernel at roughly 1MB, but the kernel is linked to be loaded at address 0xf0000000 + 1MB, so the first thing that’s done is that a new descriptor table is loaded to map all the low memory to high addresses.  This actually moves the video memory buffers up into the higher addresses as well, which is why the ArcanOS “console driver” must offset its writes by KERNBASE.

I’m going to skip over my experiences writing the “console driver” for now.  Suffice it to say, when you are implementing your own kernel, you quickly realize how little you have to work with.  I was once told that, ultimately, there are only two debugging tools — cleverness and printf( ).  I actually repeat that to rookies on my team, because, generally speaking, it’s true.  Well, guess what?  On day 1 of your kernel, you don’t have printf( ), and you have to go make one for yourself.  That starts with a functional screen console of some kind, and some bits of library code, such as itoa( ), that you’ve likely taken for granted your whole life.  So, a fairly considerable amount of time went into just getting some basic screen printing for debugging purposes.  I have no doubt that the “console driver” is, ultimately, throw-away code in favor of a far more robust console system, but necessity dictates.

The other thing I wanted to go ahead and get out of the way was a kernel heap allocator.  If there’s another function we all take for granted, it’s malloc( ).  There is no malloc( ) in your kernel if you don’t write it yourself.  This itself was an interesting journey, because the first challenge is finding a place to start the heap.  Placing the heap wrong might mean overwriting the kernel or running into memory dedicated to hardware I/O.  Fortunately, there isn’t much hardware to dodge once you’re above the 1MB offset, so the question is how to find where the kernel ends.  The way I’ve solved this is by extending the linker script called “kernel.ld”.  The PROVIDE( ) function in linker script language is particularly useful, and it lets you store information gathered in the linking process into C variables which you declare “extern”.  You can see the code which “receives” this information in kern/init.c.

What about the allocator itself?  Well, I decided that determining the amount of RAM on the target platform was a fish too big to fry right away.  RAM detection requires that I either give up my boot loader in favor of GRUB or that I add RAM detection into my boot loader before I go into protected mode (since it requires BIOS calls).  Both are non-trivial, and I also don’t plan on doing any massive allocations any time soon.  Given this, the heap just naively assumes it’s got 1MB of space in which to work.  I’ll correct this in future releases.

The allocation algorithm is incredibly simple.  The allocator will check to see if it’s got a free chunk that’s already the right size and will return that if it’s available.  Otherwise, the last member of the free list is always the “frontier” of never-before-allocated space, and it will carve off a new hunk from it.  The algorithm for freeing space it so simply move the chunk from the “allocated list” into the “free list” for future use.  There’s no provision for coalescing chunks together in the free list or anything like that, and I have some reasons for not doing this.  The first reason is because, frankly, I expect that I will rewrite this heap allocator when I have RAM detection and other goodies of this nature available to me.  The other major reason is because this allocation scheme is, in reality, enough to keep development moving along for now.  ArcanOS currently just services itself, and its allocation needs are small, and the 1MB space is likely to be plenty for a while.  Moreover, I expect most of my allocation needs will be a handful of objects of regular size.  If I were allocating lots of things of different sizes, this would eat into memory pretty badly, but if I’m just allocating structs of the same type over and over again, this system works pretty well.  Finally, a more complex algorithm means more chances for bugs…at a time when my available debugging options aren’t huge.  It’s unwise to invite extra bugs for functionality you don’t need when you expect you’re going to rewrite it anyway.

A crummy malloc( ) is, by the way, WAY better than no malloc( ) at all.

And now, armed with console out and a heap allocator, it’s time for me to attack the last feature for the FOOL milestone — writing ISRs and turning on interrupts.