iris/report/making-of.tex

\documentclass{shevek}
\begin{document}
\title{Writing a kernel from scratch}
\author{Bas Wijnen}
\date{\today}
\maketitle
\begin{abstract}
This is a report of the process of writing a kernel from scratch for
the cheap (€150) Trendtac laptop.  In a following report I shall write about
the operating system on top of it.  It is written while writing the system, so
that no steps are forgotten.  Choices are explained and problems (and their
solutions) are shown.  After reading this, you should have a thorough
understanding of the kernel, and (with significant effort) be able to write a
similar kernel yourself.  This document assumes a working Debian system with
root access (for installing packages), and some knowledge about computer
architectures.  (If you lack that knowledge, you can try to read it anyway and
check other sources when you see something new.)
\end{abstract}

\tableofcontents

\section{Hardware details}
The first step in the process of writing an operating system is finding out
what the system is you're going to program for.  While most of the work is
supposed to be platform--independant, some parts, especially in the beginning,
will depend very much on the actual hardware.  So I searched the net and found:
\begin{itemize}
\item There's a \textbf{Jz4730} chip inside, which implements most
functionality.  It has a mips core, an OHCI USB host controller (so no USB2),
an AC97 audio device, a TFT display controller, an SD card reader, a network
device, and lots of general purpose I/O pins, which are used for the LEDs and
the keyboard.  There are also two PWM outputs, one of which seems to be used
with the display.  It also has some other features, such as a digital camera
controller, which are not used in the design.
\item There's a separate 4-port USB hub inside.
\item There's a serial port which is accessible with a tiny connector inside
the battery compartiment.  It uses TTL signals, so to use it with a PC serial
port, the signals must be converted with a MAX232.  That is normal for these
boards, so I already have one handy.  The main problem in this case is that the
connector is an unusual one, so it may take some time until I can actually
connect things to the serial port.
\end{itemize}

First problem is how to write code which can be booted.  This seems easy: put a
file named \textbf{uimage} on the first partition on an SD card, which must be
formatted FAT or ext3, and hold down Fn, left shift and left control while
booting.  The partition must also not be larger than 32 MB.

The boot program is u-boot, which has good documentation on the web.  Also,
there is a Debian package named uboot-mkimage, which has the mkimage executable
to create images that can be booted using u-boot.  uimage should be in this
format.

To understand at least something of addresses, it's important to understand the
memory model of the mips architecture:
\begin{itemize}
\item usermode code will never reference anything in the upper half of the
memory (above 0x80000000).  If it does, it receives a segmentation fault.
\item access in the lower half is paged and can be cached.  This is called
kuseg when used from kernel code.  It will access the same pages as non-kernel
code finds there.
\item the upper half is divided in 3 segments.
\item kseg0 runs from 0x80000000 to 0xa0000000.  Access to this memory will
access physical memory from 0x00000000 to 0x20000000.  It is cached, but not
mapped (meaning it accesses physical, not virtual, memory)
\item kseg1 runs from 0xa0000000 to 0xc0000000.  It is identical to kseg0,
except that is is not cached.
\item kseg2 runs from 0xc0000000 to the top.  It is mapped like user memory,
differently for each process, and can be cached.  It is intended for
per-address space kernel structures.  I shall not use it in my kernel.
\end{itemize}
U-boot has some standard commands.  It can load the image from the SD card at
0x80600000.  Even though the Linux image seems to use a different address, I'll
go with this one for now.

\section{Cross-compiler}
Next thing to do is build a cross-compiler so it is possible to try out some
things.  This shouldn't need to be very complex, but it is.  I wrote a separate
document about how to do this.  Please read that if you don't have a working
cross-compiler, or if you would like to install libraries for cross-building
more easily.

\section{Choosing a language to write in}
Having a cross-compiler, the next thing to do is choose a language.  I prefer
to use C++ for most things.  I have used C for a previous kernel, though,
because it is more low-level.  This time, I decided to try C++.  But since I'm
not linking any libraries, I need to avoid things like new and delete.  For
performance reasons I also don't use exceptions.  They might need library
support as well.  So what I use C++ for is classes with member functions, and
default function arguments.  I'm not even using these all the time, and the
whole thing is very much like C anyway.

Except for one change I made: I'm using a \textit{pythonic preprocessor} I
wrote.  It changes python-style indented code into something a C compiler
accepts.  It shouldn't be too hard to understand if you see the kernel source.
Arguments to flow control instructions (if, while, for) do not need
parenthesis, but instead have a colon at the end of the line.  After a colon at
the end of a line follows a possibly empty indented block, which is put in
brackets.  Indenting a line with respect to the previous one without a colon
will not do anything: it makes it a continuation.  Any line which is not empty
or otherwise special gets a semicolon at the end, so you don't need to type
those.  When using both spaces and tabs (which I don't recommend), set the tab
width to 8 spaces.

\section{Making things run}
For loading a program, it must be a binary executable with a header.  The
header is inserted by mkimage.  It needs a load address and an entry point.
Initially at least, the load address is 0x80600000.  The entry point must be
computed from the executable.  The easiest way to do this is by making sure
that it is the first byte in the executable.  The file can then be linked as
binary, so without any headers.  This is done by giving the
\verb+--oformat binary+ switch to ld.  I think the image is loaded without the
header, so that can be completely ignored while building.  However, it might
include it.  In that case, the entry point should be 0x40 higher, because
that's the size of the header.

\section{The first version of the kernel}
This sounds better than it is.  The first version will be able to boot, and
somehow show that it did that.  Not too impressive at all, and certainly not
usable.  It is meant to find out if everything I wrote above actually works.

For this kernel I need several things: a program which can boot, and a way to
tell the user.  As the way to tell the user, I decided to use the caps-lock
LED.  The display is quite complex to program, I suppose, so I won't even try
at this stage.  The LED should be easy.  Especially because Linux can use it
too.  I copied the code from the Linux kernel patch that seemed to be about the
LED, and that gave me the macros \verb+__gpio_as_output+, \verb+__gpio_set_pin+
and \verb+__gpio_clear_pin+.  And of course there's \verb+CAPSLOCKLED_IO+,
which is the pin to set or clear.

I used these macros in a function I called \verb+kernel_entry+.  In an endless
loop, it switches the LED on 1000000 times, then off 1000000 times.  If the
time required to set the led is in the order of microseconds, the LED should be
blinking in the order of seconds.  I tried with 1000 first, but that left the
LED on seemingly permanently, so it was appearantly way too fast.

This is the code I want to run, but it isn't quite ready for that yet.  A C
function needs to have a stack when it is called.  It is possible that u-boot
provides one, but it may also not do that.  To be sure, it's best to use some
assembly as the real entry point, which sets up the stack and calls the
function.

The symbol that ld will use as its entry point must be called \verb+__start+
(on some other architectures with just one underscore).  So I created a simple
assembly file which defines some stack space and does the setting up.  It also
sets \$gp to the so-called \textit{global offset table}, and clears the .bss
section.  This is needed to make compiler-generated code run properly.

Now how to build the image file?  This is a problem.  The ELF format allows
paged memory, which means that simply loading the file may not put everything
at its proper address.  ld has an option for this, \verb+--omagic+.  This is
meant for the a.out format, which isn't supported by mipsel binutils, but that
doesn't matter.  The result is still that the .text section (with the
executable code) is first in the file, immediately followed by the .data
section.  So that means that loading the file into memory at the right address
results in all parts of the file in the proper place.  Adding
\verb+-Ttext 0x80600000+ makes everything right.  However, the result is still
an ELF file.  So I use objcopy with \verb+-Obinary+ to create a binary file
from it.  At this point, I also extract the start address (the location of
\verb+__start+) from the ELF file, and use that for building uimage.  That
way it is no longer needed that \_\_start is at the first byte of the file.

Booting from the SD card is as easy as it seemed, except that I first tried an
mmc card (which fits in the same slot, and usually works when SD is accepted)
and that didn't work.  So you really need an SD card.

\section{Context switching}
One very central thing in the kernel is context switching.  That is, we need to
know how the registers and the memory are organized when a user program is
running.  In order to understand that, we must know how paging is done.  I
already found that it is done by coprocessor 0, so now I need to find out how
that works.

On the net I found the \textit{MIPS32 architecture for developers}, version 3
of which is sub-titled \textit{the MIPS32 priviledged resource architecture}.
It explains everything there is to know about things which are not accessible
from normal programs.  In other words, it is exactly the right book for
programming a kernel or device driver using this processor.  How nice.

It explains that memory accesses to the lower 2GB are (almost always) mapped
through a TLB (translation lookaside buffer).  This is an array of some records
where virtual to physical address mappings are stored.  In case of a TLB-miss
(the virtual address cannot be found in the table), an exception is generated
and the kernel must insert the mapping into the TLB.

This is very flexible, because I get to decide how I write the kernel.  I shall
use something similar to the hardware implementation of the IBM PC: a page
directory which contains links to page tables, with each page table filled with
pointers to page information.  It is useful to have a direct mapping from
virtual address to kernel data as well.  There are several ways how this can be
achieved.  The two simplest ones each have their own drawback: making a shadow
page directory with shadow page tables with links to the kernel structures
instead of the pages wastes some memory.  Using only the shadow, and doing a
lookup of the physical address in the kernel structure (where it must be stored
anyway) wastes some cpu time during the lookup.  At this moment I do not know
what is more expensive.  I'll initially go for the cpu time wasting approach.

\section{Kernel entry}
Now that I have an idea of how a process looks in memory, I need to implement
kernel entry and exit.  A process is preempted or makes a request, then the
kernel responds, and then a process (possibly the same) is started again.

The main problem of kernel entry is to save all registers in the kernel
structure which is associated with the thread.  In case of the MIPS processor,
there is a simple solution: there are two registers, k0 and k1, which cannot be
used by the thread.  So they can be set before starting the thread, and will
still have their values when the kernel is entered again.  By pointing one of
them to the place to save the data, it becomes easy to perform the save and
restore.

As with the bootstrap process, this must be done in assembly.  In this case
this is because the user stack must not be used, and a C function will use the
current stack.  It will also mess up some registers before you can save them.

The next problem is how to get the interrupt code at its address.  I'll try to
load the thing at address 0x80000000.  It seems to work, which is good.  Linux
probably has some reason to do things differently, but if this works, it is the
easiest way.

\section{Memory organization}
Now I've reached the point where I need to create some memory structures.  To
do that, I first need to decide how to organize the memory.  There's one very
simple rule in my system: everyone must pay for what they use.  For memory,
this means that a process brings its own memory where the kernel can write
things about it.  The kernel does not need its own allocation system, because
it always works for some process.  If the process doesn't provide the memory,
the operation will fail.\footnote{There are some functions with \textit{alloc}
in their name.  However, they allocate pieces of memory which is owned by the
calling process.  The kernel never allocates anything for itself, except during
boot.}

Memory will be organized hierarchically.  It belongs to a container, which I
shall call \textit{Memory}.  The entire Memory is the property of another
Memory, its parent.  This is true for all but one, which is the top level
Memory.  The top level Memory owns all memory in the system.  Some of it
directly, most of it through other Memories.

The kernel will have a list of unclaimed pages.  For optimization, it actually
has two lists: one with pages containing only zeroes, one with pages containing
junk.  When idle, the junk pages can be filled with zeroes.

Because the kernel starts at address 0, building up the list of pages is very
easy: starting from the first page above the top of the kernel, everything is
free space.  Initially, all pages are added to the junk list.

\section{The idle task}
When there is nothing to do, an endless loop should be waiting for interrupts.
This loop is called the idle task.  I use it also to exit bootstrapping, by
enabling interrupts after everything is set up as if we're running the idle
task, and then jumping to it.

There are two options for the idle task, again with their own drawbacks.  The
idle task can run in kernel mode.  This is easy, it doesn't need any paging
machinery then.  However, this means that the kernel must read-modify-write the
Status register of coprocessor 0, which contains the operating mode, on every
context switch.  That's quite an expensive operation for such a critical path.

The other option is to run it in user mode.  The drawback there is that it
needs a page directory and a page table.  However, since the code is completely
trusted, it may be possible to sneak that in through some unused space between
two interrupt handlers.  That means there's no fault when accessing some memory
owned by others (which is a security issue), but the idle task is so trivial
that it can be assumed to run without affecting them.

\section{Intermezzo: some problems}
Some problems came up while working.  First, I found that the code sometimes
didn't work and sometimes it did.  It seemed that it had problems when the
functions I called became more complex.  Looking at the disassembly, it appears
that I didn't fully understand the calling convention used by the compiler.
Appearantly, it always needs to have register t9 set to the called function.
In all compiled code, functions are called as \verb+jalr $t9+.  It took quite
some time to figure this out, but setting t9 to the called function in my
assembly code does indeed solve the problem.

I also found that every compiled function starts with setting up gp.  This is
complete nonsense, since gp is not changed by any code (and it isn't restored
at the end of a function either).  I'll report this as a but to the compiler.
Because it is done for every function, it means a significant performance hit
for any program.

The other problem is that the machine was still doing unexpected things.
Appearantly, u-boot enables interrupts and handles them.  This is not very nice
when I'm busy setting up interrupt handlers.  So before doing anything else, I
first switch off all interrupts by writing 0 to the Status register of CP0.

This also reminded me that I need to flush the cache, so that I can be sure
everything is correct.  For that reason, I need to start at 0xa0000000, not
0x80000000, so that the startup code is not cached.  It should be fine to load
the kernel at 0x80000000, but jump in at the non-cached location anyway, if I
make sure the initial code, which clears the cache, can handle it.  After that,
I jump to the cached region, and everything should be fine.  However, at this
moment I first link the kernel at the non-cached address, so I don't need to
worry about it.

Finally, I read in the books that k0 and k1 are in fact normal general purpose
registers.  So while they are by convention used for kernel purposes, and
compilers will likely not touch them, the kernel can't actually rely on them
not being changed by user code.  So I'll need to use a different approach for
saving the processor state.  The solution is trivial: use k1 as before, but
first load it from a fixed memory location.  To be able to store k1 itself, a
page must be mapped in kseg3 (wired into the tlb), which can then be accessed
with a negative index to \$zero.

At this point, I was completely startled by crashes depending on seemingly
irrelevant changes.  After a lot of investigation, I saw that I had forgotten
that mips jumps have a delay slot, which is executed after the jump, before the
first new instruction is executed.  I was executing random instructions, which
lead to random behaviour.

\section{Back to the idle task}
With all this out of the way, I continued to implement the idle task.  I hoped
to be able to never write to the Status register.  However, this is not
possible.  The idle task must be in user mode, and it must call wait.  That
means it needs the coprocessor 0 usable bit set.  This bit may not be set for
normal processes, however, or they would be able to change the tlb and all
protection would be lost.  However, writing to the Status register is not a
problem.  First of all, it is only needed during a task switch, and they aren't
as frequent as context switches (every entry to the kernel is a context switch,
only when a different task is entered from the kernel than exited to the kernel
is it a task switch).  Furthermore, and more importantly, coprocessor 0 is
intgrated into the cpu, and writing to it is actually a very fast operation and
not something to be avoided at all.

So to switch to user mode, I set up the Status register so that it looks like
it's handling an exception, set EPC to the address of the idle task, and use
eret to ``return'' to it.

\section{Timer interrupts}
This worked well.  Now I expected to get a timer interrupt soon after jumping
to the idle task.  After all, I have set up the compare register, the timer
should be running and I enabled the interrupts.  However, nothing happened.  I
looked at the contents of the count register, and found that it was 0.  This
means that it is not actually counting at all.  Looking at the Linux sources,
they don't use this timer either, but instead use the cpu-external (but
integrated in the chip) timer.  The documentation says that they have a
different reason for this than a non-functional cpu timer.  Still, it means it
can be used as an alternative.

Having a timer is important for preemptive multitasking: a process needs to be
interrupted in order to be preempted, so there needs to be a periodic interrupt
source.

During testing it is not critical to have a timer interrupt.  Without it, the
system can still do cooperative multitasking, and all other aspects of the
system can be tested.  So I decided to leave the timer interrupts until I'm
going to write the drivers for the rest of the hardware as well.

\section{Invoke}
So now I need to accept calls from programs and handle them.  For this, I need
to decide what such a call looks like.  It will need to send a capability to
invoke, and a number of capabilities and numbers as arguments.  I chose to send
four capabilities (so five in total) and also four numbers.  The way to send
these is by setting registers before making a system call.  Similarly, when the
kernel returns a message, it sets the registers before returing to the program.

I wrote one file with assembly for receiving interrupts and exceptions
(including system calls) and one file with functions called from this assembly
to do most of the work.  For syscall, I call an arch-specific\footnote{I split
off all arch-specific parts into a limited number of files.  While I am
currently writing the kernel only for the Trendtac, I'm trying to make it easy
to port it to other machines later.} invoke function, which reads the message,
puts it in variables, and calls the real invoke function.

The real invoke function analyzes the called capability: if it is in page 0
(which is used by the interrupt handlers, and cannot hold real capabilities),
it must be a kernel-implemented object.  If not, it is a pointer to a Receiver.

Then kernel object calls are handled, and messages to receivers are sent.  When
all is done, control is returned to the current process, which may or may not
be the calling process.  If it isn't, the processor state is initialized for
the new process by setting the coprocessor 0 usable bit in the Status register
and the asid bits in the EntryHi register of CP0.

\section{Paging}
While implementing user programs, I needed to think about paging as well.  When
a TLB miss occurs, the processor must have a fast way to reload it.  For this,
page tables are needed.  On Intel processors, these need to be in the format
that Intel considered useful.  On a mips processor, the programmer can choose
whatever they want.  The Intel format is a page containing the
\textit{directory}, 1024 pointers to other pages.  Each of those pages contains
1024 pointers to the actual page.  That way, 10 bits of the virtual address
come from the directory, 10 bits from the page table, and 12 from the offset
within the page, leading to a total of 32 bits of virtual memory addressing.

On mips, we need 31 bits, because addresses with the upper bit set will always
result in an address error.  So using the same format would waste half of the
page directory.  However, it is often useful to have address to mapped page
information as well.  For this, a shadow page table structure would be needed.
It seems logical to use the upper half of the directory page for the shadow
directory.  However, I chose a different approach: I used the directory for
bits 21 to 30 (as opposed to 22 to 31).  Since there are still 12 bit
addressable pages, this leaves 9 bits for the page tables.  I split every page
table in two, with the data for EntryLo registers in the lower half, and a
pointer to page information in the upper half of the page.  This way, my page
tables are smaller, and I waste less space for mostly empty page tables.

To make a TLB refill as fast as possible, I implemented it directly in the
assembly handler.  First, I check if k0 and k1 are both zero.  If not, I use
the slow handler.  If they are, I can use them as temporaries, and simply set
them to zero before returning.  Then I read the current directory (which I save
during a task switch), get the proper entry from it, get the page table from
there, get the proper entry from that as well, and put that in the TLB.  Having
done that, I reset k0 and k1, and return.  No other registers are changed, so
they need not be saved either.  If anything unexpected happens (there is no
page table or no page entry at the faulting address), the slow handler is
called, which will fail as well, but it will handle the failure.  This is
slightly slower than handling the failure directly, but speed is no issue in
case of such a failure.

While implementing this, I have been searching for a problem for some time.  In
the end, I found that the value in the EntryLo registers does not have the bits
at their normal locations, but 6 bits back.  I was mapping the wrong page in,
and thus got invalid data when it was being used.

\section{Sharing}
The next big issue is sharing memory.  In order to have efficient
communication, it is important to use shared memory.  The question is how to
implement it.  A Page can be mapped to memory in the address space that owns
it.  It can be mapped to multiple locations in that address space.  However I
may remove this feature for performance reasons.  It doesn't happen much
anyway, and it is always possible to map the same frame (a page in physical
memory) to multiple virtual addresses by creating an multiple Pages.

For sharing, a frame must also be mappable in a different address space.  In
that case, an operation must be used which copies or moves the frame from one
Page to another.  There is a problem with rights, though: if there is an
operation which allows a frame to be filled into a Page, then the rights of
capabilities to that Page may not be appropriate for the frame.  For example,
if I have a frame which I am not allowed to write, and a frame which I am
allowed to write, I should not be able to write to the first frame by
transferring it to the second Page.  So some frame rights must be stored in the
Page, and they must be updated during copy and move frame operations.

Move frame is only an optimization.  It allows the receiver to request a
personal copy of the data without actually copying anything.  The result for
the sender is a Page without a frame.  Any mappings it has are remembered, but
until a new frame is requested, no frame will be mapped at the address.  A Page
is also able to \textit{forget} its frame, thereby freeing some of its memory
quota (if it stops paying for it as well; a payed-for frame costs quota, and is
guaranteed to be allocatable at any time).

Another optimization is to specify a minimum number of bytes for a page move.
If the page needs to be copied, this reduces the time needed to complete that
operation.  The rest of the page should not contain secret data: it is possible
that the entire page is copied, for example if it doesn't need to be copied,
but can be reused.

\section{Copy on write}
Another nice optimization is \textit{copy on write}: a page is shared
read-only, and when a page-fault happens, the kernel will copy the contents, so
that the other owner(s) don't see the changes.  For the moment, I don't
implemnt this.  I'm not sure if I want it in the kernel at all.  It can well be
implemented using an exception handler in user space, and is not used enough to
spend kernel space on, I think.  But I can change my mind on that later.

\section{Memory listing}
The last thing to do for now is allowing a memory to be listed.  That is,
having a suitably priviledged capability to a Memory should allow a program to
see what's in it.  In particular, what objects it holds, and where pages are
mapped.  Probably also what messages are in a receiver's queue.

\end{document}