A bootstrapping OS with minimal binary seed (mirror)

Giovanni Mascellani 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
asmg 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
asmg0 d6c9d1e1f4 Remove unused platform_exit API. 4 months ago
attic 541c6dc267 Remove useless multiboot mmap stuff. 4 months ago
boot f5ea8a92c4 Other debugging code. 4 months ago
contrib 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
diskfs 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
http 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
lib 5c704fe41d Use QEMU ACPI shutdown to avoid bad exit code. 2 months ago
test 0e3a67ff1b Use images in QCOW2 format. 2 months ago
.gitlab-ci.yml 5c704fe41d Use QEMU ACPI shutdown to avoid bad exit code. 2 months ago
.gitmodules 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
.travis.yml 409383f656 Start a test HTTP server in Travis. 3 months ago
ASSUMPTIONS.md 145a6c92c7 Write down low level assumptions. 4 months ago
COPYING 116c0a6d2c Add licensing details. 1 year ago
G_LANGUAGE.md f7aa688c77 Indirect function call. 4 months ago
MM0.md ccacf4b68f Introduce Type objects. 4 months ago
Makefile 0e3a67ff1b Use images in QCOW2 format. 2 months ago
README.md 1a769a0775 Fix some errors, update README, try to introduce SHA256 code verification. 1 month ago
create_diskfs.py f597115d0d Fix diskfs creation. 6 months ago
create_partition.py 976b604c0a Implement diskfs. 11 months ago
extract_debugfs.py 8d7b09508c Dump and recover tokens after preprocessing. 6 months ago


asmc, a bootstrapping OS with minimal binary seed

asmc is an extremely minimal operating system, whose binary seed (the compiled machine code that is fed to the CPU when the system boots) is very small (around 6 KiB). Such compiled code is enough to contain a minimal IO environment and compiler, which is used to compile a more powerful environment and compiler, further used to compile even more powerful things, until a fully running system is bootstrapped. In this way nearly any code running in your computer is compiled during the boot procedure, except the the initial seed that ideally is kept as small as possible.

This at least is the plan; from the moment we are not yet at the point where we manage to compile a real system environment. However, work is ongoing (and you can contribute!).

The name asmc indicates two of the most prominents languages used in this process: Assembly (with which the initial seed is written) and C (one of the first targets we aim to). The initial plan was to embed an Assembly compiler in the binary seed and then use Assembly to produce a C compiler. In the end a different path was devised: the initial seed is written in Assembly and embeds a G compiler (where G is a custom language, sitting something between Assembly and C, conceived to be very easy to compile); the G compiler is then use to produce a C compiler. Assembly is never directly used in this chain, although it is of course continuously behind the curtains.

asmc is currently able to:

  • boot from a minimal seed of around 6 KiB;

  • compile and execute source code written in the G language (see below);

  • initialize a minimal environment, including a terminal (writing to serial port and to monitor), dynamic memory allocation, simple data structures, disk access and a simple virtual file system;

  • compile a basic assembler and a basic C compiler, both of which are written in G;

  • use them to compile a lightly patched version of tinycc, using a custom basic C standard library;

  • use tinycc to compile another copy of itself (in line of principle that can be repeated as long as you want);

  • use the second tinycc to compile a patched and customized version of iPXE;

  • use iPXE to initialize the network card and be ready to download more source code from the network.

Hopefully asmc will eventually be able to:

  • download the Linux source code, patch it and compile via a custom build script;

  • compile a minimal userspace environment, which includes at least a copy of tinycc;

  • boot Linux with this userspace, and be ready to compile other source code to continue bootstrapping.

Enough talking, show me the deal!

You should use Linux to compile asmc, although some parts of it can also be built on macOS. If you use Debian, install the prerequisites with

sudo apt-get install build-essential nasm qemu-system-x86 python3 qemu-utils

If you cloned the GIT repository, you will probably want to checkout the submodules as well:

git submodule init
git submodule update --recursive

Then just call make in the toplevel directory of the repository. A subdirectory named build will be created, with all compilation artifacts inside it. Before running the virtual machine you have to start the web server that serves code for the later stages of asmc: open another terminal, enter the http directory and run

python3 -m http.server 8080


qemu-system-i386 -m 256M -hda build/boot_asmg.x86.qcow2 -serial stdio -display none

(if your host system supports it, you can add -enable-kvm to benefit of KVM acceleration)

WARNING! ATTENTION! Remember that you are giving full control over you hardware to an experimental program whose author is not an expert operating system programmer: it could have bugs and overwrite your disk, damage the hardware and whatever. I only run it on an old otherwise unused laptop that once belonged to my grandmother. No problem has ever become apparent, but care is never too much!

How to interpret all the writings

For the full story, read below and the code. However, just to have an idea of what is happening, if you use the command above to boot boot_asmg.x86 the following will happen:

  • The first log lines are written by the bootloader. At this point it is mostly concerned with loading to RAM the actual kernel, enabling some obscure features of the PC platform used to boot properly and entering the CPU's protected mode.

  • At some point asmc kernel is finally ran, and it writes Hello, asmc! to the log. There is where the asmc binary seed first enters execution. It will just initialize some data structures and then invoke its embedded G compiler to compile the file main.g and then call the main routine.

  • This is the point where for the first time code that has just been compiled is fed to the CPU, so in a sense the binary seed is not in control of the main program any more (but still gets called as a library, for example to compile other G sources). The message Hello, G! is written to the log and immediately after other G sources are compiled, first to introduce some library code (like malloc/free, code for handling dynamic vectors and maps, some basic disk/filesystem driver and other utilities) and then to compile an assembler and a C compiler. These two compilers are not meant to be complete: they are just enough to build the following step, which is tinycc.

  • Then a suite of C test programs is compiled and executed. They test part of the C compiler and the C standard library. In line of principle all the test should pass and all malloc-s should be free-ed.

  • After all tests have passed, tinycc is finally compiled. This takes a bit (around 20 seconds on my machine, my KVM enable), because the previous C compiler is quite inefficient. During preprocessing progress is indicated by open and closed square brackets, which indicate when a new file is included or finished to include. During compilation (which consists of three stages), progress is indicated by dots, where each dot correspons to a thousands tokens processed.

  • At last, tinycc is ran, by mean of its libtcc interface. A small test program is compiled and executed, showing a glimpse of the third level of compiled code from the beginning of the asmc run.

  • In the end some statistics are printed, hopefully showing that all allocated memory have been deallocated (not that it matters much, since the machine is going to be powered off anyway, but I like resources to be deinitialized properly).

How to fiddle with flags

asmg's behaviour can be customized thanks to some flags at the beginning of the asmg/main.g file. There you can enable:

  • RUN_ASM: it will compile an assembler and then run it on the file test/test.asm. The assembler is not complete at all, but ideally it should be more or less enough to assemble a lightly patched version of FASM (see next point).

  • RUN_FASM (currently unmaintained): compile the assembler and then use it to assemble FASM, as mentioned above. In theory it should work, but in practice it does not: the assembled FASM crashes for some reason I could not understand. There is definitely a bug in my assembler (or at least some unmet FASM assumption), but I could not find it so far. However, the bulk of the project is not here.

  • RUN_C: it will compile the assembler and the C compiler and then use them to compile the program in diskfs/tests/test.c. In the source code there are flags to dump debugging information, including a dump of the machine code itself. It is useful to debug the C compiler itself. Also, feel free to edit the test program to test your own C programs (but expect frequent breakages!).

  • RUN_MESCC (currently unmaintained; only this port is unmaintained, the original project is going on): it will compile a port of the mescc/M2-Planet toolchain, which is basically an indepdendent C compiler with different features and bugs than mine. This port just tracks the upstream program, no original development is done here. See below from more precise links. The test program in test/test_mes.c will then be compiled and executed.

  • RUN_MCPP (currently unmaintained): it will compile the assembler and the C compiler and then use them to try compiling a lightly patched version of mcpp, which is a complete C preprocessor. Since the preprocessor embedded in asmc's C compiler is rather lacking, the idea is that mcpp could be used instead to compile C sources that require deep preprocessing capabilities. However, at this point, mcpp itself does not compile, so at some point asmc with die with a failed assertion. Also, it nowadays seems that asmc is able to preprocess tinycc by itself, so there is no point anymore in going forward with this subproject.

  • RUN_TINYCC: here is where the juice stays! This will compile the assembler and the C compiler, and then compile tinycc, as mentioned above. Then it will use tinycc to compile and execute a little C program. In the future the bootstrapping chain will continue here.

  • TEST_MAP: there are three implementation of an associative array in asmc, of increasing complexity (see below). This tests the implementation, and was used in the past to check new implementations for correctness.

  • TEST_INT64: implementing 64 bits integers on a 32 bits platform is somewhat tricky. The G language itself only supports 32 bits numbers, so some additional Assembly code was required to implement 64 bits operations. Also, the division code is particularly tricky. However, 64 bits integers are required by tinycc, which needs support for long long types, so they were implemented at some point. This enables some tests on the resulting implemntation.

  • TEST_C: it will compile the C compiler and run the test suite.

By default thre three TEST_* flags and RUN_TINYCC are enabled in asmc.

There is also another pack of flags that control which malloc implementation asmg is going to use. There are four at this point. All of them gather memory with the platform_allocate call (see below), which is similary to UNIX' brk (and does not permit to release memory back).

  • USE_TRIVIAL_MALLOC: just map malloc to platform_allocate and discard free. Very quick, but wastes all free-ed memory.

  • USE_SIMPLE_MALLOC: a simple freelist implementation, ported from here, which is probably rather memory efficient, but can be linear in time, so it easily becomes a bottleneck.

  • USE_CHECKED_MALLOC: somewhat similar to USE_TRIVIAL_MALLOC, but checks that your program uses malloc and free correctly (i.e., that you not overflow or underflow your allocations, that you do not double free, or use after free). As a result it is very slow and memory-inefficient, but if your program runs with it it most probably means that it is correctly allocating and deallocating memory. It is a kind of valgrind checker.

  • USE_KMALLOC: a port (with some modifications, mainly due to the fact that there is no paging in asmc) of kmalloc. Very quick and rather memory-efficient. Basically to better option currently available in asmc (unless you want to debug memory allocation), so also the default one.

A third pack of flags is for controlling the associative array (map) implementation used by asmc.

  • USE_SIMPLE_MAP: the original map implementation, based on lineary arrays, which require a full trasversal of the array for basically every operation. Very slow.

  • USE_AVL_MAP: a new implementation based on AVL trees. In the end it was never finished (because at some point I decided to switch to red-black trees), so it implements a binary search tree, but without rebalancing. Not guaranteed to be balanced, but probably, since most of the times data arrive in random order, it ususally is. Practical performance are comparable with red-black trees.

  • USE_RB_MAP: the final and default implementation, using properly balanced red-black trees.

In theory all three of them should work, with different performances. In practice, only the red-black tree is routinely used and thus tested.

What is inside this repository

  • lib contains a very small kernel, designed to run on a i386 CPU in protected mode (with a flat memory model and without memory paging). The kernel offers an interface for writing to the console and to the serial port, a simple read-only ramdisk and some library routines for later stages.

The kernel can be booted with multiboot, or can simply be loaded at 1 MB and jumped in at its very beginning. The ramdisk must be appended to it in the ar format.

The kernel must be compiled with payload, inside which it jumps after loading. Three payloads are provided, detailed later.

In this directory there are also some C files that in theory enable you to host a payload directly in a Linux process. They were mostly useful in the beginning to test the code in a more friendly environment than a virtual machine, but they are far from being perfect and not very useful nowadays.

  • asmg is an G compiler written in Assembly. G is a custom language I invented for this project, described below in more details. As soon as it is ready, it compiles the file main.g and jumps to main. Here is where most of the development is concentrated nowadays. asmg can be compiled by asmasm. See above for what is implemented in the G environment.

  • asmg0 is an effort at reducing even more asmg binary seed, by introducing a smaller language called G0 between the binary seed and the G language. It is currently a very experimental effort (even more experimental than the rest) and it does not work at all.

  • boot contains a simple bootloader that can be used to boot all of the above (in the minimalistic style of the rest of the project). For the moment is cannot be compiled with asmasm, because it must use some system level opcodes that are not supported by asmasm, so you have to use NASM. It works under QEMU and in line of principle it also work under bare metal, at least those that I tried (old computers that I had around). As already outlined above, this is not tested software, you should never run it on computer that you cannot afford to be erased.

  • attic contains some earlier test code, that is not used any more and it is also probably broken. They are probably not very interesting to most users, and might be removed altogether at some point.

  • test contains some test programs for the Assembly and C compilers contained in asmg.

  • diskfs contains the file that are made available to the virtual file system in asmg.

Design considerations

Ideally the system seed written in Assembly should be as simple and small as possible. Since that is the part that must be already build when the system boots, it should be verifiable by hand, i.e., it should be so simple that a human being can take a printout of the code and of the binary hex dump and check opcode by opcode that the code was translated correctly. This is very tedious, so everything that is not strictly necessary for building later stages should be moved to later stages.

All other design criteria are of smaller concern: in particular efficiency is not a target (all first stages compilers are definitely not efficient, both in terms of their execution time and of the generated code; however, ideally they are meant to be dropped as soon as a real compiler is built).

Also coding style is very inhomogeneus, mostly because I am working with languages with which I had very small prior experience before starting this project (I had never written more than a few Assembly lines together; the G language did not even exist when I started this, because I invented it, so I could not possibly have prior experience). During writing I established my own style, but never went back to fix already written code. So in theory looking at the style you can probably reconstruct the order in which a wrote code.

Why all of this?

Well, the first and most important reason was learning. So far I learnt how to write a basic boot loader, a basic operating system and a few language compilers (for Assembly, G and C). I learnt to write simple Assembly and I invented a G language, that I found pretty satisfying for the specific domain it was written for (more on this below).

Other than that, it bothers me that the fine art of programming is currently based on a misconception: that there are two worlds, the "source world" and the "executable world", and that given the source you can build the executable. This is not completely true: to pass from the source to the executable you need another executable (the compiler). And to compile the compiler, you most often need the compiler itself. In the current situation, if all the executable binary code in the world were erased by some magic power and only the source code remained, we would not be able to rebuild the executable code, because we would not have working compilers.

The aim of the Bootstrappable project is to recover from this situation, i.e., produce a path to rebootstrap all the executable world from the source world that we already have. Source code is knowledge, executable code is a way to use this knowledge, but it is not knowledge itself. It should be derivable for knowledge without having to depend on anything else.

See the site of the Boostrappable project for additional practical and phylosophical reasons. The asmc project is my personal contribution to Bootstrappable.

Of course it is not possible to remove completely the dependency on some executable code for bootstrapping, becuase at some point you have to power up your CPU and you will have to feed it some executable code (which is called the "seed"). The target is to reduce this seed as much as possible, so that it can be directly inspected. Currently asmc is seeded by around 15 KiB of code (plus, unfortunately, the BIOS and the various microcodes and firmwares in the computer, which are not even open in most cases), which is pretty good. Maybe in the future I'll be able to shrink it even more (there is some room for optimization). At some point I would also like to convert it to a free architecture, like RISC-V, but this will require major rewriting of code generation for all compilers and assemblers. I am not aware of completely free and Linux-capable RISC-V implementations, so for the moment I am concentrating on Intel processors.

Beside the Bootstrappable projects (many are listed in the wiki page), one great inspiration for asmc was TCCBOOT, by Fabrice Bellard (the same original author of tinycc). TCCBOOT uses tinycc to build a stripped down version of the Linux kernel at boot time and then executes it, which is kind of what asmc is trying to do, expect that asmc is trying to compile the compiler as well.

The platform interface exposed by the kernel

The kernel and library in the directory lib offer some simple API to later stages, which is described here. All calls follow the usual cdecl ABI (arguments pushed right to left; caller cleans up; return value in EAX or EDX:EAX; stack aligned to 4; EAX, ECX and EDX are caller-saved and the other registers are callee-saved; objects are returned via additional first argument).

  • platform_exit() Exit successfully; it will never return.

  • platform_panic() Exit unsuccessfully, writing a panic error message; it will never return. In the earlier stages there is nearly no error diagnosing facility, so if the program terminates with a panic message you are on your own finding the problem. Next time you want an easy life please write in Java.

  • platform_write_char(int fd, int c) Write character c in file fd. Writing on a filesystem is not supported yet. There are only two virtual files: file 0, which writes in memory, at the address contained in write_mem_ptr, and then increment the address; and file 1, which writes on the console and on the serial port. There is also file 2, which just maps to 1.

  • platform_log(int fd, char *s) Write a NULL terminated string s into fd, by repeatedly calling platform_write_char.

  • platform_open_file(char *fname) Open file fname for reading, returning the associated fd number. Opened files cannot be closed, for the moment.

  • platform_read_char(int fd) Read a char and return from file fd. Return -1 (i.e., 0xffffffff) at EOF.

  • platform_reset_file(int fd) Seek back to the beginning of the file. Other seeks are not supported.

  • platform_allocate(int size) Simple memory allocator returning a pointer to a memory region of at least size bytes. It works in a similar way to sbrk on UNIX platforms, so you cannot return a memory region to the pool, unless it is the last one that was allocated. But you can implement your own malloc/free on top of it, as it is actually later done in G.

  • platform_get_symbol(char *name, int *arity) Return the address of symbol name, panicking if it does not exist. If arity is not NULL, return there the symbol arity (i.e., the number of parameters, which is relevant for the G language, but not for Assembly). The number -1 (0xffffffff) is returned if arity is undefined.

  • platform_setjmp(void *env) Copy the content of the general purpose registers in the buffer pointed by env (which must be at least 26 bytes long). This is used to implement the setjmp call in the C compiler.

  • platform_longjmp(void *env, int status) Restore the content of the general purpose registers from the buffer pointed by env, except EAX which is set to status. This is used to implement the longjmp call in the C compiler.

Another routine is provided when compiling the kernel with asmg:

  • platform_g_compile(char *filename) Compile the G program in filename, panicking if an error is found.

Symbols generated by any of the two compilers can be recovered with platform_get_symbol.

The G compiler also exports a few internal calls to give the G program a little introspection capabilities, used to generate stack traces on assertions. They are not documented and are not to be used other that in these debugging utilities.

The G language

My initial idea, when I begun working on asmc, was to embed an Assembly compiler in the initial seed and then use Assembly to write a C compiler. At some point I realized that bridging the gap between Assembly and C with just one jump is very hard: Assembly is very low level, and C, when you try to write its compiler, is much higher level that one would expect in the beginning. At the same time, Assembly is harder to compile properly then I initially expected: there are quite some syntax variations and opcode encoding can be rather quirky. So it is not the ideal thing to put in the binary seed.

Then I set out to invent a language which could offer a somewhat C like writing experience, but that was as simple as possible to parse and compile (without, of course, pretending to do any optimization whatsoever). What eventually came out is the G language.

In my experience, and once you have ditched the optimization part, the two difficult things to do to compile C are parsing the input (a lot of operators, different priorities and syntaxes) and handling the type system (functions, pointers, arrays, type decaying, lvalues, ...). So I decided that G had to be C without types and without complicated syntax. So G has just one type, which is the equivalent of a 32-bits int in C and is also used for storing pointers, and expressions are written in reverse Polish notation, so it is basically a stack machine. Of course G is very tied to the i386 architecture, but it is not meant to do much else.

The syntax of the G language is exaplained in a dedicated document.

Ported programs

This repository contains the following code ported to G:

Other programs are used by mean of Git submodules (see the contrib directory), so their exact version is encoded in the Git repository itself and it is not repeated here.


Most of the original code I wrote is covered by the GNU General Public License, version 3 or later. Code that was imported from other projects, with or without modifications, is covered by their own licenses, which are usually either the GPL again or very liberal licenses. Therefore, I believe that the combined project is again distributable under the terms of the GPL-3+ license.

Individual files' headers detail the licensing conditions for that specific file. Having taken material from many different sources, I tried my best to respect all the necessary conditions. Please contact me if you become aware of some mistake on my side.


Giovanni Mascellani gio@debian.org