Prototype Kernel and modules

Recently I've picked up one of my old projects and restarted it, pretty much from scratch. I've been sick for awhile, so I've had time to crack down hard and implement tons of functionality. However one thing that I feel would be a good idea to implement is module loading. I want to do kernel mode dynamic loading of modules.

The word modules is a bit ambiguous, the correct term would just be to load libraries, such as a miniture implementation of the C library for kernel mode drivers or standard things like the PIT and keyboard which are on IRQ 0 & 1. The method I'm trying to achieve is a bit self-sustaining; in the aspect that the modules my kernel will load, will be used in the kernel itself to get into user mode.

As an example, my kernel uses very few functions from the C library, which I've implemented myself. These functions themselfs are used in the setup of my GDTs, IDTs, IRQs, ISRs etc, etc. I would like to abstract these functions to a library that the ker开发者_运维问答nel can load and use. Which means the kernel itself will require module loading at the very first stage, before anything is setup.

Now, I've thought of a few ways to do this myself, such as adding a structure to this library with a table of function pointers that are assigned the address of the functions in the library itself. Compiling the library as an aout-kludge file, loading the library into the kernel as a void * ( which is okay since I have a working allocator ), and then figuring out the offset of the structure, stepping into the void pointer that much, and recreating the structure in the kernel. This does not sound like it would work, since the table of function pointers need to be assigned, which means there needs to be an initialize function in the library itself. How would that be called, even if I knew the address?

I'm clueless as to how I could implement such a loader, and is it even worth it? I want to abstract as much as I possibly can, my kernel has a modular design. I also do expect to load drivers and other things with this method, I'm just unsure how I would implement it. I tried various methods already, and they all failed. What should I do?

I would recommend you first write a dynamic loader in user space. The techniques needed are very similar, and you may be able to adapt much of the code to kernel space later. Also, don't use a.out and don't make up your own 'table of function pointers' - use a more modern format such as ELF. The compile-time tools already exist, so this will save you a lot of effort; you can just write an appropriate linker script and build straight from a Linux GCC.

As it happens, the Windows kernel does something very similar to what you say - the Windows kernel (ntoskrnl.exe) is a PE executable file linking in routines from various DLLs (PSHED.dll, HAL.dll, KDCOM.dll, CLFS.sys, and Cl.dll on my system). In this case, the NTLDR program loads all files required by ntoskrnl.exe into memory, and a boot stub in ntoskrnl.exe then performs dynamic linking. Later the same dynamic linker can be used to load other drivers as well.

Implementing kernel modules is not a trivial job. It is a little complicated and you will need to read to ELF documentation for coding. I will try to provide you some insight here -

In user-space, executable files required shared libraries to implement some of their functionality or code. So, the code in the executable will reference code in the shared library. This leads to the development of symbols. Symbols represent pointers to data/functions/other & have a name.

CHAR VariableName[20];

For example, in the above code a data symbol will be created with the name 'VariableName'. After the 'invention' of dynamic/shared libraries the symbol table (set of symbols in a binary) has to be loaded to resolve the references from executable file in the libraries. But a lot of debugging symbols & useless symbols are present in the symbol table.

Symbol: Main.c

For example, in the symbol table, even a symbol for the C source file will be present for debugging. But that is not required for resolving references at runtime. Here, the concept of dynamic symbols comes in the play.

Dynamic linking refers to the resolving of references between binaries. The dynamic linker will use the dynamic symbol table (which must be loaded & 'normal' symbol table doesn't have to) to resolve the references that the executable makes in the library.

Now, in the kernel core which is the executable file, there are no references in the shared libraries (kernel modules). But the shared libraries reference in the executable. So, the executable must contain dynamic symbols for resolving the references in the kernel modules. This is contrary to the case in user-space. Thus, if you are using ld,

-pie -T LinkerScript.ld

option should be used for create a dynamic symbol table in the kernel executable.

And you should create a LinkerScript.ld file -

/* File: LinkerScript.ld */
PHDRS {
  kernel PT_LOAD FILEHDR;/* This declares a segment in which your code/data is.*/
  dynamic PT_DYNAMIC;/* Segment containing the dynamic table (not DST). */
}

SECTIONS {
    /* text, data, bss sections must be implemented already */
    .dynamic ALIGN(0x1000) : AT(ADDR(.dynamic) - KERNEL_OFFSET)
    { 
       *(.dynamic) 
    } :dynamic/* add :kernel to text, data, bss*/
}

with the above structure. Make sure, your .text, .data & .bss sections are already present & :kernel is added to the end of the section descriptor.

For more information, read the ELF documentation & LD manual (for linker script insight).