How kernel, compiler, and C library work together
Kernel
The kernel is the core of an operating system. In a traditional design, it is responsible for memory management, I/O, interrupt handling, and various other things. And even while some modern designs like Microkernel or Exokernels move several of these services into user space, this matters little in the scope of this document.
The kernel makes its services available through a set of system calls; how they are called and what they do exactly differs from kernel to kernel.
C Library
One thing up front: When you begin working on your kernel, you do not have a C library available. You must not #include anything you did not write yourself. You will also have to port an existing C library. See GCC Cross-Compiler, Future Steps / Standard Library.
The C library implements the standard C functions (i.e., the things declared in <stdlib.h>, <math.h>, <stdio.h> etc.) and provides them in binary form suitable for linking with user-space applications.
In addition to standard C functions (as defined in the ISO standard), a C library might (and usually does) implement further functionality, which might or might not be defined by some standard. The standard C library says nothing about networking, for example. For Unix-alike systems, the POSIX standard defines what is expected from a C library; other systems might differ fundamentally.
It should be noted that, in order to implement its functionality, the C library must call kernel functions. So, for your own OS, you can of course take a ready-made C library and just recompile it for your OS - but that requires that you tell the library how to call your kernel functions, and your kernel to actually provide those functions. The good news is that relatively few of the library's functions do use some system call
Available libraries include the GNU C library (with info about porting the glibc), newlib (with info on the required OS functions detailed in the manual), and uClibC (although that is highly optimized to be used with an embedded Linux).
You may also wish to check the 'OS for Dummies' limited library, which should be enough for kernel programming
And finally, you have fdlibm and dietlib that can provide you minimun standard support if you're fine with GPL.
A more elaborate example of LibraryCalls is available on a different page.
Compiler / Assembler
An Assembler takes (plaintext) source code and turns it into (binary) machine code; more precisely, it turns the source into object code, which contains additional information like symbol names, relocation information etc.
A compiler takes higher-level language source code, and either directly turns it into object code, or (as is the case with GCC) turns it into Assembler source code and invokes an Assembler for the final step.
The resulting object code does not yet contain any code for standard functions called. If you included e.g. <stdio.h> and used printf(), the object code will merely contain a reference stating that a function named printf() (and taking a const char * and a number of unnamed arguments as parameters) must be linked to the object code in order to receive a complete executable.
Some compilers use standard library functions internally, which might result in object files referencing e.g. memset() or memcpy() even though you did not include the header or used a function of this name. You will have to provide an implementation of these functions to the linker, or the linking will fail (see below).
Some advanced operations (e.g. 64-bits divisions on a 32-bits system) might involve compiler-internal functions. For GCC, those functions are residing in libgcc.a. The content of this library is agnostic of what OS you use, and it won't taint your compiled kernel with licensing issues of whatever sort, so you are welcome to locate "libgcc.a" and link it with your kernel.
Linker
A linker takes the object code generated by the compiler / assembler, and links it against the C library (and / or libgcc.a or whatever link library you provide). This can be done in two ways: static, and dynamic.
Static Linking
When linking statically, the linker is invoked during the build process, just after the compiler / assembler run. It takes the object code, checks it for unresolved references, and checks if it can resolve these references from the available libraries. It then adds the binary code from these libraries to the executable; after this process, the executable is complete, i.e. when running it does not require anything but the kernel to be present.
On the downside, the executable can become quite large, and code from the libraries is duplicated over and over, both on disk and in memory.
Dynamic Linking
When linking dynamically, the linker is invoked during the loading of an executable. The unresolved references in the object code are resolved against the libraries currently present in the system. This makes the on-disk executable much smaller, and allows for in-memory space-saving strategies such as shared libraries (see below).
On the downside, the executable becomes dependent on the presence of the libraries it references; if a system does not have those libraries, the executable cannot run.
A popular strategy is to share dynamically linked libraries across multiple executables. This means that, instead of attaching the binary of the library to the executable image, the references in the executable are tweaked, so that all executables refer to the same in-memory representation of the required library.
This requires some trickery. For one, the library must either not have any state (static or global data) at all, or it must provide a separate state for each executable. This gets even trickier with multi-threaded systems, where one executable might have more than one simultaneous control flow.
Second, in a virtual memory environment, it is usually impossible to provide a library to all executables in the system at the same virtual memory address. To access library code at an arbitrary virtual address requires the library code to be position independent (which can be achieved e.g. by setting the -PIC command line option for the GCC compiler). This requires support of the feature by the binary format (relocation tables), and can result in slightly less efficient code on some architectures.
ABI - Application Binary Interface
The ABI of a system defines how library function calls and kernel system calls are actually done. This includes e.g. whether parameters are passed on the stack or in registers, how function entry points are located in libraries etc.
When using static linkage, the resulting executable is depending on the executing kernel using the same ABI as the one it was built for; when using dynamic linkage, the executable is depending on the libraries' ABI staying the same.
Unresolved Symbols
The linker is the stage where you will find out about stuff that has been added without your knowledge, and which is not provided by your environment. This can include references to alloca(), memcpy(), or several others. This is usually a sign that either your toolchain or your command line options are not correctly set up for compiling your own OS kernel - or that you are using functionality that is not yet implemented in your C library / runtime environment!
It is strongly recommended to build a GCC Cross-Compiler to avoid this and similar problems right from the start.
Other kind of symbols such as _udiv* or __builtin_saveregs are available in the OS-agnostic "libgcc.a". If get error about missing such symbols, try to link your kernel against that library too. See this thread for details.
__alloca, ___main
alloca() is a "semi-standard" C function (from BSD?, but supported by most C implementations) that is used to allocate memory from the stack. On Windows this function is also used for stack probing. As such, alloca() is referenced in PE binaries, build e.g. by the Cygwin GCC. You can set -mno-stack-arg-probe to suppress those references.
Another "specialty" of PE binaries is that, if you define int main(), a function void __main() is called first thing after entering main(). You can either define that function, or omit main() from your kernel code, using a different function as entry point.
This explanation of alloca() comes from Chris Giese, posted to alt.os.dev:
>> I think _alloca() is for stack-probing, which is required by Windows.
> What is stack probing?
By default, Windows reserves 1 meg of virtual memory for the stack. No page of stack memory is actually allocated (commited) until the page is accessed. This is demand-allocation.
The page beyond the top of the stack is the guard page. If this page is accessed, memory will be allocated for it, and the guard page moved downward by 4K (one page). Thus, the stack can grow beyond the initial 1 meg.
Windows will not, however, let you grow the stack by accessing discontiguous pages of memory. Going beyond the guard page causes an exception. Stack-probing code prevents this.
Some more information about stack-probing:
- http://groups.google.com/groups?&selm=702bki%24oki%241%40news.Eindhoven.NL.net
- http://groups.google.com/groups?&selm=3381B63D.6E39%40austin.finnigan.com
alloca() is not just used for stack-probing, but also as a sort of malloc() -- for dynamically allocating memory for variables/buffers -- without the need to manually free the reserved memory (with free() as it should be done for malloc() allocated memory).
If you search the man pages for alloca on a UNIX OS you'll find something like this: "The alloca() function allocates space in the stack frame of the caller, and returns a pointer to the allocated block. This temporary space is automatically freed when the function from which alloca() is called returns."
On Windows:
"_alloca allocates size bytes from the program stack. The allocated space is automatically freed when the calling function exits (not when the allocation merely passes out of scope). Therefore, do not pass the pointer value returned by _alloca as an argument to free."
Note that "A stack overflow exception is generated if the space cannot be allocated." so alloca() might cause the stack to grow.
Unfortunately, there are all kinds of caveats with using alloca(). Notably, Turbo~C/C++ had the limitation that functions calling alloca() needed to have at least one local variable, and Linux man-page warns the following:
The alloca function is machine and compiler dependent. On many systems its implementation is buggy. Its use is discouraged.
On many systems alloca cannot be used inside the list of arguments of a function call, because the stack space reserved by alloca would appear on the stack in the middle of the space for the function arguments.
memcpy
This function is used internally by GCC. You should set the =--no-builtin= switch, and provide your own implementation of =memcpy()= if you want to be independent of GCC (OSD library just has it ;).