> On Thu, 4 Sep 1997, Teunis Peters wrote:
>
> > Just how difficult (impossible?) would making the kernel itself modular
> > be?
> > What portions of the kernel that are currently not modular could be made
> > replacable with modules?
>
> > Of course, this whole concept might just be too silly to even look at any
> > further....
>
> Not silly at all, on the surface anyway. OpenVMS is modularized somewhat
> like this. I think the "modular Exec" came in with version 5.0, and from
> 5.whatever to 6.0 DEC claimed that they only changed three of the 19 modules.
I run one kernel on several systems. You need one filesystem, one device
driver, and one binary image interpreter to begin. I have an initial
ramdisk, ROMFS, and binfmt_elf as the only nonmodular components since
they are the smallest and most flexible. On the initrd I place my ide
drivers (ide, the disk and the probe modules) and a script to load them.
The main problem is my .config makes *everything* as a module, so when
something breaks (e.g. umsdos), the make fails, and they don't have a real
make -k for the kernel build.
The only major chunk of code that isn't a module is IPV4. I had a patch
(back to 2.1.42) that didn't yet work as a module (it loaded, and gave
kernel panics), but didn't hurt anything when not compiled as a module -
mainly the symbols are scattered between net and the net/ipv4 files. This
probably should be done before 2.2, but I haven't gone much further.
Most of the rest of the kernel except symbols (symbols.o? - it should
work) is needed from the earliest part of the boot and then used all the
time. Other things are the virtualizers that make it possible to load the
real functions (e.g. filesystems).
--- reply to tzeruch - at - ceddec - dot - com ---