Stefan Reinauer <[EMAIL PROTECTED]> writes: > * Nick Jarmany <[EMAIL PROTECTED]> [020211 15:40]: > > This sounds like a great idea > > > > -----Original Message----- > > From: Eric Seppanen [mailto:[EMAIL PROTECTED]] > > Sent: 08 February 2002 17:23 > > To: LinuxBios > > Subject: Re: Questions > > > > > > On Fri, Feb 08, 2002 at 11:19:58AM +0800, Chung Ming wrote: > > > Ron, I wonder if there are possibility to split LinuxBios in two > > > chips, sais a cheapy 64K linear flash ROM that hold basic system > > > initialization and a ATA Flash (coz it's cheaper than DoC) attached at > > > IDE with kernel, and after IDE and RAM got initialized load kernel > > > from ATA flash.. > > > > Here's another odd idea: maybe someone could figure out a way to split > > the linux kernel into two pieces. The first piece could contain all the > > bits needed to initialize hardware and find the boot media, then the > > second part would contain everything else (multiuser support, etc.). > > Probably impossible, but impressive... > > No, no, no! Please don't even think about that. The result can be seen when > looking at the Alpha ARC bootloader Milo. It links drivers from the Linux > kernel directly by emulating the neccessary backend, resulting in 180k lines > of code for the interfaces and machine setup plus an additional amount of > code from the linux kernel that is used to initialize devices. The result > is not really by means smaller than a Linux kernel itself, it won't fit in > a standard 2Mbit flash part. > For such an approach a microkernel design where drivers and core kernel are > strictly seperated would be much easier to use. > Big disadvantage: almost every change in the kernel results in a patch set > for that particular kernel. I'm pretty sure, nobody wants to maintain such > a beast.
Though there is some argument that shrinking the core kernel is good, so if you could get it to be a kernel config option it might be o.k. I can definentily agree about Milo however, we don't want to follow that path. Which is why my rule has been either reuse it all or don't use it. The most interesting argument against using a kernel is that simple polled device drivers tend to be smaller and more reliable then general purpose high performance kernel drivers. As long as you can load a couple of megs in a second the performance is fine. Plan9 looks interesting in that it may be small enough and reusing drivers that are developed for other purposes is a real time savings. Eric
