When the system is booted, the BIOS reads sector 0 (known as the MBR - the Master Boot Record) from the first disk (or from floppy or CDROM), and jumps to the code found there - usually some bootstrap loader. These small bootstrap programs found there typically have no own disk drivers and use BIOS services. This means that a Linux kernel can only be booted when it is entirely located within the first 1024 cylinders, unless you both have a modern BIOS (a BIOS that supports the Extended INT13 functions), and a modern bootloader (a bootloader that uses these functions when available).
This problem (if it is a problem) is very easily solved: make sure that the kernel (and perhaps other files used during bootup, such as LILO map files) are located on a partition that is entirely contained in the first 1024 cylinders of a disk that the BIOS can access - probably this means the first or second disk.
Thus: create a small partition, say 10 MB large, so that there is room for a handful of kernels, making sure that it is entirely contained within the first 1024 cylinders of the first or second disk. Mount it on /boot so that LILO will put its stuff there.
Most systems from 1998 or later will have a modern BIOS.
Executive summary: If you use LILO as boot loader, make sure you have LILO version 21.4 or later. (It can be found at ftp://metalab.unc.edu/pub/Linux/system/boot/lilo/.) Always use the lba32 option.
An invocation of /sbin/lilo (the boot map installer) stores a list of addresses in the boot map, so that LILO (the boot loader) knows from where to read the kernel image. By default these addresses are stored in (c,h,s) form, and ordinary INT13 calls are used at boot time.
When the configuration file specifies lba32 or linear, linear addresses are stored. With lba32 also linear addresses are used at boot time, when the BIOS supports extended INT13. With linear, or with an old BIOS, these linear addresses are converted back to (c,h,s) form, and ordinary INT13 calls are used.
Thus, with lba32 there are no geometry problems and there is no 1024 cylinder limit. Without it there is a 1024 cylinder limit. What about the geometry?
The boot loader and the BIOS must agree as to the disk geometry. /sbin/lilo asks the kernel for the geometry, but there is no guarantee that the Linux kernel geometry coincides with what the BIOS will use. Thus, often the geometry supplied by the kernel is worthless. In such cases it helps to give LILO the `linear' option. The advantage is that the Linux kernel idea of the geometry no longer plays a role. The disadvantage is that lilo cannot warn you when part of the kernel was stored above the 1024 cylinder limit, and you may end up with a system that does not boot.
With LILO versions below v21 there is another disadvantage: the address conversion done at boot time has a bug: when c*H is 65536 or more, overflow occurs in the computation. For H larger than 64 this causes a stricter limit on c than the well-known c < 1024; for example, with H=255 and an old LILO one must have c < 258. (c=cylinder where kernel image lives, H=number of heads of disk)
Tim Williams writes: `I had my Linux partition within the first 1024 cylinders and still it wouldnt boot. First when I moved it below 1 GB did things work.' How can that be? Well, this was a SCSI disk with AHA2940UW controller which uses either H=64, S=32 (that is, cylinders of 1 MiB = 1.05 MB), or H=255, S=63 (that is, cylinders of 8.2 MB), depending on setup options in firmware and BIOS. No doubt the BIOS assumed the former, so that the 1024 cylinder limit was found at 1 GiB, while Linux used the latter and LILO thought that this limit was at 8.4 GB.
The nuni boot loader does not use BIOS services but accesses IDE drives directly. So, one can put it on a floppy or in the MBR and boot from anywhere on any IDE drive (not only from the first two). Find it at //metalab.unc.edu/pub/Linux/system/boot/loaders/.