Linux/i386 boot protocol version 2.04

I have been doing some very weird things with booting the Linux kernel for a long time.
- Entering the kernel in 32bit mode to avoid 16bit BIOS calls.
- Converting bzImage into static ELF executables.
- Hard coding a kernel command-line
- Going back to 16bit mode to make BIOS calls if necessary.

This version of the boot protocol should be fully backwards compatible but has new capabilities so I can do all of the above cleanly.

The current plan is to send this to Linus in the next couple of days as soon as he gets back.

The patch series is at:
ftp://download.lnxi.com/pub/src/linux-kernel-patches/boot/

The overall patch is:
ftp://download.lnxi.com/pub/src/linux-kernel-patches/boot/linux-2.5.7.boot.diff

Anyway please tell me what you think.

Eric

This is a log of a series of patches that cleans up and enhances the x86 boot process.

2.5.7.boot.linuxbios 8
============================================================
Support for reading information from the linuxbios table. For now I just get the memory size more to come as things evolve.

2.5.7.boot.proto 7
============================================================
Update the boot protocol to include:
- A compiled in command line
- A 32bit entry point
- File and memory usage information enabling a 1 to 1 conversion between the bzImage format and the static ELF executable file format.

- In setup.c split the parameters between those that are compiled in and those that are

2.5.7.boot.build 6
============================================================
Rework the actual build/link step for kernel images.
- remove the need for objcopy
- Kill the ROOT_DEV Makefile variable, the implementation was only half correct and there are much better ways to specify your root device than modifying the kernel.
- Don't loose information when the executable is built

Except for a few extra fields in setup.S the binary image is unchanged.

2.5.7.boot.heap 5
============================================================
Modify video.S so that mode_list is also allocated from the boot time heap. This probably saves a little memory, and makes a compiled in command line a sane thing to implement.

- Made certain we don't overwrite code with the E820_MAP

- Changed the lables around the setup.S to _setup && _esetup

2.5.7.boot.pic16 4
============================================================
All changes are syntactic the generate code should not be affected at all.

- Modify the 16 bit code files bootsect.S video.S setup.S so they may linked with any virtual address, not just 0. The code is already PIC this just makes the build process the same.

- e820.h Add define E820ENTRY_SIZE

- Add define KERNEL_START in setup.S so if I need this value more than once it is easy to get at.

2.5.7.boot.32bit_entry 3
============================================================
- trampoline.S fix comments, and enter the kernel at secondary_startup_32 instead of startup_32
- trampoline.S fix gdt_48 to have the correct gdt limit
- Save all of the registers we get from any 32bit entry point, and don't assume they have any particular value.
- head.S split up startup_32
- secondary_startup_32 handles the SMP case
- move finding the command line to startup.c
- Don't copy the kernel parameters to the initial_zero_page, instead just pass setup.c where they are located.
All of these are what it takes to remove the assumptions of what register values we get on entry. And let's us handle those assumptions up in C code.
- Seperate the segments used by setup.S from the rest of the kernel. This way bootloader can continue to make assumptions about which segments setup.S uses while the rest of the kernel can do whatever is convinient.
- Move boot specific defines into boot.h

2.5.7.boot.vmlinuxlds 2
============================================================
- i386/Makefile remove bogus linker command line of -e stext
- Fix vmlinux.lds so vmlinux knows it loads at 0x100000 (1MB)
- Fix vmlinux.lds so we correctly use startup_32 for our entry point

2.5.7.boot.boot_params 1
============================================================
- Introduce asm-i386/boot_param.h and struct boot_params
- Implement struct boot_params in misc.c & setup.c

This removes a lot of magic macros and by keeping all of the boot parameters in a structure it becomes much easier to track which boot_paramters we have and where they live. Additionally this keeps the names more uniform making grepping easier.

On 30 Mar 2002, Eric W. Biederman wrote:

| I have been doing some very weird things with booting the Linux kernel | for a long time.

so you are saying that we want these 'weird' things in the baseline kernel? ;)

Will you please provide a one-sentence explanation for each of these items? (not "what," but "why" it's good to have it)

|   - Entering the kernel in 32bit mode to avoid 16bit BIOS calls.
|   - Converting bzImage into static ELF executables.
|   - Hard coding a kernel command-line
|   - Going back to 16bit mode to make BIOS calls if necessary.
|
| This version of the boot protocol should be fully backwards compatible
| but has new capabilities so I can do all of the above cleanly.
|
| The current plan is to send this to Linus in the next couple of days
| as soon as he gets back.
|
...
|
| Anyway please tell me what you think.
|
| Eric
|
|
| This is a log of a series of patches that cleans up and enhances the
| x86 boot process.

[snippage]

| 2.5.7.boot.proto 7
| ============================================================
| Update the boot protocol to include:
|    - A compiled in command line

when and how is the command line specified? at build time?

maybe in a kernel.command.line file so that I don't have to type it in every time?

|
| 2.5.7.boot.boot_params 1
| ============================================================
| - Introduce asm-i386/boot_param.h and struct boot_params
| - Implement struct boot_params in misc.c & setup.c

Yep, I like that one.

| This removes a lot of magic macros and by keeping all of the
| boot parameters in a structure it becomes much easier to track
| which boot_paramters we have and where they live.  Additionally
| this keeps the names more uniform making grepping easier.

I'll try to look over the patch files too.

~Randy

We want the ability yes. The overhead is quite small, the code is clean, and it makes the kernel more maintainable. The mainline handles weird cases so Linux will just work. Without these we aren't handling the hard cases.

If you don't have an pc compatible BIOS making 16bit BIOS calls is fatal. Additionally there are boot environments where switching to 16bit to enter the kernel is just silly.

Bean counting for bootloaders, my patch simply increases the visibility of what the Linux kernel does while loading. Being able to actually understand what the kernel is doing while loading makes it much easier to trouble shot problems.

I'm a big fan of making a giant static binary with the command line, ramdisk and everything included and just asking the bootloader to read it from wherever and put it in memory exactly where I said. And failing that to give me an Error message. The ELF file format doesn't have much overhead and provides a general way for me to do that.

The biggest thing this allows is with some fixing on how we allocate the bootmem bitmap bootloaders will be able to put the initial ramdisk right after where the kernel is in memory, without a decrease in bootability for the low memory case. Which if we had done it that way in the first place would have prevented our problems on high memory machines.

There is one byte of overhead with this a terminating null. For cases with stupid bootloaders, or users who can't figure out their
bootloaders this is handy. Additionally doing this reduces the
memory fragmentation potentially allowing video.S to have a larger heap.

At compile time. You can put in .config . I copied the config interface from the other Linux ports that already have this functionality. There is enough information present in the bzImage header that you can edit have a utility to edit it later if you want.

Thanks.

Eric