OSDev.org

So, I'm compiling my 32-bit bootstrap code with the following commands:


Then compiling my 64-bit kernel with:


Finally, I link everything with:


At which point I get an error message:


And then, a couple undefined references in boot32.o to symbols defined in the 64-bit code. Now, what is the best way to link this together? I know I could manually parse the 64-bit symbol table, or even use objcopy somehow to move the 64-bit sections into the 32-bit object file by hand, with the two exported symbols at defined addresses. But this looks extremely ugly to do, and I think I'm missing something about 32- and 64-bit executable formats.

Also, I'm using Multiboot, so the bootstrap code *must* be 32-bit, and the entire image loadable by the bootloader.

Compile everything as 64 bit. Link as 64 bit too. In your .S files add .code32 directive in places where you have 32 bit code and .code64 where you have 64 bit code (I am not very familiar with GNU assembler, but I think, this is how you do it). There is even .code16 directive if you need it. Loader file can be either raw binary, 32 bit ELF or 64 bit ELF. QEMU's built-in loader can't load 64 bit ELF files (not without a patch), but GRUB can. You will still start in 32 bit mode, but ELF file itself can be 64 bit.
Try lower half kernel first, since high half one needs some extra linker script tweaks and compiler flags.

Have a look at this wiki article, but be sure to read the linked material too: https://wiki.osdev.org/Creating_a_64-bit_kernel

I use Grub to start a 32-bit "kernel" which then jumps to a 64-bit module, as described in the article. Some information needs to be passed to the 64-bit code; I do this in a structure located at a fixed address (which can be overwritten once the data has been transferred).

I think I've come up with a solution:



It links everything together and emits no errors. All symbols are linked properly and have the right addresses as far as I've tested. I still haven't managed to get the 64-bit code to run though...

Okay, now it works perfectly. Does anybody want me to put this in the wiki? I mean, the solution wasn't really that obvious, it would be useful to have it in "Creating a 64-bit kernel" in case somebody wants to code (parts of) their 32-bit bootstrap in C and not just use assembly and ".code32" or have to write a dedicated loader.

Correct me if I am wrong, but I think this method may be a little bit problematic for high half kernels and address spans larger that 4GiB. I mean, 32 bit ELF expects 32 bit addresses. Some relocation types may be impossible if you use 64 bit addresses in 32 bit ELF.

Don't worry, my original idea was to put my kernel in the first 4 GiB so addresses were the same in 32 and 64 bit modes. In fact, I'll possibly make a lower half kernel. I know it's a bad idea, but I think I know what I'm doing. Well, if someone finds out how to link this better, please tell me