GSoC'24: Linux x86 Boot Protocol Support

The Linux boot protocol plays an important role in the initialization of the Linux operating system, emphasizing the importance of system optimization and scalability.

Setup in Unikraft#

First of all, I added the following lines to the lxboot.S file from Unikraft:

diff --git a/plat/kvm/x86/lxboot.S b/plat/kvm/x86/lxboot.S
index 0ebe18a5..8140804f 100644
--- a/plat/kvm/x86/lxboot.S
+++ b/plat/kvm/x86/lxboot.S
@@ -21,6 +21,13 @@ lcpu_boot_startup_args:
        .quad   0
        .quad   0
 
+.code32
+.section .text.32.boot
+ENTRY(_lxboot_entry32)
+    cli
+    hlt
+END(_lxboot_entry32)
+
 /**
  * 64-bit Linux Boot Protocol entry function
  *

This way, if my header is correct and I do indeed reach Unikraft, the guest should hang.

Starting to debug#

There are 2 binary files that are of interest here: linuxboot.bin and linuxboot_dma.bin. The linuxboot.bin file is optained by compiling linuxboot.S and the linuxboot_dma.bin file is optained by compiling linuxboot_dma.c. As you can see, one is written in x86 AT&T assembly and the other is written in C, by using inline assembly. Naturally, the one written in C would be easier to debug, and luckily, it seems like I do no need the other one.

By adding in linuxboot.S the following code, the QEMU instance should stay in a loop:

diff --git a/pc-bios/optionrom/linuxboot.S b/pc-bios/optionrom/linuxboot.S
index ba821ab922..26df825db8 100644
--- a/pc-bios/optionrom/linuxboot.S
+++ b/pc-bios/optionrom/linuxboot.S
@@ -30,6 +30,8 @@ run_linuxboot:
 
     cli
     cld
+0:
+    jmp        0b
 
     jmp        copy_kernel
 boot_kernel:

I observed that nothing changes in its behaviour, so I added the following lines in linuxboot_dma.c:

diff --git a/pc-bios/optionrom/linuxboot_dma.c b/pc-bios/optionrom/linuxboot_dma.c
index cbcf6679d9..4f12fe359e 100644
--- a/pc-bios/optionrom/linuxboot_dma.c
+++ b/pc-bios/optionrom/linuxboot_dma.c
@@ -55,6 +55,8 @@ asm(
 "_bev:\n"
 "   cli\n"
 "   cld\n"
+"some_label:\n"
+"   jmp some_label\n"
 "   jmp load_kernel\n"
 );

This time the QEMU instance stayed in the loop, so everything was fine. Now, all I had to do was to change the jmp with two nop instructions in gdb and from there to start debugging. Sounds easy, right?

Unexpected challenges#

For some (at the time) unknown reason, when compiling the .bin files didn't update. Was it my fault? Did I have something wrong with my setup?

After a lot of time, I figured out along with my mentors that I needed to have gcc-multilib instead of the one that I had previously. As different versions of gcc clash (installing one uninstalls the other), I had to replace it.

I observed that there are 3 linuxboot.bin and 3 linuxboot_dma.bin in the filesystem. Only one of each was updated when compiling, and it wasn't the one that it searched for by default. Using strace, I figured out that QEMU looks for the .bin file in the directory from which the command runs. Because of this, I copied the linuxboot_dma.bin that was updated where I needed it.

Next steps#

I am going to look what and where the problem is inside the linuxboot_dma.bin file, using gdb. Everything took much more than I expected, so I have to do this now.