Re: [PATCH qemu] x86: don't let decompressed kernel image clobber setup_data

[H. Peter Anvin]

On December 30, 2022 11:54:11 AM PST, Borislav Petkov <bp@alien8.de> wrote:
>On Fri, Dec 30, 2022 at 06:07:24PM +0100, Jason A. Donenfeld wrote:
>> Look closer at the boot process. The compressed image is initially at
>> 0x100000, but it gets relocated to a safer area at the end of
>> startup_64:
>
>That is the address we're executing here from, rip here looks like 0x100xxx.
>
>> /*
>>  * Copy the compressed kernel to the end of our buffer
>>  * where decompression in place becomes safe.
>>  */
>>         pushq   %rsi
>>         leaq    (_bss-8)(%rip), %rsi
>>         leaq    rva(_bss-8)(%rbx), %rdi
>
>when you get to here, it looks something like this:
>
>        leaq    (_bss-8)(%rip), %rsi           # 0x9e7ff8
>        leaq    rva(_bss-8)(%rbx), %rdi                # 0xc6eeff8
>
>so the source address is that _bss thing and we copy...
>
>>         movl    $(_bss - startup_32), %ecx
>>         shrl    $3, %ecx
>>         std
>
>... backwards since DF=1.
>
>Up to:
>
># rsi = 0xffff8
># rdi = 0xbe06ff8
>
>Ok, so the source address is 0x100000. Good.
>
>> HOWEVER, qemu currently appends setup_data to the end of the
>> compressed kernel image,
>
>Yeah, you mean the kernel which starts executing at 0x100000, i.e., that part
>which is compressed/head_64.S and which does the above and the relocation etc.
>
>> and this part isn't moved, and setup_data links aren't walked/relocated. So
>> that means the original address remains, of 0x100000.
>
>See above: when it starts copying the kernel image backwards to a higher
>address, that last byte is at 0x9e7ff8 so I'm guessing qemu has put setup_data
>*after* that address. And that doesn't get copied ofc.
>
>So far, so good.
>
>Now later, we extract the compressed kernel created with the mkpiggy magic:
>
>input_data:
>.incbin "arch/x86/boot/compressed/vmlinux.bin.gz"
>input_data_end:
>
>by doing
>
>/*
> * Do the extraction, and jump to the new kernel..
> */
>
>        pushq   %rsi                    /* Save the real mode argument */      
> 0x13d00
>        movq    %rsi, %rdi              /* real mode address */                
>         0x13d00
>        leaq    boot_heap(%rip), %rsi   /* malloc area for uncompression */    
> 0xc6ef000
>        leaq    input_data(%rip), %rdx  /* input_data */                       
> 0xbe073a8
>        movl    input_len(%rip), %ecx   /* input_len */                        
>         0x8cfe13
>        movq    %rbp, %r8               /* output target address */            
> 0x1000000
>        movl    output_len(%rip), %r9d  /* decompressed length, end of relocs 
> */
>        call    extract_kernel          /* returns kernel location in %rax */
>        popq    %rsi
>
>(actual addresses at the end.)
>
>Now, when you say you triplefault somewhere in initialize_identity_maps() when
>trying to access setup_data, then if you look a couple of lines before that 
>call
>we do
>
>       call load_stage2_idt
>
>which sets up a boottime #PF handler do_boot_page_fault() and it actually does
>call kernel_add_identity_map() so *actually* it should map any unmapped
>setup_data addresses.
>
>So why doesn't it do that and why do you triplefault?
>
>Hmmm.
>

See the other thread fork. They have identified the problem already.