self-modifying code in 2.6 kernel for ppc writes into readonly section

[Becky Bruce]

On Aug 18, 2008, at 3:51 PM, Michael Neuling wrote:

> In message <78ef7ce10808181257u637c5597xaa992b9e4e7a0925 at mail.gmail.com 
> > you wrote:
>> The mmu is still disabled at this point.
>>
>> What is marked as readonly is the text section of the vmlinux file
>> generated when compiling the kernel. And since the code tries to  
>> write
>> to the text section, I assumed it was the reason for the segmentation
>> fault.
>
> Seriously, a seg fault in your emulator is a bug in the emulator!

Mikey is likely right here. I've (unfortunately) done a lot of  
emulator work, and every time I've hit a problem like this, the  
problem has been with the emulator or the emulation environment.  Have  
you isolated the faulting instruction, verified that it's to a  
reasonable address, and tried examining memory at the faulting address  
using your emulator's command interface?

>
>
>> I'm not sure how this is dealt with on real hardware.
>
> The CPU seg faults... :-P

But only if the page is mapped non-writeable.  Even with the MMU on,  
Linux maps itself in as writeable.  It's the OS, it can do whatever it  
wants.  So it just works on real hardware, and it should just work in  
your emulator.

>
>
>> Can somebody please explain how is it supposed to work ? Is it ok to
>> write to text section that you load on real hardware as readonly ?
>> (again, no mmu involved, as it is still turned off, so i'm not sure
>> who's guarding this section against writing)
>
> I'm not sure how this works for 32 bit CPUs, so I can't speak to the
> details of it.
>
> For the 64bit MMU, if you're in real mode (MMU off), nothing can stop
> this from being written.  The kernel ignores the elf sections
> permissions and does it's own mapping but this can only be enforced  
> once
> the MMU is on.

The same is true on 32-bit ppc - the basic MMU architecture is very  
similar if you have a part that has "real mode" (i.e. non-booke).   
There is no way to restrict stores in real mode.

-Becky

>
>
> Mikey
>
>> On Mon, Aug 18, 2008 at 10:19 PM, Michael Neuling  
>> <mikey at neuling.org> wrote:
>>> In message <78ef7ce10808180901v6c694e63xefc37dd97485533 at mail.gmail.com 
>>> > you
> wrote:
>>>> Hello,
>>>>
>>>> First, I'm talkin about the 2.6.11 version. I know arch/ppc is  
>>>> gone in
>>>> latest versions,
>>>> but i assume the code is still the same and just moved to powerpc.
>>>>
>>>> There is a piece of code in the early initialization of the 2.6  
>>>> kernel
>>>> that identifies the cpu type and then tries to eliminate code that
>>>> does not apply to the current cpu. This is done by writing nop's  
>>>> over
>>>> sections of code that are not needed (do_cpu_ftr_fixups in
>>>> arch/ppc/kernel/misc.S)
>>>>
>>>> When I try to run the kernel in a ppc emulator, I get a  
>>>> segmentation
>>>> fault in do_cpu_ftr_fixups. From examining the section headers of  
>>>> the
>>>> vmlinux, the text section is marked as readonly. The piece of code
>>>> above mentioned is trying to write a nop to memory location  
>>>> inside the
>>>> text section which is readonly, so that explains the sigsegv   
>>>> error.
>>>
>>> Any segv in the emulator sounds like a bug in the emulator.
>>>
>>> If the page really is marked read only, then writing to it should  
>>> cause
>>> a page fault.
>>>
>>>> Since the kernel does run on boards with ppc cpu's, can somebody
>>>> explain how come this is actually working ? Or if/where I am  
>>>> mistaking
>>>> with my assumptions ?
>>>>
>>>> Thank you
>>>>
>>>> P.S. please add me in cc in a reply to this message
>>>> _______________________________________________
>>>> Linuxppc-dev mailing list
>>>> Linuxppc-dev at ozlabs.org
>>>> https://ozlabs.org/mailman/listinfo/linuxppc-dev
>>>>
>>>
>>
> _______________________________________________
> Linuxppc-dev mailing list
> Linuxppc-dev at ozlabs.org
> https://ozlabs.org/mailman/listinfo/linuxppc-dev