x86_32: CONFIG_PHYSICAL_START problem

x86_32: CONFIG_PHYSICAL_START problem

[Randy Dunlap]

background:

I was trying to debug a MIPS build error, but it wasn't MIPS-specific,
so I did this, using the MIPS .config file:

make ARCH=i386 O=xx32 olddefconfig

Little to my knowledge, this came up with
CONFIG_PHYSICAL_START=0x81000000

Then I built the i386 kernel, and got this message:
ld: kernel image bigger than KERNEL_IMAGE_SIZE

so I promptly changed many =y drivers etc. to =m
and still got the same ld error message.

Well, it must be something else, he said.

I tracked it down to this large value of CONFIG_PHYSICAL_START
and changed it back to its default value, then the kernel
built with no problems.

So far I haven't been able to track the chain of values/changes
that involve PHYSICAL_START, __PAGE_OFFSET, LOAD_OFFSET, etc.

Anyway, I would like to see PHYSICAL_START limited to some
acceptable range of values in arch/x86/Kconfig,
or at a minimum, a little bit better error message coming
from arch/x86/kernel/vmlinux.lds.S:

. = ASSERT((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
	   "kernel image bigger than KERNEL_IMAGE_SIZE");

so maybe:

. = ASSERT((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
	   "kernel image bigger than KERNEL_IMAGE_SIZE or load address is too large");
(or start address)

Comments?

thanks.
-- 
~Randy
Reported-by: Randy Dunlap <rdunlap@infradead.org>

Re: x86_32: CONFIG_PHYSICAL_START problem

[Thomas Gleixner]

On Sat, Dec 12 2020 at 22:29, Randy Dunlap wrote:
> Little to my knowledge, this came up with
> CONFIG_PHYSICAL_START=0x81000000
...
> I tracked it down to this large value of CONFIG_PHYSICAL_START
> and changed it back to its default value, then the kernel
> built with no problems.
>
> So far I haven't been able to track the chain of values/changes
> that involve PHYSICAL_START, __PAGE_OFFSET, LOAD_OFFSET, etc.

It's convoluted :)

> Anyway, I would like to see PHYSICAL_START limited to some
> acceptable range of values in arch/x86/Kconfig,

Its the combination of CONFIG_PHYSICAL_START and CONFIG_PAGE_OFFSET so
having checks in Kconfig is going to be a challenge.

> or at a minimum, a little bit better error message coming
> from arch/x86/kernel/vmlinux.lds.S:
>
> . = ASSERT((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
> 	   "kernel image bigger than KERNEL_IMAGE_SIZE");
>
> so maybe:
>
> . = ASSERT((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
> 	   "kernel image bigger than KERNEL_IMAGE_SIZE or load address is too large");
> (or start address)

Some reasonable error message is definitely due, but we probably can
catch that at the compile stage already once we computed the combined
values. Then we can sanity check those for resonable ranges.

Thanks,

        tglx