Re: ppc64le STRICT_MODULE_RWX and livepatch apply_relocate_add() crashes

[Christophe Leroy <christophe.leroy@csgroup.eu>]

Le 13/12/2021 à 18:26, Joe Lawrence a écrit :
> On 12/13/21 11:36 AM, Christophe Leroy wrote:
>>
>>
>> Le 13/12/2021 à 15:47, Joe Lawrence a écrit :
>>> On 12/13/21 2:42 AM, Christophe Leroy wrote:
>>>>
>>>> Hello Joe,
>>>>
>>>> I'm implementing LIVEPATCH on PPC32 and I wanted to test with
>>>> STRICT_MODULE_RWX enabled so I took your branch as suggested, but I'm
>>>> getting the following errors on build. What shall I do ?
>>>>
>>>>      CALL    scripts/checksyscalls.sh
>>>>      CALL    scripts/atomic/check-atomics.sh
>>>>      CHK     include/generated/compile.h
>>>>      KLP     lib/livepatch/test_klp_convert1.ko
>>>> klp-convert: section .rela.klp.module_relocs.test_klp_convert_mod length
>>>> beyond nr_entries
>>>>
>>>> klp-convert: Unable to load user-provided sympos
>>>> make[2]: *** [scripts/Makefile.modfinal:79:
>>>> lib/livepatch/test_klp_convert1.ko] Error 255
>>>>      KLP     lib/livepatch/test_klp_convert2.ko
>>>> klp-convert: section .rela.klp.module_relocs.test_klp_convert_mod length
>>>> beyond nr_entries
>>>>
>>>> klp-convert: Unable to load user-provided sympos
>>>> make[2]: *** [scripts/Makefile.modfinal:79:
>>>> lib/livepatch/test_klp_convert2.ko] Error 255
>>>>      KLP     lib/livepatch/test_klp_convert_sections.ko
>>>> klp-convert: section .rela.klp.module_relocs.test_klp_convert_mod length
>>>> beyond nr_entries
>>>>
>>>> klp-convert: Unable to load user-provided sympos
>>>> make[2]: *** [scripts/Makefile.modfinal:79:
>>>> lib/livepatch/test_klp_convert_sections.ko] Error 255
>>>> make[2]: Target '__modfinal' not remade because of errors.
>>>> make[1]: *** [scripts/Makefile.modpost:145: __modpost] Error 2
>>>> make: *** [Makefile:1770: modules] Error 2
>>>>
>>>
>>> Hi Christophe,
>>>
>>> Interesting failure mode.  That's klp-convert complaining that it found
>>> more relocations in a .klp.module_relocs.<objname> section than
>>> expected, i.e. nr_entries = sec->size / sizeof(struct klp_module_reloc).
>>>
>>> A few possibilities: the ELF sec->size was incorrectly set/read by
>>> build/libelf (I doubt that).  Or maybe the layout/size of struct
>>> klp_module_reloc is not consistent between kernel and userspace (I'm
>>> more suspicious of this).
>>>
>>> Can you post a copy of the build's symbols.klp and
>>> lib/livepatch/test_klp_convert1.tmp.ko somewhere?  I should be able to
>>> start debug with those files.
>>>
>>
>> I sent you both files off list.
>>
>> It looks like klp-convert doesn't use the correct size. It finds a
>> struct of size 12 hence 3 entries for a section of size 40.
>>
>> On PPC32 the struct has size 8 (void * is 4 and int is 4).
>>
>> But I'm cross-building from x86_64 where the struct is 8 + 4 = 12.
>>
>> Can it be the reason ?
>>
> 
> I'm pretty sure that is it.  I haven't had much runtime with klp-convert
> and cross-building (I've only found one big/little endian bug with
> x86_64->s390x) and was going to ask you how you were testing :)
> 
> Do you know if there are other kernel build tools that deal with similar
> situations?  This seems like a tricky job for the userspace build tool
> to determine non-native target struct layout.
> 
> In the meantime, hacking in:
> 
>   struct klp_module_reloc {
> -       void *sym;
> +       uint32_t sym;
>          unsigned int sympos;
>   } __packed;
> 
> gets me generating an output .ko file, but the readelf output doesn't
> look right.
> 
> I'll add this to the patchset TODO list, but may not get to it for a
> while -- is there any chance the above hack works or could you test a
> local non-cross build?
> 

No I have no way to do a non-cross build. My target is an embedded board 
with slow CPU and little memory.

I tested with your hack, I get:

root@vgoip:~# insmod /lib/modules/test_klp_convert1.ko
insmod: can't insert '/lib/modules/test_klp_convert1.ko': unknown symbol 
in module, or unknown parameter
root@vgoip:~# insmod /lib/modules/test_klp_livepatch.ko
insmod: can't insert '/lib/modules/test_klp_livepatch.ko': unknown 
symbol in module, or unknown parameter


I agree with you readelf shows something went wrong with relocations.

Christophe