Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv

[Mark Hatle <mark.hatle@windriver.com>]

On 3/26/13 5:38 AM, Phil Blundell wrote:
> On Mon, 2013-03-25 at 16:47 -0500, Mark Hatle wrote:
>> I've looked at the code some more.  It does checksum the header itself and then
>> if it's got contents, it also adds the contents to the checksum.. That is where
>> the failure appears to be happening:
>>
>>             if (u.shdr.sh_type != SHT_NOBITS)
>>               {
>>                 Elf_Data *d = elf_rawdata (dso->scn[i], NULL);
>>                 if (d == NULL)
>>                   goto bad;
>>                 process (d->d_buf, d->d_size);
>>               }
>>
>> So it's specifically checking for SHT_NOBITS, but it's matching so it falls
>> through and d->d_buf == 0, causing the failure.  I'll keep investigating, but
>> somehow that value (u.shdr.sh_type) is wrong [or at least unexpected!].
>
> Ah.  If a NOBITS section is getting reported as something else then it
> seems that this must clearly be a bug in libelf and ought to be fixed
> there rather than working around it in rpm.  What do you actually get as
> u.shdr.sh_type?

I posted an updated patch last night.  debugedit was re-translating (byte 
swapping) the header elements during the buildid calculations.  This was causing 
the value of sh_type to be in the ELF binaries native endian during the check. 
So big endian binaries failed to match the SHT_NOBITS, thus the problem being 
discovered on PPC and MIPS.

(Fix was to not translate the original values, but instead only translate a 
copy.  I verified the produced buildid was expected, as I now have a better 
understanding of how the buildid is generated.)

--Mark

> p.
>