[yocto] BBB doesn't boot

[William Mills]

On 04/15/2014 11:31 PM, Khem Raj wrote:
> On Tue, Apr 15, 2014 at 6:20 PM, Denys Dmytriyenko <denis at denix.org> wrote:
>> On Tue, Apr 15, 2014 at 09:12:30PM -0400, Denys Dmytriyenko wrote:
>>> On Wed, Apr 16, 2014 at 12:36:58AM +0100, Richard Purdie wrote:
>>>> On Tue, 2014-04-15 at 15:43 -0400, Denys Dmytriyenko wrote:
>>>>> I don't yet know what is going on, but building in the same directory with
>>>>> sources (B = S) makes it work regarless of the path length:
>>>>>
>>>>> /OE/RAM/poky-111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111/22222222222222222222222222222222222222222222222222222222222222222222/3333333333333333333333333333333333333333333333333333/tmp/work/beaglebone-poky-linux-gnueabi/linux-yocto/3.14+gitAUTOINC+928d7b2dda_0143c6ebb4-r0/linux
>>>>>
>>>>> So, I just commented out setting kernel-specific B in linux-yocto.inc and any
>>>>> kernel now boots with long path:
>>>>>
>>>>> #B = "${WORKDIR}/linux-${PACKAGE_ARCH}-${LINUX_KERNEL_TYPE}-build"
>>>>>
>>>>> I'm copying Richard and Bruce directly to see if they may have a quick insight
>>>>> and/or accept it as a workaround for the release. I'll keep digging further,
>>>>> but if anyone cares to verify the above workaround works for them, I would
>>>>> appreciate. Thanks!
>>>>
>>>> I'm travelling and don't have hardware so I'm limited in what I can look
>>>> at right now. I suspect this workaround "works" because it makes the
>>>> "beaglebone-standard-build" extra characters on the path. I have a
>>>> feeling your -1111111 test above may have reused sstate or something
>>>> like that and given misleading results. I'd be interested in the vmlinux
>>>> file from the build above to see if the poky-111111 pathnames are in
>>>> there (they get stripped out when you create the zImage).
>>>
>>> Nope, I was fooled by sstate once, so this time I tested with cleansstate and
>>> compared timestamps of the kernel when it boots - it is definitely the new
>>> one. And to make sure 'beaglebone-standard-build' extra suffix doesn't affect
>>> it, I increased the path length even further - that extra level with 333 is
>>> there to over-compensate, as it was failing before even with just 111 and 222
>>> levels only... Looks like Gary was also able to verify it.
>>>
>>> But, you are right about vmlinux - it doesn't have those paths in there any
>>> more! I've seen them there when building with B != S, but they are gone when
>>> building with B = S. Wondering why. You can check it for yourself:
>>>
>>> http://arago-project.org/files/short-term/misc/vmlinux-3.14.0-yocto-standard
>>
>> And, as a follow up, all those long paths in vmlinux when built with B != S
>> point to sources. Here are few examples:
>>
>> B != S strings:
>>
>> /OE/RAM/poky/tmp/work/beaglebone-poky-linux-gnueabi/linux-yocto/3.14+gitAUTOINC+928d7b2dda_0143c6ebb4-r0/linux/init/main.c
>> earlycon
>> 4Malformed early option '%s'
>> 3Starting init: %s exists but couldn't execute it (error %d)
>> ...
>> 6Trying to unpack rootfs image as initramfs...
>> /OE/RAM/poky/tmp/work/beaglebone-poky-linux-gnueabi/linux-yocto/3.14+gitAUTOINC+928d7b2dda_0143c6ebb4-r0/linux/init/initramfs.c
>> 6rootfs image is not initramfs (%s); looks like an initrd
>> TRAILER!!!
>> ...
>> %lu.%02lu BogoMIPS (lpj=%lu)
>> /OE/RAM/poky/tmp/work/beaglebone-poky-linux-gnueabi/linux-yocto/3.14+gitAUTOINC+928d7b2dda_0143c6ebb4-r0/linux/arch/arm/vfp/vfpmodule.c
>> 6VFP support v0.3:
>>
>>
>> B = S same strings:
>>
>> init/main.c
>> earlycon
>> 4Malformed early option '%s'
>> 3Starting init: %s exists but couldn't execute it (error %d)
>> ...
>> 6Trying to unpack rootfs image as initramfs...
>> init/initramfs.c
>> 6rootfs image is not initramfs (%s); looks like an initrd
>> TRAILER!!!
>> ...
>> %lu.%02lu BogoMIPS (lpj=%lu)
>> arch/arm/vfp/vfpmodule.c
>> 6VFP support v0.3:
>>
>>
>> So, that explains why B = S works regardless of the location, as it doesn't
>> embed full path to source files...
>>
>
> seemingly there could be rodata segment overflow issue. Can you do the
> kernel build with say linker from dora build
> and see if it still fails.
>

It would be interesting to see if the rodata segment (or segment it 
contributes to) between working and non-working versions were crossing
some magic power of 2 value.

Are there two images published from the same build machines with just
the path difference?  I would like to see one that was just under the
wire and one just a bit too long.  Denys, you have that?