starting with 2.6.26-rc1 cell_defconfig fails on QS22

starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Adrian Reber]

I have a custom config to build a kernel which runs on JS21/QS21/QS22.
Starting with 2.6.26-rc1 that fails on JS21 (with SLOF). Bisecting lead
me to the following commit:

 commit 366234f657879aeb7a1e2ca582f2f24f3fae9269
 Author: Kumar Gala <galak@kernel.crashing.org>
 Date:   Wed Apr 16 05:52:28 2008 +1000

     [POWERPC] Update linker script to properly set physical addresses

I then tried it with cell_defconfig and it fails with almost the same
error on QS22. I have not restared the bisecting on QS22 because the
error is so similar:

zImage starting: loaded at 0x00400000 (sp: 0x0e16aea0)
Allocating 0x6790c8 bytes for kernel ...
OF version = 'IBM,SLOF,HEAD'
gunzipping (0x01400000 <- 0x00407000:0x0062ce4a)...done 0x60405a bytes

Linux/PowerPC load: 
Finalizing device tree... using OF tree (promptr=0e1004c4)
 

( 700 ) Program Exception [ e1004c4 ]


    R0 .. R7           R8 .. R15         R16 .. R23         R24 .. R31
00000000014073b0   000000000e974200   0000000000000000 0000000000000000   
000000000e16aea0   0000000000638c04   0000000000000000 0000000000000000   
00000000019e2b98   000000000e974200   0000000000000000 0000000001400000   
0000000000000000   b000000000003000   0000000000000000 0000000000638804   
0000000000000000   2000000000000000   0000000000000000 000000000e96f3c0   
000000000e1004c4   0000000000000000   0000000000000000 000000000e1004c4   
000000000e96f3c0   0000000000000000   0000000000000000 0000000000000000   
0000000000638804   0000000000000000   000000000eac5d70 0000000000000000   

    CR / XER           LR / CTR          SRR0 / SRR1        DAR / DSISR
        80000022   00000000014073e8   000000000189e99c 0000000000000000   
0000000020000000   0000000001400000   9000000000083000 00000000   

I have tried it with gcc-3.4.2, gcc-4.1.1 and gcc-4.2.4.

Somehow I expect that I am doing something wrong, because nobody else
has reported something like this yet.

It looks like 2.6.25 was the last version which worked on SLOF based
systems.

		Adrian

Re: starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Benjamin Herrenschmidt]

On Mon, 2008-06-23 at 11:45 +0200, Adrian Reber wrote:
> I have a custom config to build a kernel which runs on JS21/QS21/QS22.
> Starting with 2.6.26-rc1 that fails on JS21 (with SLOF). Bisecting lead
> me to the following commit:
> 
>  commit 366234f657879aeb7a1e2ca582f2f24f3fae9269
>  Author: Kumar Gala <galak@kernel.crashing.org>
>  Date:   Wed Apr 16 05:52:28 2008 +1000
> 
>      [POWERPC] Update linker script to properly set physical addresses
> 
> I then tried it with cell_defconfig and it fails with almost the same
> error on QS22. I have not restared the bisecting on QS22 because the
> error is so similar:
> 
> zImage starting: loaded at 0x00400000 (sp: 0x0e16aea0)
> Allocating 0x6790c8 bytes for kernel ...
> OF version = 'IBM,SLOF,HEAD'
> gunzipping (0x01400000 <- 0x00407000:0x0062ce4a)...done 0x60405a bytes
> 
> Linux/PowerPC load: 
> Finalizing device tree... using OF tree (promptr=0e1004c4)
>  
> 
> ( 700 ) Program Exception [ e1004c4 ]

The program check exception happens at 0xe1004c4 ? That looks like the
OF entry point (promptr)... could it be possible that it got corrupted
somewhat ? The only thing I see above there would be the stack but
I fail to see how it would use that much... 

>     R0 .. R7           R8 .. R15         R16 .. R23         R24 .. R31
> 00000000014073b0   000000000e974200   0000000000000000 0000000000000000   
> 000000000e16aea0   0000000000638c04   0000000000000000 0000000000000000   
> 00000000019e2b98   000000000e974200   0000000000000000 0000000001400000   
> 0000000000000000   b000000000003000   0000000000000000 0000000000638804   
> 0000000000000000   2000000000000000   0000000000000000 000000000e96f3c0   
> 000000000e1004c4   0000000000000000   0000000000000000 000000000e1004c4   
> 000000000e96f3c0   0000000000000000   0000000000000000 0000000000000000   
> 0000000000638804   0000000000000000   000000000eac5d70 0000000000000000   
> 
>     CR / XER           LR / CTR          SRR0 / SRR1        DAR / DSISR
>         80000022   00000000014073e8   000000000189e99c 0000000000000000   
> 0000000020000000   0000000001400000   9000000000083000 00000000   
> 
> I have tried it with gcc-3.4.2, gcc-4.1.1 and gcc-4.2.4.
> 
> Somehow I expect that I am doing something wrong, because nobody else
> has reported something like this yet.
> 
> It looks like 2.6.25 was the last version which worked on SLOF based
> systems.
> 
> 		Adrian
> _______________________________________________
> Linuxppc-dev mailing list
> Linuxppc-dev@ozlabs.org
> https://ozlabs.org/mailman/listinfo/linuxppc-dev

Re: starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Jeremy Kerr]

Hi Adrian,

> Somehow I expect that I am doing something wrong, because nobody else
> has reported something like this yet.

Seems to work fine here - just booted 945754a on a QS22 with 
cell_defconfig:

[jk@qs22 ~]$ cat /proc/version
Linux version 2.6.26-rc7-09992-g945754a (jk@build) (gcc version 4.1.3 
20080114 (prerelease) (Debian 4.1.2-19)) #154 SMP Tue Jun 24 09:32:59 
EST 2008

Reverting that commit definitely fixes the problem for you? Perhaps a 
binutils problem then?

Cheers,


Jeremy

Re: starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Segher Boessenkool]

>> Finalizing device tree... using OF tree (promptr=0e1004c4)
>>
>>
>> ( 700 ) Program Exception [ e1004c4 ]
>
> The program check exception happens at 0xe1004c4 ? That looks like the
> OF entry point (promptr)... could it be possible that it got corrupted
> somewhat ? The only thing I see above there would be the stack but
> I fail to see how it would use that much...

I don't know what this [ e1004c4 ] is, if I read the current public
SLOF code (for JS2x) correctly it seems to be whatever was on the
stack (or just below the stack) below the error code that was thrown.
Maybe some errors put something interesting there, dunno.

/me looks deeper...  Nastiness :-)  So, a decrementer exception gives
the current decrementer value as reason code; an external exception
reads some BE-specific registers (on any system!); and all other
exceptions use whatever was in GPR4?

Anyway, the register dump shows:

>>     CR / XER           LR / CTR          SRR0 / SRR1        DAR / 
>> DSISR
>>         80000022   00000000014073e8   000000000189e99c 
>> 0000000000000000
>> 0000000020000000   0000000001400000   9000000000083000 00000000

so SRR0 is 0189e99c, which is where the exception happened.  Does
objdump show what's going on?  It seems to happen almost immediately
after the kernel starts, given the CTR value.  If the SRR0 address
doesn't help, the LR address should.

>> I have tried it with gcc-3.4.2, gcc-4.1.1 and gcc-4.2.4.

The binutils version is more interesting here.  2.18?


Segher

Re: starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Adrian Reber]

On Tue, Jun 24, 2008 at 09:45:32AM +1000, Jeremy Kerr wrote:
> > Somehow I expect that I am doing something wrong, because nobody else
> > has reported something like this yet.
> 
> Seems to work fine here - just booted 945754a on a QS22 with 
> cell_defconfig:
> 
> [jk@qs22 ~]$ cat /proc/version
> Linux version 2.6.26-rc7-09992-g945754a (jk@build) (gcc version 4.1.3 
> 20080114 (prerelease) (Debian 4.1.2-19)) #154 SMP Tue Jun 24 09:32:59 
> EST 2008
> 
> Reverting that commit definitely fixes the problem for you? Perhaps a 
> binutils problem then?

D'oh. I spent so much time and the solution is so easy.

Thanks everybody. I have updated to binutils-2.18.50.0.6-2.ppc and now
it works again on QS22 and JS21.

		Adrian

Re: starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Segher Boessenkool]

> D'oh. I spent so much time and the solution is so easy.
>
> Thanks everybody. I have updated to binutils-2.18.50.0.6-2.ppc and now
> it works again on QS22 and JS21.

So I checked your binary, and the only differences between working and
not-working are a) some section offsets in the file, and b) the load
address of the segment (0 vs. 0xc000...).

a) should be harmless, and b) likely is a binutils bug that got fixed.

What is the exact binutils version you used for the non-working?  Should
we test for it in our Makefile?  If we require 2.18, we can drop the 
tests
for 2.12 and 2.14 ;-)


Segher

Re: starting with 2.6.26-rc1 cell_defconfig fails on QS22

[Adrian Reber]

On Wed, Jun 25, 2008 at 05:59:33PM +0200, Segher Boessenkool wrote:
>> D'oh. I spent so much time and the solution is so easy.
>>
>> Thanks everybody. I have updated to binutils-2.18.50.0.6-2.ppc and now
>> it works again on QS22 and JS21.
>
> So I checked your binary, and the only differences between working and
> not-working are a) some section offsets in the file, and b) the load
> address of the segment (0 vs. 0xc000...).
>
> a) should be harmless, and b) likely is a binutils bug that got fixed.
>
> What is the exact binutils version you used for the non-working?  Should

I had two old Fedora installations. Both installations have been
updated, but looking at the installation tree it must have been
something like 2.16.91.0.6 and 2.17.50.0.3

> we test for it in our Makefile?  If we require 2.18, we can drop the tests
> for 2.12 and 2.14 ;-)

		Adrian