On 13 May 2021 at 07:00pm, Christophe Leroy wrote:

Le 13/05/2021 à 18:35, Christian Zigotzky a écrit :

quoted

On 13 May 2021 at 5:51pm, Christophe Leroy wrote:

quoted

Le 13/05/2021 à 17:19, Christian Zigotzky a écrit :

quoted

On 13 May 2021 at 12:01 pm, Christophe Leroy wrote:

quoted

Le 13/05/2021 à 08:47, Christian Zigotzky a écrit :

quoted

Hi Christophe,

quoted

On 9. May 2021, at 19:37, Christophe Leroy 
[off-list ref] wrote:

Did I do an error in my analysis ?

No, you didn’t. On the contrary you have found the issue. ;-) The 
issue is somewhere in the new interrupt code.

I'm not convinced, but let's give it a try.

quoted

quoted

ZZ | * | ceff77efa4f8 powerpc/64e/interrupt: Use new interrupt 
context tracking scheme

Could you please create a patch for reverting the new interrupt 
code? I would like to confirm your result.

Please fetch https://github.com/chleroy/linux.git and try the 
branch for_christian.

This is a revert of approx a dozen of commits around the changes 
to 64e on top of v5.13-rc1.

If the top of the branch works for you, it would be great if you 
can find out which one of the reverts fixes the problem for real.

Thanks
Christophe

It's working! Great! I can use the RC1 on my FSL P5040. Thank you! 
The issue is definitely somewhere in the interrupt code. Please 
tell me the next steps.

Can you bisect between 5.13-rc1 and the top of the 'for_christian' 
branch to find the first 'good' commit ?

Take care it is an "up side down" bisect, a 'good' is one that does 
_not_ work and a 'bad' is a commit that works.

git bisect start
git bisect bad 1c8f441f1485
git bisect good 6efb943b8616

Christophe

Hi Christophe,

Yes, I can. Shall I use the branch 'for_christian' or the default 
linux git for bisecting? I have tried it already with the branch 
'for_christian' but it doesn't compile.

git bisect start
git bisect bad 1c8f441f1485
git bisect good 6efb943b8616

Output: [d66b1d1aab0c1caad11eca417f515b86ec0bebe9] Revert 
"powerpc/64e/interrupt: Use new interrupt context tracking scheme"

Result:

arch/powerpc/kernel/interrupt.o: In function `.syscall_exit_prepare':
interrupt.c:(.text+0x278): undefined reference to `.schedule_user'
arch/powerpc/kernel/interrupt.o: In function 
`.interrupt_exit_user_prepare':
interrupt.c:(.text+0x340): undefined reference to `.schedule_user'
Makefile:1191: recipe for target 'vmlinux' failed
make: *** [vmlinux] Error 1

Ah yes, I remember this problem.

Can you select CONFIG_VIRT_CPU_ACCOUNTING_GEN in your configuration ?

Otherwise, I can try to fix the branch.

Christophe

I selected this. After that it compiles.

1. git bisect good - Xorg restarts again and again
     Output: [f9aa0ac1e9e82b60401ad567bdabc30598325bc1] Revert 
"powerpc/64e/interrupt: use new interrupt return"
2. git bisect good - Xorg restarts again and again
     Output: [cd6d259a14704741bf0cd1dcadb84c0de22d7f77] Revert 
"powerpc/64e/interrupt: always save nvgprs on interrupt"
3. git bisect bad - Xorg works
     Output: [9bfa20ef2ae54d3b9088dfbcde4ef97062cf5ef2] Revert 
"powerpc/interrupt: update common interrupt code for"
4. git bisect good - Xorg restarts again and again
     Output:

cd6d259a14704741bf0cd1dcadb84c0de22d7f77 is the first bad commit
commit cd6d259a14704741bf0cd1dcadb84c0de22d7f77
Author: Christophe Leroy [off-list ref]
Date:   Thu May 13 09:52:06 2021 +0000

     Revert "powerpc/64e/interrupt: always save nvgprs on interrupt"

     This reverts commit 4228b2c3d20e9f80b847f809c38e6cf82864fa50.

:040000 040000 156542c857ad72776b69bb67b2f244afeeb7abd3 
92ea86ed097fce16238b0c2f2b343473894e4e8e M    arch

- Christian