Re: [Linux PPC] Disable PREEMPT

[Shrikanth Hegde <sshegde@linux.ibm.com>]

Hi Michal,

On 9/21/26 10:16 PM, Michal Suchánek wrote:
> On Mon, Sep 21, 2026 at 10:06:22AM +0200, Michal Suchánek wrote:
>> On Thu, Sep 03, 2026 at 03:28:26PM +0200, Michal Suchánek wrote:
>>> On Thu, Sep 03, 2026 at 06:06:16PM +0530, Harsh Prateek Bora wrote:
>>>> + Amit, Gautam - since they recently came across a similar regression which
>>>> is seen with non KVM workload as well.
>>>> + Venkat - for CI improvement suggestion below.
>>>>
>>>> Hi Michal,
>>>>
>>>> On 03/09/26 5:55 pm, Michal Suchánek wrote:
>>>
>>>>>
>>>>> There is another problem that Linux 7.2.2 often locks up on boot inside
>>>>> the guest, and as a result a different kernel is needed in the guest to
>>>>> reproduce the host crash. This is more recent, linux 6.12 does not have
>>>>> this problem. Will try some different kernel combinations to see if I
>>>>> can narrow this down. For this the repro is merely booting the kernel in
>>>>> the VM, it typically fails early before userspace produces any messages.
>>>
>>> This one regressed between 6.19.12
>>> https://github.com/openSUSE/kernel-source/blob/9f6830f/config/ppc64le/default
>>> and 7.0.12
>>> https://github.com/openSUSE/kernel-source/blob/2ebf0bc/config/ppc64le/default
>>
>> Which is when preempt was enabled.
>>
>> Very likely preexisting problem that was exposed by enabling preemption
>> again.
>>
>> @@ -121,9 +122,7 @@ CONFIG_BPF_LSM=y
>>   
>>   CONFIG_PREEMPT_BUILD=y
>>   CONFIG_ARCH_HAS_PREEMPT_LAZY=y
>> -# CONFIG_PREEMPT_NONE is not set
>> -CONFIG_PREEMPT_VOLUNTARY=y
>> -# CONFIG_PREEMPT is not set
>> +CONFIG_PREEMPT=y
>>   # CONFIG_PREEMPT_LAZY is not set
>>   CONFIG_PREEMPT_COUNT=y
>>   CONFIG_PREEMPTION=y
> 
> Looking more into this the problem is not reproducible with 6.19 with
> full preempt, and can be reproduced with 7.0 with full preempt where
> preemtion is mandatory. It only so happens that 1.19 was last kernel
> where preemption was not mandatory and where this problem also happened
> to not manifest with the Tumbleweed config.
> 

I am bit confused. Is it different problem than the kprobe/bpf issue
which was observed on host?

Is this kvm/guest specific problem you are observing? Is it still seen
even with that kprobe/bpf fix?

> It is not deterministic enough to bisect. It's probably some kind of
> race condition or code generation problem that can surface and go away
> or become too difficult to reproduce with seemingly unrelated changes.
> 

Can you try latest kernel?

There was change recently to move to use
generic transfer to guest mode work. i.e VIRT_XFER_TO_GUEST_WORK.

> Crash happens very early but it's eemu so the memory can be saved
> regardelss. The problem is that the crash tool refuses to load the core
> bacause it detects the core as SMP and vmlinux as non-SMP or the other
> way around.
> 
> Thanks
> 
> Michal
>