Re: [REPORT] Softlockups on PowerNV with upstream

[Gavin Shan <gshan@redhat.com>]

On 4/10/25 10:22 PM, Aditya Gupta wrote:
> Cc +donet
> 
> On 25/04/10 07:44PM, Gavin Shan wrote:
>> On 4/10/25 6:23 PM, Oscar Salvador wrote:
>>> On Thu, Apr 10, 2025 at 03:35:19PM +1000, Gavin Shan wrote:
>>>> Thanks, Oscar. You're correct that the overhead is introduced by for_each_present_section_nr().
>>>> I already had the fix, working on IBM's Power9 machine, where the issue can be
>>>> reproduced. Please see the attached patch.
>>>>
>>>> I'm having most tests on ARM64 machine for the fix.
>>>
>>> Looks good to me.
>>> But we need a comment explaining why block_id is set to ULONG_MAX
>>> at the beginning as this might not be obvious.
>>>
>>> Also, do we need
>>>    if (block_id != ULONG_MAX && memory_block_id(nr) == block_id) ?
>>>
>>> Cannot just be
>>>
>>>    if (memory_block_id(nr) == block_id) ?
>>>
>>> AFAICS, the first time we loop through 'memory_block_id(nr) == ULONG_MAX'
>>> will evaluate false and and we will set block_id afterwards.
>>>
>>> Either way looks fine to me.
>>> Another way I guess would be:
>>>
>>
>> Yeah, we need to record the last handled block ID by @block_id. For the
>> first time to register the block memory device in the loop, @block_id needs
>> to be invalid (ULONG_MAX), bypassing the check of 'memory_block_id(nr) == block_id'.
>> I will post the fix for review after Aditya confirms it works for him, with extra
>> comment to explain why @block_id is initialized to ULONG_MAX.
>>
>> Aditya, please have a try when you get a chance, thanks! I verified it on Power9
>> machine where the issue exists and on one of my ARM64 machine.
> 
> I don't see any softlockups now with your patch as well as Oscar's patch.
> 
> Tested on PowerNV Power10.
> 
> Thanks for the quick replies Gavin.

Nice, thanks for the quick test, Aditya. I will send the fix for reivew, with
you copied.

Thanks,
Gavin