Re: [PATCH] powerpc/vtime: Initialize starttime at boot for native accounting

["Christophe Leroy (CS GROUP)" <chleroy@kernel.org>]

Le 04/06/2026 à 16:51, Christophe Leroy (CS GROUP) a écrit :
> 
> 
> Le 04/06/2026 à 15:24, Shrikanth Hegde a écrit :
>> It was observed that /proc/stat had very large value for one ore more
>> CPUs. It was more visible after recent code simplifications around
>> cpustats.
>>
>> System has 240 CPUs.
>>
>> cat /proc/uptime;
>> 194.18 46500.55
>> cat /proc/stat
>> cpu  5966 39 837032887 4650070 164 185 100 0 0 0
>> cpu0 108 0 837030890 19109 24 4 23 0 0 0
>>
>> Since uptime is 194s, system time of each CPU can't be more than 19400.
>> Sum of system time  of all CPUs can't be more than 19400*240 4656000.
>> In fact huge value is close to mftb(). Note mftb doesn't reset on powerVM
>> when the LPAR restart. It only resets when whole system resets. The same
>> issue exists for kexec too.
>>
>> This happens since starttime is not setup at init time. Once it is set
>> then subsequent vtime_delta will return the right delta.
>>
>> Fix it by initializing the starttime during CPU initialization. This
>> fixes the large times seen.
>>
>> cat /proc/uptime; cat /proc/stat
>> 15.78 3694.63
>> cpu  6035 35 1347 369479 23 144 49 0 0 0
>> cpu0 19 0 38 1508 0 1 14 0 0 0
>>
>> Now, system time is reported as expected.
>>
>> Suggested-by: Christophe Leroy (CS GROUP) <chleroy@kernel.org>
>> Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
> 
> Reviewed-by: Christophe Leroy (CS GROUP) <chleroy@kernel.org>
> 
>> ---
>>
>> Christophe, I have taken the patch as is from the discussion we had.
>> Let me know if i should send it with your signed-off-by tag. I have just
>> written the changelog. I sent it like this since tag was not there.
> 
> Suggested-by is fine for me.
> 
>>
>> discussion thread:
>> https://eur01.safelinks.protection.outlook.com/? 
>> url=https%3A%2F%2Flore.kernel.org%2Fall%2Fcd10be19- 
>> e0bc-4e0c-8dac-4f1c05d0de8f%40kernel.org%2F&data=05%7C02%7Cchristophe.leroy%40csgroup.eu%7C1702c7a5ff63417da4ef08dec248be1d%7C8b87af7d86474dc78df45f69a2011bb5%7C0%7C0%7C639161814851391682%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=DjLhS3imHryT613JPVG0mnkWjILSGGrt1tLvmqFHtVM%3D&reserved=0
>>
>> Also, does this warrant Fixes tag? I found these two likely candidates.
>> Likely this issues exists since beginning.
>> c223c90386bc powerpc32: provide VIRT_CPU_ACCOUNTING
> 
> You say system has 240 CPU so I suppose this is not ppc32. That commit 
> wsa not supposed to change anything for ppc64, did you identify anything 
> special in that commit related to ppc64 ?
> 
>> b38a181c11d0 powerpc/time: isolate scaled cputime accounting in 
>> dedicated functions.
> 
> This one is also pure code re-organisation, unless you've been able to 
> spot a particular issue ?

Maybe commit cf9efce0ce31 ("powerpc: Account time using timebase rather 
than PURR")

It removed snapshot_timebases() and I can't see anything to replace it.


> 
>>
>>   arch/powerpc/kernel/time.c | 4 +++-
>>   1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/powerpc/kernel/time.c b/arch/powerpc/kernel/time.c
>> index 3460d1a5a97c..11145c40183d 100644
>> --- a/arch/powerpc/kernel/time.c
>> +++ b/arch/powerpc/kernel/time.c
>> @@ -377,7 +377,6 @@ void vtime_task_switch(struct task_struct *prev)
>>       }
>>   }
>> -#ifdef CONFIG_NO_HZ_COMMON
>>   /**
>>    * vtime_reset - Fast forward vtime entry clocks
>>    *
>> @@ -394,6 +393,7 @@ void vtime_reset(void)
>>   #endif
>>   }
>> +#ifdef CONFIG_NO_HZ_COMMON
>>   /**
>>    * vtime_dyntick_start - Inform vtime about entry to idle-dynticks
>>    *
>> @@ -933,6 +933,7 @@ static void __init set_decrementer_max(void)
>>   static void __init init_decrementer_clockevent(void)
>>   {
>>       register_decrementer_clockevent(smp_processor_id());
>> +    vtime_reset();
>>   }
>>   void secondary_cpu_time_init(void)
>> @@ -948,6 +949,7 @@ void secondary_cpu_time_init(void)
>>       /* FIME: Should make unrelated change to move snapshot_timebase
>>        * call here ! */
>>       register_decrementer_clockevent(smp_processor_id());
>> +    vtime_reset();
>>   }
>>   /*
>