Re: [PATCH v1 3/5] powerpc/pseries: Fix stolen time accounting when dtl debugfs is used

["Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com>]

Paul Mackerras wrote:
> On Fri, Oct 26, 2018 at 01:55:44AM +0530, Naveen N. Rao wrote:
>> When the dtl debugfs interface is used, we usually set the
>> dtl_enable_mask to 0x7 (DTL_LOG_ALL). When this happens, we start seeing
>> DTL entries for all preempt reasons, including CEDE. In
>> scan_dispatch_log(), we add up the times from all entries and account
>> those towards stolen time. However, we should only be accounting stolen
>> time when the preemption was due to HDEC at the end of our time slice.
> 
> It's always been the case that stolen time when idle has been
> accounted as idle time, not stolen time.  That's why we didn't check
> for this in the past.
> 
> Do you have a test that shows different results (as in reported idle
> and stolen times) with this patch compared to without?

Ah ok, that makes sense now and explains why I couldn't observe much of 
a difference in practice. However, I also went by the fact that there 
are 7 other preemption reasons, which could impact our calculation.  
Looking at the list again, it looks like H_CONFER/H_PROD and some faults 
can also have an impact here, though they may be rare?

Thanks,
Naveen