Re: [PATCH] intel_idle: Adjust the SKX C6 latency and residency if PC6 is disabled

[Chen Yu]

On Thu, May 27, 2021 at 11:25:18AM +0300, Artem Bityutskiy wrote:
> On Thu, 2021-05-27 at 12:56 +0800, Chen Yu wrote:
> 
> ... snip ...
> 
> > Exit latency:
> > The C6 exit latency is measured when woken up from CC6/PC6. In the past,
> > if PC6 is disabled, CPU would be demoted to CC6/PC3, which is close to
> > the latency from CC6/PC6 and there is no need to update the C6 exit latency.
> > However on newer platform there is no CC3/PC3 anymore, then the C6 exit
> > latency with PC6 disabled should be CC6/PC0.
> > 
> > Target residency:
> > With PC6 disabled and C3/PC3 supported, the OS requests C3 if idle
> > duration is within [CC6, PC6) target_residency. On new CPU generations
> > with C3/PC3 deprecated, the OS would request C1E. This would cause
> > low energy-efficiency. In summary, the question is, should we lower
> > the bar to request C6 when PC6 is disabled? The answer is yes.
> ... snip ...
> 
> Hi Yu,
>
Hi Artem, 
> Thanks for this patch, it is very actual and helpful.
> 
> Comments about the commit message below.
> 
> This patch is specifically about SKX. It also covers CLX and CPX,
> because they have the same ID.
> 
> Now, this platforms do not have C3 and PC3. So I would avoid talking
> about these states
> in the commit message. Why making a simple thing more complex?
> 
I agree, in theory this issue is irrelevant of the C3/PC3, however I was
thinking the issue would become more significant to user on platform without
C3/PC3, and this increase the necessity of this patch. But let me
refine the commit log to only mention C3 a little just for future reference.
> Here are all the SKX C-states.
> 
> 1. Linux-level C-states (linux can ask for): C1, C1E, C6.
> 2. HW-level C-states (HW supports under the hood): C1, C1E, CC6, PC2,
> PC6.
> 
And PC0?
> Here is the story of this patch in my understanding.
> 
> 1. C6 maps to CC6 and PC6.
> 2. CC6 is "shallower" than PC6.
> 3. Linux assumes worst case - PC6.
> 4. Many datacenters and users disable PC6.
> 5. We can optimize intel_idle in this case: adjust C6 latency and
> target residency to match (faster) CC6.
> 
> That's it.
> 
Okay.
> Then may be it is worth mentioning that CC6 vs PC2 difference is not
> really measurable, so
> the adjustment is only for PC6.
> 
Okay.

thanks,
Chenyu
> Artem.
>