AW: RE: RE: RE: RE: RE: No C-States any longer...(it's not over!)

AW: RE: RE: RE: RE: RE: No C-States any longer...(it's not over!)

[Carsten Schiers]

I thought by buying a new mainboard, I could solve the issue. It is not the case.

>> >In your case there's no _CST found, and then no PBLK found.
>...
>w/o PBLK, FADT info is incomplete to construct a full Cstate information.

The new mainboard now implements a PBLK for the first CPU in FADT, but latencies for C2/C3 
are > 100/1000, which is the semantic to disable it (I read). 

But no change in xenpm output, there are still no C-States. 

I think acpi_processor_get_power_info_default is called too late. In old, working 2.6.18 code, 
a similar function acpi_processor_get_power_info_default_c1 is called as very first function in 
acpi_processor_get_power_info, to setup C1 which is mandatory for all CPUs.

Some other thoughts where I see the code needs changes:

  - even if spec says that either PBLK or _CST should implement, C1 should be set up (solves issue with mainboard 1)
  - when C2/C3 are above limits (100/1000), C1 still should be set up (solves issue with mainboard 2)
  - Spec allows PBLK for one CPU only (definition as of mainboard 2, I don't think code is respecting this case).

BTW:

  - I have C1E support in CPU and BIOS, but see no difference in DSDT and FADT when on/off in BIOS
  - Just curious: how many C-States should AMD X3 400e normally support? Why is my BIOS switching off C2/C3?

Carsten.

RE: RE: RE: RE: RE: RE: No C-States any longer...(it's not over!)

[Tian, Kevin]

> From: Carsten Schiers [mailto:carsten@schiers.de]
> Sent: Monday, June 27, 2011 7:13 PM
> 
> I thought by buying a new mainboard, I could solve the issue. It is not the case.
> 
> >> >In your case there's no _CST found, and then no PBLK found.
> >...
> >w/o PBLK, FADT info is incomplete to construct a full Cstate information.
> 
> The new mainboard now implements a PBLK for the first CPU in FADT, but
> latencies for C2/C3
> are > 100/1000, which is the semantic to disable it (I read).
> 
> But no change in xenpm output, there are still no C-States.
> 
> I think acpi_processor_get_power_info_default is called too late. In old,
> working 2.6.18 code,
> a similar function acpi_processor_get_power_info_default_c1 is called as very
> first function in
> acpi_processor_get_power_info, to setup C1 which is mandatory for all CPUs.

this is the generic Linux ACPI logic, so you may report to Linux upstream for their
opinion.

> 
> Some other thoughts where I see the code needs changes:
> 
>   - even if spec says that either PBLK or _CST should implement, C1 should be
> set up (solves issue with mainboard 1)
>   - when C2/C3 are above limits (100/1000), C1 still should be set up (solves
> issue with mainboard 2)

yes, I agree that C1 should be always present regardless of platform difference,
as long as cpuidle is enabled in Xen. I'll try to work out a patch and CC you when
it's ready.

Thanks
Kevin