Re: [PATCH 5/5] Package Level Thermal Control and Power Limit Notification: pkgtemp doc

[Guenter Roeck <guenter.roeck@ericsson.com>]

On Thu, Aug 19, 2010 at 04:51:20PM -0400, Fenghua Yu wrote:
> On Thu, Aug 19, 2010 at 09:27:19AM -0700, Guenter Roeck wrote:
> > On Thu, Aug 19, 2010 at 11:46:32AM -0400, Jean Delvare wrote:
> > > Hi Fenghua, Guenter,
> > > 
> > > Sorry for joining the discussion a little late, I was on vacation when
> > > it happened. I'll comment now, it's probably "too late" as the patch
> > > set was merged meanwhile, but still...
> > > 
> > There was no discussion at all, unfortunately.
> > 
> > > On Thu, 22 Jul 2010 14:21:11 -0700, Fenghua Yu wrote:
> > > > On Thu, Jul 22, 2010 at 11:58:14AM -0700, Guenter Roeck wrote:
> > > > > You use the argument that there may be other package level sensors in the future.
> > > > > Are there any plans for this, or is this just a theory ?
> > > > 
> > > > Not just a theory. Sandy Bridge already implements other package level sensors.
> > > > If really need to know exactly which sensors are implemented, we might go
> > > > through a channel before releasing the info.
> > > > 
> > > > > Next question is how to handle future sensor types. One hwmon instance per sensor,
> > > > > additional sensors in this driver, or even a new driver ?
> > > > 
> > > > Currently package level thermal just reports the maximum temperature across
> > > > the package. Which sensor is reporting the highest temperature is not exposed.
> > > 
> > > So this isn't a real physical sensor, but more of a meta-sensor? If
> > > this is a case, then we don't need support for this at all. User-space
> > > can compute a maximum by itself, we don't need a dedicated kernel
> > > driver for that.
> > > 
> The pkgtemp reports thermal status for a set of sensors in a package. Please
> note the sensors in a package are not limited to processor sensors which are
> handled by coretemp. The sensors in a package also include gfx sensors, cache
> sensors, memory controllor sensors which are not handled by any hwmon drivers.
> 
> OS gets maximum package thermal status only from MSR PACKAGE_THERM_STATUS.
> There is no detailed thermal info for each sensor in a package. User-space
> can't compute a maximum by itself. So a piece of kernel driver code, whether
> a seperate driver or a integrated driver, is necessary if user-space wants to
> know thermal status of a package.
> 
> > > > > We had was a separate discussion if the coretemp driver should be redesigned
> > > > > to one instance per CPU. The package sensor would fit into that model,
> > > > > since you would have
> > > > > 
> > > > > coretemp-isa-0000
> > > > > Core0
> > > > > Core1
> > > > > ...
> > > > > CoreN
> > > > > Package
> > > > > 
> > > > > coretemp-isa-0001
> > > > > Core0
> > > > > Core1
> > > > > ...
> > > > > CoreM
> > > > > Package
> > > > > 
> > > > > I personally would prefer that approach. It would avoid ambiguity associating Package X
> > > > > with specific cores, and it would also easily expand to additional non-core future sensors.
> > > 
> > > For the records, I totally support this approach. I want the coretemp
> > > driver to be updated to present a single hwmon device per CPU, no
> > > matter what happens to the "package temperature".
> > > 
> > I might spend some time rewriting the coretemp driver as described above,
> > unless someone else picks it up, and unless there is opposition. 
> > Obviously, that won't include the package sensor since there is now
> > a separate driver for it.
> I agree with this method too. On a multiple socket system, the current coretemp
> output will cause confusion since it only outputs core# without package#.
> 
> If it's ok for you, I can rewrite this part to have hwmon device per CPU with
> both core and package thermal info and send out RFC patch soon.
> 
Sounds good to me.

Guenter