Re: [linux-pm] [PATCH v2 0/8] RFC: CPU frequency min/max as PM QoS params

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Thursday, January 19, 2012, mark gross wrote:
> On Thu, Jan 19, 2012 at 12:24:26AM +0100, Rafael J. Wysocki wrote:
> > On Wednesday, January 18, 2012, mark gross wrote:
> > > On Mon, Jan 16, 2012 at 10:38:57PM +0100, Rafael J. Wysocki wrote:
> > > > Hi,
> > > > 
> > > > On Monday, January 16, 2012, Antti P Miettinen wrote:
> > > > > [did not reach linux-pm as I sent to wrong address, sorry for
> > > > > duplicates]
> > > > > 
> > > > > The inspiration for this patch series is the N9 CPU frequency boost
> > > > > upon input events:
> > > > > 
> > > > > http://www.spinics.net/lists/cpufreq/msg00667.html
> > > > > 
> > > > > and the related changes in git://codeaurora.org/kernel/msm.git tree.
> > > > > Those patches modify the ondemand cpufreq governor. This patch series
> > > > > adds minimum and maximum CPU frequency as PM QoS parameters and
> > > > > modifies the cpufreq core to enforce the PM QoS limits.
> > > > 
> > > > If that hasn't been clear enough so far, I'm still not convinced that using
> > > > PM QoS for that is a good idea.
> > > > 
> > > > First off, frequency as a unit of throughput is questionable to say the least,
> > > > because it isn't portable from one system to another.  Moreover, even on a
> > > > given system it isn't particularly clear what the exact correspondence
> > > > between frequency and throughput actually is.
> > > 
> > > You are right.  The notion of throughput of a CPU is really hard to
> > > quantify.  Perhaps not using the term "throughput" would help?
> > 
> > Yes, it would.
> > 
> > > The base issue I see, the Intel platform, is needing is that sometimes
> > > we need to block the lowest P-states that the ondemand governor goes for
> > > because those P-states result in media / graphics workloads dropping
> > > frames.  However; GPU intensive workloads do not stress the CPU so the
> > > ondemand governor goes for the low p-state.
> > > 
> > > I could use some way of constraining the PM-throttling of the
> > > cpu-freq that can be hit from kernel or user mode.  So the graphics
> > > driver can dynamically adjust the constraint request on the cpufreq
> > > subsystem.
> > > 
> > > It is problematic that any driver requesting a given frequency request
> > > is not portable across ISA's or even processor families in the same ISA.
> > > But, maybe such a driver should use a module parameter to work around
> > > this lack of portability?
> > 
> > Well, it seems to me that we're trying to add a backdoor to the (apparently
> > inadequate) governors here.  Arguably, the governors should be able to
> > make the right decisions on the basis of the information they receive
> > through their own interfaces.
> 
> the failings of governors to have the information needed is why pm_qos
> was created in the first place.

Well, that's interesting. :-)

> It can be seen as a limitation on the governor from some perspectives.  But,
> I like to think of if as updating existing governors to account for new use
> case requirements as hardware get bigger power management / performance
> dynamic ranges.

The current patchset doesn't seem to update governors, though.

> > > > Second, it's not particularly clear what the meaning of the "min" frequency
> > > > is supposed to be in terms of throughput.
> > > 
> > > It should mean "please cpufreq do not put the cpu into a state where its
> > > clock runs slower than min".  I don't think we should talk about it as
> > > throughput because thats not what the cpufreq controls.
> > 
> > Perhaps we need a new cpufreq governor that would take use PM QoS internally
> > to store requests from different sources, but that would work on a per-CPU
> > basis (not globally) and would provide a new interface for user space?
> > 
> 
> I don' think we need a new cpufreq governor, the parts of this patchset
> that I agree with evolve the governor to account for pm-qos requests
> but, globally for all cpu's.

As I said, it doesn't really evolve governors.  It adds a mechanism for
influencing policy limits in a kind of convoluted fashion.

I mean, there are scaling_min_freq and scaling_max_freq in the per-CPU
cpufreq's sysfs interface that can be used to set min/max policy
limits.  The only problem with those I see, which the current patchset
is kind of trying to address, is that they don't contain information
about who requested those limits.

> Hmm, your right this patch set is global in its request and not
> "per-cpu".  I need to think on that.  Making it per-cpu would likely
> infer we need to make the qos request per cpu as well.  
> 
> Do you think it needs to be per-cpu?  (I'm starting to think "yes" it
> does)
> 
> How do we scale the pm_qos ABI to support per/cpu?  (maybe we don't
> export those types of qos classes to the user mode?)

I think we should focus on iproving the existing cpufreq inteface in the
first place.

Thanks,
Rafael