Re: [PATCH 5/5] powercap/rapl: change floor frequency for vallewview

[Jacob Pan <jacob.jun.pan@linux.intel.com>]

On Tue, 29 Apr 2014 14:40:37 +0000
"R, Durgadoss" <durgadoss.r@intel.com> wrote:

> > -----Original Message-----
> > From: Jacob Pan [mailto:jacob.jun.pan@linux.intel.com]
> > Sent: Tuesday, April 29, 2014 6:33 PM
> > To: R, Durgadoss
> > Cc: Linux PM; Wysocki, Rafael J; LKML; David E. Box; Alan Cox;
> > Accardi, Kristen C Subject: Re: [PATCH 5/5] powercap/rapl: change
> > floor frequency for vallewview
> > 
> > On Tue, 29 Apr 2014 02:45:22 +0000
> > "R, Durgadoss" <durgadoss.r@intel.com> wrote:
> > 
> > > Hi Jacob,
> > >
> > > > -----Original Message-----
> > > > From: Jacob Pan [mailto:jacob.jun.pan@linux.intel.com]
> > > > Sent: Monday, April 28, 2014 7:35 PM
> > > > To: Linux PM; Wysocki, Rafael J; LKML
> > > > Cc: David E. Box; Alan Cox; R, Durgadoss; Accardi, Kristen C;
> > > > Jacob Pan Subject: [PATCH 5/5] powercap/rapl: change floor
> > > > frequency for vallewview
> > > >
> > > > RAPL power limit reduce power by limiting CPU P-state and
> > > > other techniques. On Valleyview, RAPL power limit cannot
> > > > go to LFM (low frequency mode) if we don't set the floor
> > > > frequency via IOSF mailbox.
> > > >
> > > > This patch enables setting of floor frquency such that
> > > > RAPL power limit is more effective.
> > > >
> > > > Signed-off-by: Jacob Pan <jacob.jun.pan@linux.intel.com>
> > > > ---
> > > >  drivers/powercap/intel_rapl.c | 27 +++++++++++++++++++--------
> > > >  1 file changed, 19 insertions(+), 8 deletions(-)
> > > >
> > > > diff --git a/drivers/powercap/intel_rapl.c
> > > > b/drivers/powercap/intel_rapl.c index b1cda6f..13e4776 100644
> > > > --- a/drivers/powercap/intel_rapl.c
> > > > +++ b/drivers/powercap/intel_rapl.c
> > > > @@ -32,6 +32,7 @@
> > > >
> > > >  #include <asm/processor.h>
> > > >  #include <asm/cpu_device_id.h>
> > > > +#include <asm/iosf_mbi.h>
> > > >
> > > >  /* bitmasks for RAPL MSRs, used by primitive access functions
> > > > */ #define ENERGY_STATUS_MASK      0xffffffff
> > > > @@ -336,11 +337,17 @@ static int find_nr_power_limit(struct
> > > > rapl_domain *rd) return i;
> > > >  }
> > > >
> > > > +#define VLV_CPU_POWER_BUDGET_CTL (0x2)
> > > > +static const struct x86_cpu_id valleyview_id[] = {
> > > > +	{ X86_VENDOR_INTEL, 6, 0x37},
> > > > +	{}
> > > > +};
> > >
> > > There are other platforms that have this FloorFreq register as
> > > well. And those addresses are not '0x02'. So, we need to have a
> > > cpu_id based table to define the address of the floor freq
> > > register as well. [This is not specific to valleyview.]
> > >
> > Sounds like I need to add an abstraction to capture this. So far,
> > there are only two exceptions so i was hesitate to do so. Thanks
> > for the input.
> 
> Yes, We at least have few platforms that need this.
> 
> > > Also, is there a plan to expose this floor freq ratio through
> > > Sysfs for runtime configuration. ? May be through a standard
> > > thermal cooling device interface ?
> > >
> > why would that be necessary? who will use it? floor freq only
> > affects RAPL, AFAIK. In Linux there is no guaranteed freq anyway.
> > My original patch to enable RAPL as cooling device was abandoned in
> > favor of powercap framework, I am not sure if we should go back.
> 
> There are user space thermal controls which change RAPL Power limits
> according to platform's thermal condition as you might be aware.
> 
> The floor frequency is not used only to transition to LFM ratio. We
> can transition to any frequency ratio by adjusting this floor
> frequency (at least on VLV and couple more platforms)
> 
> Hence while changing RAPL Power Limits, there is a need to adjust
> this also, to specify which ratio is our Floor (basically we will not
> go below that). That's why we need an interface for modifying this
> at run time (along with Power Limits).
> 
I understand. What I am proposing here is to have a single knob for
user control power, instead of two knobs (power limit and floor freq)
which may have conflicts. When thermal throttling is needed, user only
cares about power limit, that is why I think it is better to set floor
to LFM and let power limit be the only knob. It is simpler. In case
freq is a constraint, user should use cpufreq interface.

> Thanks,
> Durga
> 
> > > > +
> > > >  static int set_domain_enable(struct powercap_zone *power_zone,
> > > > bool mode) {
> > > >  	struct rapl_domain *rd =
> > > > power_zone_to_rapl_domain(power_zone); int nr_powerlimit;
> > > > -
> > > > +	u32 mdata = 0;
> > > >  	if (rd->state & DOMAIN_STATE_BIOS_LOCKED)
> > > >  		return -EACCES;
> > > >  	get_online_cpus();
> > > > @@ -350,7 +357,16 @@ static int set_domain_enable(struct
> > > > powercap_zone *power_zone, bool mode)
> > > >  	/* always enable clamp such that p-state can go below
> > > > OS requested
> > > >  	 * range. power capping priority over guranteed
> > > > frequency. */
> > > > -	rapl_write_data_raw(rd, PL1_CLAMP, mode);
> > > > +	if (x86_match_cpu(valleyview_id)) {
> > > > +		iosf_mbi_read(BT_MBI_UNIT_PMC, BT_MBI_PMC_READ,
> > > > +			VLV_CPU_POWER_BUDGET_CTL, &mdata);
> > > > +		mdata &= ~(0x7f << 8);
> > > > +		mdata |= 1 << 8;
> > > > +		iosf_mbi_write(BT_MBI_UNIT_PMC,
> > > > BT_MBI_PMC_WRITE,
> > > > +			VLV_CPU_POWER_BUDGET_CTL, mdata);
> > > > +	} else
> > > > +		rapl_write_data_raw(rd, PL1_CLAMP, mode);
> > > > +
> > > >  	/* some domains have pl2 */
> > > >  	if (nr_powerlimit > 1) {
> > > >  		rapl_write_data_raw(rd, PL2_ENABLE, mode);
> > > > @@ -833,11 +849,6 @@ static int rapl_write_data_raw(struct
> > > > rapl_domain *rd, return 0;
> > > >  }
> > > >
> > > > -static const struct x86_cpu_id energy_unit_quirk_ids[] = {
> > > > -	{ X86_VENDOR_INTEL, 6, 0x37},/* Valleyview */
> > > > -	{}
> > > > -};
> > >
> > > Same thing here. There are other Atom platforms that need this
> > > conversion quirk. So, please keep the table as is.
> > >
> > > Thanks,
> > > Durga
> > >
> > > > -
> > > >  static int rapl_check_unit(struct rapl_package *rp, int cpu)
> > > >  {
> > > >  	u64 msr_val;
> > > > @@ -859,7 +870,7 @@ static int rapl_check_unit(struct
> > > > rapl_package *rp, int cpu) */
> > > >  	value = (msr_val & ENERGY_UNIT_MASK) >>
> > > > ENERGY_UNIT_OFFSET; /* some CPUs have different way to
> > > > calculate energy unit */
> > > > -	if (x86_match_cpu(energy_unit_quirk_ids))
> > > > +	if (x86_match_cpu(valleyview_id))
> > > >  		rp->energy_unit_divisor = 1000000 / (1 <<
> > > > value); else
> > > >  		rp->energy_unit_divisor = 1 << value;
> > > > --
> > > > 1.8.1.2
> > >
> > 
> > [Jacob Pan]

[Jacob Pan]