[PATCH] processor passthru - upload _Cx and _Pxx data to hypervisor (v5).

[PATCH] processor passthru - upload _Cx and _Pxx data to hypervisor (v5).

[Konrad Rzeszutek Wilk]

The problem these three patches try to solve is to provide ACPI power management
information to the Xen hypervisor. The hypervisor lacks the ACPI DSDT parser so it can't
get that data without some help - and the initial domain can provide that. One
approach (https://lkml.org/lkml/2011/11/30/245) augments the ACPI code to call
an external PM code - but there were no comments about it so I decided to see
if another approach could solve it.

This module (processor-passthru)  collects the information that the cpufreq
drivers and the ACPI processor code save in the 'struct acpi_processor' and
then uploads it to the hypervisor.

The driver can be either an module or compiled in. In either mode the driver
launches a thread that checks whether an cpufreq driver is registered. If so
it reads all the 'struct acpi_processor' data for all online CPUs and sends
it to hypervisor. The driver also register a CPU hotplug component - so if a new
CPU shows up - it would send the data to the hypervisor for it as well.
Furthermore it also verifies whether the ACPI ID count is different than what
the kernel sees (which is possible with dom0_max_vcpus) and if so uploads
the data for the other ACPI IDs.

The patches are available in this git tree:

git://git.kernel.org/pub/scm/linux/kernel/git/konrad/xen.git stable/processor-passthru.v5


Konrad Rzeszutek Wilk (3):
      xen/setup/pm/acpi: Remove the call to boot_option_idle_override.
      xen/enlighten: Expose MWAIT and MWAIT_LEAF if hypervisor OKs it.
      xen/processor-passthru: Provide an driver that passes struct acpi_processor data to the hypervisor.

 arch/x86/xen/enlighten.c         |   92 +++++++-
 arch/x86/xen/setup.c             |    1 -
 drivers/xen/Kconfig              |   14 +
 drivers/xen/Makefile             |    2 +-
 drivers/xen/processor-passthru.c |  485 ++++++++++++++++++++++++++++++++++++++
 include/xen/interface/platform.h |    4 +-
 6 files changed, 594 insertions(+), 4 deletions(-)


P.S.
On the hypervisor side, it requires this patch on AMD:
# HG changeset patch
# Parent aea8cfac8cf1afe397f2e1d422a852008d8a83fe
traps: AMD PM RDMSRs (MSR_K8_PSTATE_CTRL, etc)

The restriction to read and write the AMD power management MSRs is gated if the
domain 0 is the PM domain (so FREQCTL_dom0_kernel is set). But we can
relax this restriction and allow the privileged domain to read the MSRs
(but not write). This allows the priviliged domain to harvest the power
management information (ACPI _PSS states) and send it to the hypervisor.

This patch works fine with older classic dom0 (2.6.32) and with
AMD K7 and K8 boxes.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
diff -r aea8cfac8cf1 xen/arch/x86/traps.c
--- a/xen/arch/x86/traps.c	Thu Feb 23 13:23:02 2012 -0500
+++ b/xen/arch/x86/traps.c	Thu Feb 23 13:29:00 2012 -0500
@@ -2484,7 +2484,7 @@ static int emulate_privileged_op(struct 
         case MSR_K8_PSTATE7:
             if ( boot_cpu_data.x86_vendor != X86_VENDOR_AMD )
                 goto fail;
-            if ( !is_cpufreq_controller(v->domain) )
+            if ( !is_cpufreq_controller(v->domain) && !IS_PRIV(v->domain) )
             {
                 regs->eax = regs->edx = 0;
                 break;

[PATCH 1/3] xen/setup/pm/acpi: Remove the call to boot_option_idle_override.

[Konrad Rzeszutek Wilk]

We needed that call in the past to force the kernel to use
default_idle (which called safe_halt, which called xen_safe_halt).

But set_pm_idle_to_default() does now that, so there is no need
to use this boot option operand.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/xen/setup.c |    1 -
 1 files changed, 0 insertions(+), 1 deletions(-)

diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index e03c636..1236623 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -420,7 +420,6 @@ void __init xen_arch_setup(void)
 	boot_cpu_data.hlt_works_ok = 1;
 #endif
 	disable_cpuidle();
-	boot_option_idle_override = IDLE_HALT;
 	WARN_ON(set_pm_idle_to_default());
 	fiddle_vdso();
 }
-- 
1.7.9.48.g85da4d

[PATCH 2/3] xen/enlighten: Expose MWAIT and MWAIT_LEAF if hypervisor OKs it.

[Konrad Rzeszutek Wilk]

For the hypervisor to take advantage of the MWAIT support it needs
to extract from the ACPI _CST the register address. But the
hypervisor does not have the support to parse DSDT so it relies on
the initial domain (dom0) to parse the ACPI Power Management information
and push it up to the hypervisor. The pushing of the data is done
by the processor_harveset_xen module which parses the information that
the ACPI parser has graciously exposed in 'struct acpi_processor'.

For the ACPI parser to also expose the Cx states for MWAIT, we need
to expose the MWAIT capability (leaf 1). Furthermore we also need to
expose the MWAIT_LEAF capability (leaf 5) for cstate.c to properly
function.

The hypervisor could expose these flags when it traps the XEN_EMULATE_PREFIX
operations, but it can't do it since it needs to be backwards compatible.
Instead we choose to use the native CPUID to figure out if the MWAIT
capability exists and use the XEN_SET_PDC query hypercall to figure out
if the hypervisor wants us to expose the MWAIT_LEAF capability or not.

Note: The XEN_SET_PDC query was implemented in c/s 23783:
"ACPI: add _PDC input override mechanism".

With this in place, instead of
 C3 ACPI IOPORT 415
we get now
 C3:ACPI FFH INTEL MWAIT 0x20

Note: The cpu_idle which would be calling the mwait variants for idling
never gets set b/c we set the default pm_idle to be the hypercall variant.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/xen/enlighten.c         |   92 +++++++++++++++++++++++++++++++++++++-
 include/xen/interface/platform.h |    4 +-
 2 files changed, 94 insertions(+), 2 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 12eb07b..4c82936 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -62,6 +62,14 @@
 #include <asm/reboot.h>
 #include <asm/stackprotector.h>
 #include <asm/hypervisor.h>
+#include <asm/mwait.h>
+
+#ifdef CONFIG_ACPI
+#include <asm/acpi.h>
+#include <acpi/pdc_intel.h>
+#include <acpi/processor.h>
+#include <xen/interface/platform.h>
+#endif
 
 #include "xen-ops.h"
 #include "mmu.h"
@@ -200,13 +208,17 @@ static void __init xen_banner(void)
 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
 
+static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
+static __read_mostly unsigned int cpuid_leaf5_ecx_val;
+static __read_mostly unsigned int cpuid_leaf5_edx_val;
+
 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
 		      unsigned int *cx, unsigned int *dx)
 {
 	unsigned maskebx = ~0;
 	unsigned maskecx = ~0;
 	unsigned maskedx = ~0;
-
+	unsigned setecx = 0;
 	/*
 	 * Mask out inconvenient features, to try and disable as many
 	 * unsupported kernel subsystems as possible.
@@ -214,9 +226,18 @@ static void xen_cpuid(unsigned int *ax, unsigned int *bx,
 	switch (*ax) {
 	case 1:
 		maskecx = cpuid_leaf1_ecx_mask;
+		setecx = cpuid_leaf1_ecx_set_mask;
 		maskedx = cpuid_leaf1_edx_mask;
 		break;
 
+	case CPUID_MWAIT_LEAF:
+		/* Synthesize the values.. */
+		*ax = 0;
+		*bx = 0;
+		*cx = cpuid_leaf5_ecx_val;
+		*dx = cpuid_leaf5_edx_val;
+		return;
+
 	case 0xb:
 		/* Suppress extended topology stuff */
 		maskebx = 0;
@@ -232,9 +253,75 @@ static void xen_cpuid(unsigned int *ax, unsigned int *bx,
 
 	*bx &= maskebx;
 	*cx &= maskecx;
+	*cx |= setecx;
 	*dx &= maskedx;
+
 }
 
+static bool __init xen_check_mwait(void)
+{
+#if CONFIG_ACPI
+	struct xen_platform_op op = {
+		.cmd			= XENPF_set_processor_pminfo,
+		.u.set_pminfo.id	= -1,
+		.u.set_pminfo.type	= XEN_PM_PDC,
+	};
+	uint32_t buf[3];
+	unsigned int ax, bx, cx, dx;
+	unsigned int mwait_mask;
+
+	/* We need to determine whether it is OK to expose the MWAIT
+	 * capability to the kernel to harvest deeper than C3 states from ACPI
+	 * _CST using the processor_harvest_xen.c module. For this to work, we
+	 * need to gather the MWAIT_LEAF values (which the cstate.c code
+	 * checks against). The hypervisor won't expose the MWAIT flag because
+	 * it would break backwards compatibility; so we will find out directly
+	 * from the hardware and hypercall.
+	 */
+	if (!xen_initial_domain())
+		return false;
+
+	ax = 1;
+	cx = 0;
+
+	native_cpuid(&ax, &bx, &cx, &dx);
+
+	mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
+		     (1 << (X86_FEATURE_MWAIT % 32));
+
+	if ((cx & mwait_mask) != mwait_mask)
+		return false;
+
+	/* We need to emulate the MWAIT_LEAF and for that we need both
+	 * ecx and edx. The hypercall provides only partial information.
+	 */
+
+	ax = CPUID_MWAIT_LEAF;
+	bx = 0;
+	cx = 0;
+	dx = 0;
+
+	native_cpuid(&ax, &bx, &cx, &dx);
+
+	/* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
+	 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
+	 */
+	buf[0] = ACPI_PDC_REVISION_ID;
+	buf[1] = 1;
+	buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
+
+	set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
+
+	if ((HYPERVISOR_dom0_op(&op) == 0) &&
+	    (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
+		cpuid_leaf5_ecx_val = cx;
+		cpuid_leaf5_edx_val = dx;
+	}
+	return true;
+#else
+	return false;
+#endif
+}
 static void __init xen_init_cpuid_mask(void)
 {
 	unsigned int ax, bx, cx, dx;
@@ -261,6 +348,9 @@ static void __init xen_init_cpuid_mask(void)
 	/* Xen will set CR4.OSXSAVE if supported and not disabled by force */
 	if ((cx & xsave_mask) != xsave_mask)
 		cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
+
+	if (xen_check_mwait())
+		cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
 }
 
 static void xen_set_debugreg(int reg, unsigned long val)
diff --git a/include/xen/interface/platform.h b/include/xen/interface/platform.h
index c168468..6220b98 100644
--- a/include/xen/interface/platform.h
+++ b/include/xen/interface/platform.h
@@ -200,7 +200,7 @@ DEFINE_GUEST_HANDLE_STRUCT(xenpf_getidletime_t);
 #define XEN_PM_CX   0
 #define XEN_PM_PX   1
 #define XEN_PM_TX   2
-
+#define XEN_PM_PDC  3
 /* Px sub info type */
 #define XEN_PX_PCT   1
 #define XEN_PX_PSS   2
@@ -286,6 +286,7 @@ struct xen_processor_performance {
 };
 DEFINE_GUEST_HANDLE_STRUCT(xen_processor_performance);
 
+DEFINE_GUEST_HANDLE(uint32_t);
 struct xenpf_set_processor_pminfo {
 	/* IN variables */
 	uint32_t id;    /* ACPI CPU ID */
@@ -293,6 +294,7 @@ struct xenpf_set_processor_pminfo {
 	union {
 		struct xen_processor_power          power;/* Cx: _CST/_CSD */
 		struct xen_processor_performance    perf; /* Px: _PPC/_PCT/_PSS/_PSD */
+		GUEST_HANDLE(uint32_t)              pdc;
 	};
 };
 DEFINE_GUEST_HANDLE_STRUCT(xenpf_set_processor_pminfo);
-- 
1.7.9.48.g85da4d

Re: [PATCH 2/3] xen/enlighten: Expose MWAIT and MWAIT_LEAF if hypervisor OKs it.

[Jan Beulich]

>>> On 23.02.12 at 23:31, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> For the hypervisor to take advantage of the MWAIT support it needs
> to extract from the ACPI _CST the register address. But the
> hypervisor does not have the support to parse DSDT so it relies on
> the initial domain (dom0) to parse the ACPI Power Management information
> and push it up to the hypervisor. The pushing of the data is done
> by the processor_harveset_xen module which parses the information that
> the ACPI parser has graciously exposed in 'struct acpi_processor'.
> 
> For the ACPI parser to also expose the Cx states for MWAIT, we need
> to expose the MWAIT capability (leaf 1). Furthermore we also need to
> expose the MWAIT_LEAF capability (leaf 5) for cstate.c to properly
> function.
> 
> The hypervisor could expose these flags when it traps the XEN_EMULATE_PREFIX
> operations, but it can't do it since it needs to be backwards compatible.
> Instead we choose to use the native CPUID to figure out if the MWAIT
> capability exists and use the XEN_SET_PDC query hypercall to figure out
> if the hypervisor wants us to expose the MWAIT_LEAF capability or not.
> 
> Note: The XEN_SET_PDC query was implemented in c/s 23783:
> "ACPI: add _PDC input override mechanism".
> 
> With this in place, instead of
>  C3 ACPI IOPORT 415
> we get now
>  C3:ACPI FFH INTEL MWAIT 0x20
> 
> Note: The cpu_idle which would be calling the mwait variants for idling
> never gets set b/c we set the default pm_idle to be the hypercall variant.
> 
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> ---
>  arch/x86/xen/enlighten.c         |   92 
> +++++++++++++++++++++++++++++++++++++-
>  include/xen/interface/platform.h |    4 +-
>  2 files changed, 94 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
> index 12eb07b..4c82936 100644
> --- a/arch/x86/xen/enlighten.c
> +++ b/arch/x86/xen/enlighten.c
> @@ -62,6 +62,14 @@
>  #include <asm/reboot.h>
>  #include <asm/stackprotector.h>
>  #include <asm/hypervisor.h>
> +#include <asm/mwait.h>
> +
> +#ifdef CONFIG_ACPI
> +#include <asm/acpi.h>
> +#include <acpi/pdc_intel.h>
> +#include <acpi/processor.h>
> +#include <xen/interface/platform.h>
> +#endif
>  
>  #include "xen-ops.h"
>  #include "mmu.h"
> @@ -200,13 +208,17 @@ static void __init xen_banner(void)
>  static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0;
>  static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0;
>  
> +static __read_mostly unsigned int cpuid_leaf1_ecx_set_mask;
> +static __read_mostly unsigned int cpuid_leaf5_ecx_val;
> +static __read_mostly unsigned int cpuid_leaf5_edx_val;
> +
>  static void xen_cpuid(unsigned int *ax, unsigned int *bx,
>  		      unsigned int *cx, unsigned int *dx)
>  {
>  	unsigned maskebx = ~0;
>  	unsigned maskecx = ~0;
>  	unsigned maskedx = ~0;
> -
> +	unsigned setecx = 0;
>  	/*
>  	 * Mask out inconvenient features, to try and disable as many
>  	 * unsupported kernel subsystems as possible.
> @@ -214,9 +226,18 @@ static void xen_cpuid(unsigned int *ax, unsigned int 
> *bx,
>  	switch (*ax) {
>  	case 1:
>  		maskecx = cpuid_leaf1_ecx_mask;
> +		setecx = cpuid_leaf1_ecx_set_mask;
>  		maskedx = cpuid_leaf1_edx_mask;
>  		break;
>  
> +	case CPUID_MWAIT_LEAF:
> +		/* Synthesize the values.. */
> +		*ax = 0;
> +		*bx = 0;
> +		*cx = cpuid_leaf5_ecx_val;
> +		*dx = cpuid_leaf5_edx_val;
> +		return;
> +
>  	case 0xb:
>  		/* Suppress extended topology stuff */
>  		maskebx = 0;
> @@ -232,9 +253,75 @@ static void xen_cpuid(unsigned int *ax, unsigned int 
> *bx,
>  
>  	*bx &= maskebx;
>  	*cx &= maskecx;
> +	*cx |= setecx;
>  	*dx &= maskedx;
> +
>  }
>  
> +static bool __init xen_check_mwait(void)
> +{
> +#if CONFIG_ACPI

#ifdef

> +	struct xen_platform_op op = {
> +		.cmd			= XENPF_set_processor_pminfo,
> +		.u.set_pminfo.id	= -1,
> +		.u.set_pminfo.type	= XEN_PM_PDC,
> +	};
> +	uint32_t buf[3];
> +	unsigned int ax, bx, cx, dx;
> +	unsigned int mwait_mask;
> +
> +	/* We need to determine whether it is OK to expose the MWAIT
> +	 * capability to the kernel to harvest deeper than C3 states from ACPI
> +	 * _CST using the processor_harvest_xen.c module. For this to work, we
> +	 * need to gather the MWAIT_LEAF values (which the cstate.c code
> +	 * checks against). The hypervisor won't expose the MWAIT flag because
> +	 * it would break backwards compatibility; so we will find out directly
> +	 * from the hardware and hypercall.
> +	 */
> +	if (!xen_initial_domain())
> +		return false;
> +
> +	ax = 1;
> +	cx = 0;
> +
> +	native_cpuid(&ax, &bx, &cx, &dx);
> +
> +	mwait_mask = (1 << (X86_FEATURE_EST % 32)) |
> +		     (1 << (X86_FEATURE_MWAIT % 32));
> +
> +	if ((cx & mwait_mask) != mwait_mask)
> +		return false;
> +
> +	/* We need to emulate the MWAIT_LEAF and for that we need both
> +	 * ecx and edx. The hypercall provides only partial information.
> +	 */
> +
> +	ax = CPUID_MWAIT_LEAF;
> +	bx = 0;
> +	cx = 0;
> +	dx = 0;
> +
> +	native_cpuid(&ax, &bx, &cx, &dx);
> +
> +	/* Ask the Hypervisor whether to clear ACPI_PDC_C_C2C3_FFH. If so,
> +	 * don't expose MWAIT_LEAF and let ACPI pick the IOPORT version of C3.
> +	 */
> +	buf[0] = ACPI_PDC_REVISION_ID;
> +	buf[1] = 1;
> +	buf[2] = (ACPI_PDC_C_CAPABILITY_SMP | ACPI_PDC_EST_CAPABILITY_SWSMP);
> +
> +	set_xen_guest_handle(op.u.set_pminfo.pdc, buf);
> +
> +	if ((HYPERVISOR_dom0_op(&op) == 0) &&
> +	    (buf[2] & (ACPI_PDC_C_C1_FFH | ACPI_PDC_C_C2C3_FFH))) {
> +		cpuid_leaf5_ecx_val = cx;
> +		cpuid_leaf5_edx_val = dx;
> +	}
> +	return true;
> +#else
> +	return false;
> +#endif
> +}
>  static void __init xen_init_cpuid_mask(void)
>  {
>  	unsigned int ax, bx, cx, dx;
> @@ -261,6 +348,9 @@ static void __init xen_init_cpuid_mask(void)
>  	/* Xen will set CR4.OSXSAVE if supported and not disabled by force */
>  	if ((cx & xsave_mask) != xsave_mask)
>  		cpuid_leaf1_ecx_mask &= ~xsave_mask; /* disable XSAVE & OSXSAVE */
> +
> +	if (xen_check_mwait())
> +		cpuid_leaf1_ecx_set_mask = (1 << (X86_FEATURE_MWAIT % 32));
>  }
>  
>  static void xen_set_debugreg(int reg, unsigned long val)
> diff --git a/include/xen/interface/platform.h 
> b/include/xen/interface/platform.h
> index c168468..6220b98 100644
> --- a/include/xen/interface/platform.h
> +++ b/include/xen/interface/platform.h
> @@ -200,7 +200,7 @@ DEFINE_GUEST_HANDLE_STRUCT(xenpf_getidletime_t);
>  #define XEN_PM_CX   0
>  #define XEN_PM_PX   1
>  #define XEN_PM_TX   2
> -
> +#define XEN_PM_PDC  3
>  /* Px sub info type */
>  #define XEN_PX_PCT   1
>  #define XEN_PX_PSS   2
> @@ -286,6 +286,7 @@ struct xen_processor_performance {
>  };
>  DEFINE_GUEST_HANDLE_STRUCT(xen_processor_performance);
>  
> +DEFINE_GUEST_HANDLE(uint32_t);

Do you really need to introduce (step by step) handles for all those
uintNN_t types in a way different from Xen's
(__DEFINE_XEN_GUEST_HANDLE(uint32, uint32_t))?

Looks good to me otherwise.

Jan

>  struct xenpf_set_processor_pminfo {
>  	/* IN variables */
>  	uint32_t id;    /* ACPI CPU ID */
> @@ -293,6 +294,7 @@ struct xenpf_set_processor_pminfo {
>  	union {
>  		struct xen_processor_power          power;/* Cx: _CST/_CSD */
>  		struct xen_processor_performance    perf; /* Px: _PPC/_PCT/_PSS/_PSD */
> +		GUEST_HANDLE(uint32_t)              pdc;
>  	};
>  };
>  DEFINE_GUEST_HANDLE_STRUCT(xenpf_set_processor_pminfo);
> -- 
> 1.7.9.48.g85da4d

Re: [PATCH 2/3] xen/enlighten: Expose MWAIT and MWAIT_LEAF if hypervisor OKs it.

[Konrad Rzeszutek Wilk]

> > +DEFINE_GUEST_HANDLE(uint32_t);
> 
> Do you really need to introduce (step by step) handles for all those
> uintNN_t types in a way different from Xen's
> (__DEFINE_XEN_GUEST_HANDLE(uint32, uint32_t))?

Oh, no. Good eye - I will change it be uint32.

> 
> Looks good to me otherwise.

Thank you for time!

Re: [PATCH 2/3] xen/enlighten: Expose MWAIT and MWAIT_LEAF if hypervisor OKs it.

[Jan Beulich]

>>> On 25.02.12 at 00:52, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
>> > +DEFINE_GUEST_HANDLE(uint32_t);
>> 
>> Do you really need to introduce (step by step) handles for all those
>> uintNN_t types in a way different from Xen's
>> (__DEFINE_XEN_GUEST_HANDLE(uint32, uint32_t))?
> 
> Oh, no. Good eye - I will change it be uint32.

Might be worth converting the recently added counterpart of uint64_t
then too.

Jan

[PATCH 3/3] xen/processor-passthru: Provide an driver that passes struct acpi_processor data to the hypervisor.

[Konrad Rzeszutek Wilk]

The ACPI processor processes the _Pxx and the _Cx state information
which are populated in the 'struct acpi_processor' per-cpu structure.
We read the contents of that structure and pass it up the Xen hypervisor.

The ACPI processor along with the CPU freq driver does all the heavy-lifting
for us (filtering, calling ACPI functions, etc) so that the contents is correct.
After we are done parsing the information, we wait in case of hotplug CPUs
get loaded and then pass that information to the hypervisor.

[v1-v2: Initial RFC implementations that were posted]
[v3: Changed the name to passthru suggested by Pasi Kärkkäinen <pasik@iki.fi>]
[v4: Added vCPU != pCPU support - aka dom0_max_vcpus support]
[v5: Cleaned up the driver, fix bug under Athlon XP]
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 drivers/xen/Kconfig              |   14 +
 drivers/xen/Makefile             |    2 +-
 drivers/xen/processor-passthru.c |  492 ++++++++++++++++++++++++++++++++++++++
 3 files changed, 507 insertions(+), 1 deletions(-)
 create mode 100644 drivers/xen/processor-passthru.c

diff --git a/drivers/xen/Kconfig b/drivers/xen/Kconfig
index a1ced52..af5e062 100644
--- a/drivers/xen/Kconfig
+++ b/drivers/xen/Kconfig
@@ -178,4 +178,18 @@ config XEN_PRIVCMD
 	depends on XEN
 	default m
 
+config XEN_PROCESSOR_PASSTHRU
+	tristate "Processor passthrough driver for Xen"
+	depends on XEN
+	depends on ACPI_PROCESSOR
+	depends on X86
+	depends on CPU_FREQ
+	help
+	  This driver parses the processor structure and passes the information
+	  to the Xen hypervisor. It is used to allow the Xen hypervisor to have the
+	  full power management data and be able to select proper Cx and Pxx states.
+
+	  The driver should be loaded after acpi processor and cpufreq drivers have
+	  been loaded. If you do not know what to choose, select M here.
+
 endmenu
diff --git a/drivers/xen/Makefile b/drivers/xen/Makefile
index aa31337..ce235e7a 100644
--- a/drivers/xen/Makefile
+++ b/drivers/xen/Makefile
@@ -20,7 +20,7 @@ obj-$(CONFIG_SWIOTLB_XEN)		+= swiotlb-xen.o
 obj-$(CONFIG_XEN_DOM0)			+= pci.o
 obj-$(CONFIG_XEN_PCIDEV_BACKEND)	+= xen-pciback/
 obj-$(CONFIG_XEN_PRIVCMD)		+= xen-privcmd.o
-
+obj-$(CONFIG_XEN_PROCESSOR_PASSTHRU)	+= processor-passthru.o
 xen-evtchn-y				:= evtchn.o
 xen-gntdev-y				:= gntdev.o
 xen-gntalloc-y				:= gntalloc.o
diff --git a/drivers/xen/processor-passthru.c b/drivers/xen/processor-passthru.c
new file mode 100644
index 0000000..e4dff42
--- /dev/null
+++ b/drivers/xen/processor-passthru.c
@@ -0,0 +1,492 @@
+/*
+ * Copyright 2012 by Oracle Inc
+ * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
+ *
+ * This code borrows ideas from https://lkml.org/lkml/2011/11/30/249
+ * so many thanks go to Kevin Tian <kevin.tian@intel.com>
+ * and Yu Ke <ke.yu@intel.com>.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ */
+
+#include <linux/cpumask.h>
+#include <linux/cpufreq.h>
+#include <linux/kernel.h>
+#include <linux/kthread.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/types.h>
+#include <acpi/acpi_bus.h>
+#include <acpi/acpi_drivers.h>
+#include <acpi/processor.h>
+
+#include <xen/interface/platform.h>
+#include <asm/xen/hypercall.h>
+
+#define DRV_NAME "xen-processor-thru"
+MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>");
+MODULE_DESCRIPTION("ACPI Power Management driver to pass Cx and Pxx data to Xen hypervisor");
+MODULE_LICENSE("GPL");
+
+
+static int no_hypercall;
+MODULE_PARM_DESC(off, "Inhibit the hypercall.");
+module_param_named(off, no_hypercall, int, 0400);
+
+/*
+ * Mutex to protect the acpi_ids_done.
+ */
+static DEFINE_MUTEX(acpi_ids_mutex);
+/*
+ * Don't think convert this to cpumask_var_t or use cpumask_bit - as those
+ * shrink to nr_cpu_bits (which is dependent on possible_cpu), which can be
+ * less than what we want to put in.
+ */
+#define NR_ACPI_CPUS	NR_CPUS
+#define MAX_ACPI_BITS	(BITS_TO_LONGS(NR_ACPI_CPUS))
+static unsigned long *acpi_ids_done;
+/*
+ * Again, don't convert to cpumask - as we are reading the raw ACPI CPU ids
+ * which can go beyond what we presently see.
+ */
+static unsigned long *acpi_id_present;
+
+
+#define POLL_TIMER	msecs_to_jiffies(5000 /* 5 sec */)
+static struct task_struct *xen_processor_thread;
+
+static int xen_push_cxx_to_hypervisor(struct acpi_processor *_pr)
+{
+	struct xen_platform_op op = {
+		.cmd			= XENPF_set_processor_pminfo,
+		.interface_version	= XENPF_INTERFACE_VERSION,
+		.u.set_pminfo.id	= _pr->acpi_id,
+		.u.set_pminfo.type	= XEN_PM_CX,
+	};
+	struct xen_processor_cx *xen_cx, *xen_cx_states = NULL;
+	struct acpi_processor_cx *cx;
+	int i, ok, ret = 0;
+
+	xen_cx_states = kcalloc(_pr->power.count,
+				sizeof(struct xen_processor_cx), GFP_KERNEL);
+	if (!xen_cx_states)
+		return -ENOMEM;
+
+	for (ok = 0, i = 1; i <= _pr->power.count; i++) {
+		cx = &_pr->power.states[i];
+		if (!cx->valid)
+			continue;
+
+		xen_cx = &(xen_cx_states[ok++]);
+
+		xen_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO;
+		if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
+			xen_cx->reg.bit_width = 8;
+			xen_cx->reg.bit_offset = 0;
+			xen_cx->reg.access_size = 1;
+		} else {
+			xen_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE;
+			if (cx->entry_method == ACPI_CSTATE_FFH) {
+				/* NATIVE_CSTATE_BEYOND_HALT */
+				xen_cx->reg.bit_offset = 2;
+				xen_cx->reg.bit_width = 1; /* VENDOR_INTEL */
+			}
+			xen_cx->reg.access_size = 0;
+		}
+		xen_cx->reg.address = cx->address;
+
+		xen_cx->type = cx->type;
+		xen_cx->latency = cx->latency;
+		xen_cx->power = cx->power;
+
+		xen_cx->dpcnt = 0;
+		set_xen_guest_handle(xen_cx->dp, NULL);
+#ifdef DEBUG
+		pr_debug(DRV_NAME ": CX: ID:%d [C%d:%s] entry:%d\n", _pr->acpi_id,
+			 cx->type, cx->desc, cx->entry_method);
+#endif
+	}
+	if (!ok) {
+		pr_err(DRV_NAME ": No available Cx info for cpu %d\n", _pr->acpi_id);
+		kfree(xen_cx_states);
+		return -EINVAL;
+	}
+	op.u.set_pminfo.power.count = ok;
+	op.u.set_pminfo.power.flags.bm_control = _pr->flags.bm_control;
+	op.u.set_pminfo.power.flags.bm_check = _pr->flags.bm_check;
+	op.u.set_pminfo.power.flags.has_cst = _pr->flags.has_cst;
+	op.u.set_pminfo.power.flags.power_setup_done =
+		_pr->flags.power_setup_done;
+
+	set_xen_guest_handle(op.u.set_pminfo.power.states, xen_cx_states);
+
+	if (!no_hypercall)
+		ret = HYPERVISOR_dom0_op(&op);
+
+	if (ret)
+		pr_err(DRV_NAME "(CX): Hypervisor returned (%d) for ACPI ID: %d\n",
+		       ret, _pr->acpi_id);
+
+	kfree(xen_cx_states);
+
+	return ret;
+}
+static struct xen_processor_px *xen_copy_pss_data(struct acpi_processor *_pr,
+						  struct xen_processor_performance *xen_perf)
+{
+	struct xen_processor_px *xen_states = NULL;
+	int i;
+
+	BUILD_BUG_ON(sizeof(struct xen_processor_px) !=
+		     sizeof(struct acpi_processor_px));
+
+	xen_states = kcalloc(_pr->performance->state_count,
+			     sizeof(struct xen_processor_px), GFP_KERNEL);
+	if (!xen_states)
+		return ERR_PTR(-ENOMEM);
+
+	xen_perf->state_count = _pr->performance->state_count;
+	for (i = 0; i < _pr->performance->state_count; i++) {
+		/* Fortunatly for us, they are both the same size */
+		memcpy(&(xen_states[i]), &(_pr->performance->states[i]),
+		       sizeof(struct acpi_processor_px));
+	}
+	return xen_states;
+}
+static int xen_copy_psd_data(struct acpi_processor *_pr,
+			     struct xen_processor_performance *xen_perf)
+{
+	BUILD_BUG_ON(sizeof(struct xen_psd_package) !=
+		     sizeof(struct acpi_psd_package));
+
+	if (_pr->performance->shared_type != CPUFREQ_SHARED_TYPE_NONE) {
+		xen_perf->shared_type = _pr->performance->shared_type;
+
+		memcpy(&(xen_perf->domain_info), &(_pr->performance->domain_info),
+		       sizeof(struct acpi_psd_package));
+	} else {
+		if ((&cpu_data(0))->x86_vendor != X86_VENDOR_AMD)
+			return -EINVAL;
+
+		/* On AMD, the powernow-k8 is loaded before acpi_cpufreq
+		 * meaning that acpi_processor_preregister_performance never
+		 * gets called which would parse the _PSD. The only relevant
+		 * information from _PSD we need is whether it is HW_ALL or any
+		 * other type. AMD K8 >= are SW_ALL or SW_ANY, AMD K7<= HW_ANY.
+		 * This driver checks at the start whether it is K8 so it
+		 * if we get here it can only be K8.
+		 */
+		xen_perf->shared_type = CPUFREQ_SHARED_TYPE_ANY;
+		xen_perf->domain_info.coord_type = DOMAIN_COORD_TYPE_SW_ANY;
+		xen_perf->domain_info.num_processors = num_online_cpus();
+	}
+	return 0;
+}
+static int xen_copy_pct_data(struct acpi_pct_register *pct,
+			     struct xen_pct_register *_pct)
+{
+	/* It would be nice if you could just do 'memcpy(pct, _pct') but
+	 * sadly the Xen structure did not have the proper padding
+	 * so the descriptor field takes two (_pct) bytes instead of one (pct).
+	 */
+	_pct->descriptor = pct->descriptor;
+	_pct->length = pct->length;
+	_pct->space_id = pct->space_id;
+	_pct->bit_width = pct->bit_width;
+	_pct->bit_offset = pct->bit_offset;
+	_pct->reserved = pct->reserved;
+	_pct->address = pct->address;
+	return 0;
+}
+static int xen_push_pxx_to_hypervisor(struct acpi_processor *_pr)
+{
+	int ret = 0;
+	struct xen_platform_op op = {
+		.cmd			= XENPF_set_processor_pminfo,
+		.interface_version	= XENPF_INTERFACE_VERSION,
+		.u.set_pminfo.id	= _pr->acpi_id,
+		.u.set_pminfo.type	= XEN_PM_PX,
+	};
+	struct xen_processor_performance *xen_perf;
+	struct xen_processor_px *xen_states = NULL;
+
+	xen_perf = &op.u.set_pminfo.perf;
+
+	xen_perf->platform_limit = _pr->performance_platform_limit;
+	xen_perf->flags |= XEN_PX_PPC;
+	xen_copy_pct_data(&(_pr->performance->control_register),
+			  &xen_perf->control_register);
+	xen_copy_pct_data(&(_pr->performance->status_register),
+			  &xen_perf->status_register);
+	xen_perf->flags |= XEN_PX_PCT;
+	xen_states = xen_copy_pss_data(_pr, xen_perf);
+	if (!IS_ERR_OR_NULL(xen_states)) {
+		set_xen_guest_handle(xen_perf->states, xen_states);
+		xen_perf->flags |= XEN_PX_PSS;
+	}
+	if (!xen_copy_psd_data(_pr, xen_perf))
+		xen_perf->flags |= XEN_PX_PSD;
+
+	if (!no_hypercall)
+		ret = HYPERVISOR_dom0_op(&op);
+
+	if (ret)
+		pr_err(DRV_NAME "(_PXX): Hypervisor returned (%d) for ACPI ID %d\n",
+		       ret, _pr->acpi_id);
+
+	if (!IS_ERR_OR_NULL(xen_states))
+		kfree(xen_states);
+
+	return ret;
+}
+/*
+ * We read out the struct acpi_processor, and serialize access
+ * so that there is only one caller. This is so that we won't
+ * race with the CPU hotplug code (xen_cpu_soft_notify).
+ */
+static int xen_process_data(struct acpi_processor *_pr)
+{
+	int err = 0;
+
+	mutex_lock(&acpi_ids_mutex);
+	if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) {
+		mutex_unlock(&acpi_ids_mutex);
+		return -EBUSY;
+	}
+	if (_pr->flags.power)
+		err = xen_push_cxx_to_hypervisor(_pr);
+
+	if (_pr->performance && _pr->performance->states)
+		err |= xen_push_pxx_to_hypervisor(_pr);
+
+	mutex_unlock(&acpi_ids_mutex);
+	return err;
+}
+static acpi_status
+xen_read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv)
+{
+	u32 acpi_id;
+	acpi_status status;
+	acpi_object_type acpi_type;
+	unsigned long long tmp;
+	union acpi_object object = { 0 };
+	struct acpi_buffer buffer = { sizeof(union acpi_object), &object };
+
+	status = acpi_get_type(handle, &acpi_type);
+	if (ACPI_FAILURE(status))
+		return AE_OK;
+
+	switch (acpi_type) {
+	case ACPI_TYPE_PROCESSOR:
+		status = acpi_evaluate_object(handle, NULL, NULL, &buffer);
+		if (ACPI_FAILURE(status))
+			return AE_OK;
+		acpi_id = object.processor.proc_id;
+		break;
+	case ACPI_TYPE_DEVICE:
+		status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp);
+		if (ACPI_FAILURE(status))
+			return AE_OK;
+		acpi_id = tmp;
+		break;
+	default:
+		return AE_OK;
+	}
+	if (acpi_id > NR_ACPI_CPUS) {
+		WARN_ONCE(1, "There are %d ACPI processors, but kernel can only do %d!\n",
+		     acpi_id, NR_ACPI_CPUS);
+		return AE_OK;
+	}
+	__set_bit(acpi_id, acpi_id_present);
+
+	return AE_OK;
+}
+static unsigned int xen_acpi_ids_more(void)
+{
+	unsigned int n = 0;
+
+	acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
+			    ACPI_UINT32_MAX,
+			    xen_read_acpi_id, NULL, NULL, NULL);
+	acpi_get_devices("ACPI0007", xen_read_acpi_id, NULL, NULL);
+
+	mutex_lock(&acpi_ids_mutex);
+	if (!bitmap_equal(acpi_id_present, acpi_ids_done, MAX_ACPI_BITS))
+		n = bitmap_weight(acpi_id_present, MAX_ACPI_BITS);
+	mutex_unlock(&acpi_ids_mutex);
+
+	return n;
+}
+
+static int xen_processor_check(void)
+{
+	struct cpufreq_policy *policy;
+	struct acpi_processor *pr_backup = NULL;
+	int cpu, err = 0;
+
+	cpu = get_cpu();
+	put_cpu();
+	policy = cpufreq_cpu_get(cpu);
+	if (!policy)
+		return -EBUSY;
+
+	get_online_cpus();
+	for_each_online_cpu(cpu) {
+		struct acpi_processor *_pr;
+
+		_pr = per_cpu(processors, cpu /* APIC ID */);
+		if (!_pr)
+			continue;
+
+		if (!pr_backup) {
+			pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL);
+			memcpy(pr_backup, _pr, sizeof(struct acpi_processor));
+		}
+		(void)xen_process_data(_pr);
+	}
+	put_online_cpus();
+
+	cpufreq_cpu_put(policy);
+
+	/* All online CPUs have been processed at this stage. Now verify
+	 * whether in fact "online CPUs" == physical CPUs.
+	 */
+	acpi_id_present = kcalloc(MAX_ACPI_BITS, sizeof(unsigned long), GFP_KERNEL);
+	if (!acpi_id_present) {
+		err = -ENOMEM;
+		goto err_out;
+	}
+	memset(acpi_id_present, 0, MAX_ACPI_BITS * sizeof(unsigned long));
+
+	if (xen_acpi_ids_more() && pr_backup) {
+		for_each_set_bit(cpu, acpi_id_present, MAX_ACPI_BITS) {
+			pr_backup->acpi_id = cpu;
+			/* We will get -EBUSY if it has been programmed already. */
+			(void)xen_process_data(pr_backup);
+		}
+	}
+	kfree(acpi_id_present);
+	acpi_id_present = NULL;
+err_out:
+	kfree(pr_backup);
+	pr_backup = NULL;
+	return err;
+}
+/*
+ * The purpose of this timer/thread is to wait for the ACPI processor
+ * and CPUfreq drivers to load up and parse the Pxx and Cxx information
+ * before we attempt to read it.
+ */
+static void xen_processor_timeout(unsigned long arg)
+{
+	wake_up_process((struct task_struct *)arg);
+}
+static int xen_processor_thread_func(void *dummy)
+{
+	struct timer_list timer;
+	int err = 0;
+
+	setup_deferrable_timer_on_stack(&timer, xen_processor_timeout,
+					(unsigned long)current);
+	do {
+		__set_current_state(TASK_INTERRUPTIBLE);
+		mod_timer(&timer, jiffies + POLL_TIMER);
+		schedule();
+		err = xen_processor_check();
+		if (err != -EBUSY)
+			break;
+	} while (!kthread_should_stop());
+
+	if (err)
+		pr_err(DRV_NAME ": Failed to upload data (%d)!\n", err);
+	del_timer_sync(&timer);
+	destroy_timer_on_stack(&timer);
+	return 0;
+}
+
+static int xen_cpu_soft_notify(struct notifier_block *nfb,
+			       unsigned long action, void *hcpu)
+{
+	unsigned int cpu = (unsigned long)hcpu;
+	struct acpi_processor *_pr = per_cpu(processors, cpu);
+
+	if (action == CPU_ONLINE && _pr)
+		(void)xen_process_data(_pr);
+
+	return NOTIFY_OK;
+}
+
+static struct notifier_block xen_cpu_notifier = {
+	.notifier_call = xen_cpu_soft_notify,
+	.priority = -1, /* Be the last one */
+};
+
+static int __init check_prereq(void)
+{
+	struct cpuinfo_x86 *c = &cpu_data(0);
+
+	if (!xen_initial_domain())
+		return -ENODEV;
+
+	if (!acpi_gbl_FADT.smi_command)
+		return -ENODEV;
+
+	if (c->x86_vendor == X86_VENDOR_INTEL) {
+		if (!cpu_has(c, X86_FEATURE_EST))
+			return -ENODEV;
+
+		return 0;
+	}
+	if (c->x86_vendor == X86_VENDOR_AMD) {
+		u32 hi = 0, lo = 0;
+		/* Copied from powernow-k8.h, can't include ../cpufreq/powernow
+		 * as we get compile warnings for the static functions.
+		 */
+#define MSR_PSTATE_CUR_LIMIT    0xc0010061 /* pstate current limit MSR */
+		rdmsr(MSR_PSTATE_CUR_LIMIT, lo, hi);
+
+		/* If the MSR cannot provide the data, the powernow-k8
+		 * won't process the data properly either.
+		 */
+		if (hi || lo)
+			return 0;
+	}
+	return -ENODEV;
+}
+
+static int __init xen_processor_passthru_init(void)
+{
+	int rc = check_prereq();
+
+	if (rc)
+		return rc;
+
+	acpi_ids_done = kcalloc(MAX_ACPI_BITS, sizeof(unsigned long), GFP_KERNEL);
+	if (!acpi_ids_done)
+		return -ENOMEM;
+	memset(acpi_ids_done, 0, MAX_ACPI_BITS * sizeof(unsigned long));
+	xen_processor_thread = kthread_run(xen_processor_thread_func, NULL, DRV_NAME);
+	if (IS_ERR(xen_processor_thread)) {
+		pr_err(DRV_NAME ": Failed to create thread. Aborting.\n");
+		return -ENOMEM;
+	}
+	register_hotcpu_notifier(&xen_cpu_notifier);
+	return 0;
+}
+static void __exit xen_processor_passthru_exit(void)
+{
+	unregister_hotcpu_notifier(&xen_cpu_notifier);
+	if (xen_processor_thread)
+		kthread_stop(xen_processor_thread);
+	kfree(acpi_ids_done);
+}
+late_initcall(xen_processor_passthru_init);
+module_exit(xen_processor_passthru_exit);
-- 
1.7.9.48.g85da4d

Re: [PATCH] processor passthru - upload _Cx and _Pxx data to hypervisor (v5).

[Jan Beulich]

>>> On 23.02.12 at 23:31, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> This module (processor-passthru)  collects the information that the cpufreq
> drivers and the ACPI processor code save in the 'struct acpi_processor' and
> then uploads it to the hypervisor.

Thus looks conceptually wrong to me - there shouldn't be a need for a
CPUFreq driver to be loaded in Dom0 (or your module should masquerade
as the one and only suitable one).

> On the hypervisor side, it requires this patch on AMD:
> # HG changeset patch
> # Parent aea8cfac8cf1afe397f2e1d422a852008d8a83fe
> traps: AMD PM RDMSRs (MSR_K8_PSTATE_CTRL, etc)
> 
> The restriction to read and write the AMD power management MSRs is gated if 
> the
> domain 0 is the PM domain (so FREQCTL_dom0_kernel is set). But we can
> relax this restriction and allow the privileged domain to read the MSRs
> (but not write). This allows the priviliged domain to harvest the power
> management information (ACPI _PSS states) and send it to the hypervisor.

Why would accessing these MSRs be necessary here, when it isn't
for non-pvops? Perhaps only because you want a CPUFreq driver
loaded?

Jan

> This patch works fine with older classic dom0 (2.6.32) and with
> AMD K7 and K8 boxes.
> 
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> diff -r aea8cfac8cf1 xen/arch/x86/traps.c
> --- a/xen/arch/x86/traps.c	Thu Feb 23 13:23:02 2012 -0500
> +++ b/xen/arch/x86/traps.c	Thu Feb 23 13:29:00 2012 -0500
> @@ -2484,7 +2484,7 @@ static int emulate_privileged_op(struct 
>          case MSR_K8_PSTATE7:
>              if ( boot_cpu_data.x86_vendor != X86_VENDOR_AMD )
>                  goto fail;
> -            if ( !is_cpufreq_controller(v->domain) )
> +            if ( !is_cpufreq_controller(v->domain) && !IS_PRIV(v->domain) )
>              {
>                  regs->eax = regs->edx = 0;
>                  break;

Re: [PATCH] processor passthru - upload _Cx and _Pxx data to hypervisor (v5).

[Konrad Rzeszutek Wilk]

On Fri, Feb 24, 2012 at 10:23:42AM +0000, Jan Beulich wrote:
> >>> On 23.02.12 at 23:31, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> > This module (processor-passthru)  collects the information that the cpufreq
> > drivers and the ACPI processor code save in the 'struct acpi_processor' and
> > then uploads it to the hypervisor.
> 
> Thus looks conceptually wrong to me - there shouldn't be a need for a
> CPUFreq driver to be loaded in Dom0 (or your module should masquerade
> as the one and only suitable one).

I piggyback on the generic cpufreq drivers to collect the information they
have evaluated.

I can make the driver a cpufreq one but there does not seem to be a way
from the kernel to force a specific driver to say "use me". I could write
it naturally, but not sure what the usage case is except for the driver
I wrote. But perhaps there is also for the cpufreq powernow-k8 and acpi-processor
so that they can function without the need for strict compile order
(where powernow-k8 MUST be loaded before acpi-processor).

> 
> > On the hypervisor side, it requires this patch on AMD:
> > # HG changeset patch
> > # Parent aea8cfac8cf1afe397f2e1d422a852008d8a83fe
> > traps: AMD PM RDMSRs (MSR_K8_PSTATE_CTRL, etc)
> > 
> > The restriction to read and write the AMD power management MSRs is gated if 
> > the
> > domain 0 is the PM domain (so FREQCTL_dom0_kernel is set). But we can
> > relax this restriction and allow the privileged domain to read the MSRs
> > (but not write). This allows the priviliged domain to harvest the power
> > management information (ACPI _PSS states) and send it to the hypervisor.
> 
> Why would accessing these MSRs be necessary here, when it isn't
> for non-pvops? Perhaps only because you want a CPUFreq driver
> loaded?

Correct. The powernow-k8
> 
> Jan
> 
> > This patch works fine with older classic dom0 (2.6.32) and with
> > AMD K7 and K8 boxes.
> > 
> > Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
> > diff -r aea8cfac8cf1 xen/arch/x86/traps.c
> > --- a/xen/arch/x86/traps.c	Thu Feb 23 13:23:02 2012 -0500
> > +++ b/xen/arch/x86/traps.c	Thu Feb 23 13:29:00 2012 -0500
> > @@ -2484,7 +2484,7 @@ static int emulate_privileged_op(struct 
> >          case MSR_K8_PSTATE7:
> >              if ( boot_cpu_data.x86_vendor != X86_VENDOR_AMD )
> >                  goto fail;
> > -            if ( !is_cpufreq_controller(v->domain) )
> > +            if ( !is_cpufreq_controller(v->domain) && !IS_PRIV(v->domain) )
> >              {
> >                  regs->eax = regs->edx = 0;
> >                  break;
> 
> 

Re: [PATCH] processor passthru - upload _Cx and _Pxx data to hypervisor (v5).

[Konrad Rzeszutek Wilk]

On Fri, Feb 24, 2012 at 10:23:42AM +0000, Jan Beulich wrote:
> >>> On 23.02.12 at 23:31, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> > This module (processor-passthru)  collects the information that the cpufreq
> > drivers and the ACPI processor code save in the 'struct acpi_processor' and
> > then uploads it to the hypervisor.
> 
> Thus looks conceptually wrong to me - there shouldn't be a need for a
> CPUFreq driver to be loaded in Dom0 (or your module should masquerade
> as the one and only suitable one).

So before your email I had been thinking that b/c of the cpuidle rework
by Len it meant that when the cpufreq drivers are active - they would be started
from the cpu_idle call - and since  cpu_idle call ends up being default_idle on
pvops (which calls safe_halt) that would be fine. This is the work that Len did
"cpuidle: replace xen access to x86 pm_idle and default_idle" and
"cpuidle: stop depending on pm_idle" 

But cpufreq != cpuidle != cpufreq governor, and they all are run by different rules.
The ondemand cpufreq governor for example runs a timer and calls the appropiate cpufreq
driver. So with these patches I posted we end up with a cpufreq driver in the kernel
and in Xen hypervisor - both of them trying to change Pstates. Not good (to be fair,
if powernow-k8/acpi-cpufreq would try it via WRMSR -  those would up being trapped and
ignored by the hypervisor. I am not sure about the outw though).

The pre-RFC version of this posted driver implemented a cpufreq governor that was
nop and for future work was going to make a hypercall to get the true cpufreq value
to report properly in /proc/cpuinfo - but I hadn't figured out a way to make it be
the default one dynamically.

Perhaps having xencommons do 
echo "xen" > /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

And s/processor-passthru/cpufreq-xen/ would do it? That would eliminate the [performance,
ondemand,powersave,etc] cpufreq governors from calling into the cpufreq drivers to alter P-states.

Let me CC Dave Jones and the cpufreq mailing list - perhaps they might have
some ideas?
[The patch is http://lwn.net/Articles/483668/]

Re: [PATCH] processor passthru - upload _Cx and _Pxx data to hypervisor (v5).

[Jan Beulich]

>>> On 25.02.12 at 01:21, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> wrote:
> But cpufreq != cpuidle != cpufreq governor, and they all are run by 
> different rules.
> The ondemand cpufreq governor for example runs a timer and calls the 
> appropiate cpufreq
> driver. So with these patches I posted we end up with a cpufreq driver in 
> the kernel
> and in Xen hypervisor - both of them trying to change Pstates. Not good (to 
> be fair,
> if powernow-k8/acpi-cpufreq would try it via WRMSR -  those would up being 
> trapped and
> ignored by the hypervisor. I am not sure about the outw though).

I'm not aware of any trapping that would be done on the I/O port here;
it could be added, though (i.e. the ports removed from the list of
allowed ports of Dom0 once they become known to the hypervisor).

> The pre-RFC version of this posted driver implemented a cpufreq governor that 
> was
> nop and for future work was going to make a hypercall to get the true 
> cpufreq value
> to report properly in /proc/cpuinfo - but I hadn't figured out a way to make 
> it be
> the default one dynamically.
> 
> Perhaps having xencommons do 
> echo "xen" > /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
> 
> And s/processor-passthru/cpufreq-xen/ would do it? That would eliminate the 
> [performance,
> ondemand,powersave,etc] cpufreq governors from calling into the cpufreq 
> drivers to alter P-states.

Except that you want this to be a cpufreq driver, not a governor.

Jan