Andre Przywara
2012-May-29 11:07 UTC
[PATCH] xen: filter APERFMPERF feature for kernel usage
Xen PV kernels allow access to the APERF/MPERF registers to read the effective frequency. Access to the MSRs is however redirected to the currently scheduled physical CPU, making consecutive read and compares unreliable. In addition each rdmsr traps into the hypervisor. So to avoid bogus readouts and expensive traps, disable the kernel internal feature flag for APERF/MPERF if running under Xen. This will a) remove the aperfmperf flag from /proc/cpuinfo b) not mislead the power scheduler (arch/x86/kernel/cpu/sched.c) to use the feature to improve scheduling (by default disabled) c) not mislead the cpufreq driver to use the MSRs This does not cover userland programs which access the MSRs via the device file interface, but this will be addressed separately. Signed-off-by: Andre Przywara <andre.przywara@amd.com> Cc: stable@vger.kernel.org # v3.0+ --- arch/x86/xen/enlighten.c | 8 ++++++++ 1 files changed, 8 insertions(+), 0 deletions(-) diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c index 95dccce..dfbe1af 100644 --- a/arch/x86/xen/enlighten.c +++ b/arch/x86/xen/enlighten.c @@ -207,6 +207,9 @@ static void __init xen_banner(void) xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : ""); } +#define CPUID_THERM_POWER_LEAF 6 +#define APERFMPERF_PRESENT 0 + static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0; static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0; @@ -240,6 +243,11 @@ static void xen_cpuid(unsigned int *ax, unsigned int *bx, *dx = cpuid_leaf5_edx_val; return; + case CPUID_THERM_POWER_LEAF: + /* Disabling APERFMPERF for kernel usage */ + maskecx = ~(1 << APERFMPERF_PRESENT); + break; + case 0xb: /* Suppress extended topology stuff */ maskebx = 0; -- 1.7.4.4
Konrad Rzeszutek Wilk
2012-May-31 16:17 UTC
Re: [PATCH] xen: filter APERFMPERF feature for kernel usage
On Tue, May 29, 2012 at 01:07:31PM +0200, Andre Przywara wrote:> Xen PV kernels allow access to the APERF/MPERF registers to read the > effective frequency. Access to the MSRs is however redirected to the > currently scheduled physical CPU, making consecutive read and > compares unreliable. In addition each rdmsr traps into the hypervisor. > So to avoid bogus readouts and expensive traps, disable the kernel > internal feature flag for APERF/MPERF if running under Xen. > This will > a) remove the aperfmperf flag from /proc/cpuinfo > b) not mislead the power scheduler (arch/x86/kernel/cpu/sched.c) to > use the feature to improve scheduling (by default disabled) > c) not mislead the cpufreq driver to use the MSRs > > This does not cover userland programs which access the MSRs via the > device file interface, but this will be addressed separately. > > Signed-off-by: Andre Przywara <andre.przywara@amd.com> > Cc: stable@vger.kernel.org # v3.0+applied.> --- > arch/x86/xen/enlighten.c | 8 ++++++++ > 1 files changed, 8 insertions(+), 0 deletions(-) > > diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c > index 95dccce..dfbe1af 100644 > --- a/arch/x86/xen/enlighten.c > +++ b/arch/x86/xen/enlighten.c > @@ -207,6 +207,9 @@ static void __init xen_banner(void) > xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : ""); > } > > +#define CPUID_THERM_POWER_LEAF 6 > +#define APERFMPERF_PRESENT 0 > + > static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0; > static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0; > > @@ -240,6 +243,11 @@ static void xen_cpuid(unsigned int *ax, unsigned int *bx, > *dx = cpuid_leaf5_edx_val; > return; > > + case CPUID_THERM_POWER_LEAF: > + /* Disabling APERFMPERF for kernel usage */ > + maskecx = ~(1 << APERFMPERF_PRESENT); > + break; > + > case 0xb: > /* Suppress extended topology stuff */ > maskebx = 0; > -- > 1.7.4.4 >