|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH v2] x86/msr: fix handling of MSR_IA32_PERF_{STATUS/CTL}
On Mon, Nov 09, 2020 at 05:38:19PM +0000, Andrew Cooper wrote:
> From: Roger Pau Monné <roger.pau@xxxxxxxxxx>
>
> Currently a PV hardware domain can also be given control over the CPU
> frequency, and such guest is allowed to write to MSR_IA32_PERF_CTL.
> However since commit 322ec7c89f6 the default behavior has been changed
> to reject accesses to not explicitly handled MSRs, preventing PV
> guests that manage CPU frequency from reading
> MSR_IA32_PERF_{STATUS/CTL}.
>
> Additionally some HVM guests (Windows at least) will attempt to read
> MSR_IA32_PERF_CTL and will panic if given back a #GP fault:
>
> vmx.c:3035:d8v0 RDMSR 0x00000199 unimplemented
> d8v0 VIRIDIAN CRASH: 3b c0000096 fffff806871c1651 ffffda0253683720 0
>
> Move the handling of MSR_IA32_PERF_{STATUS/CTL} to the common MSR
> handling shared between HVM and PV guests, and add an explicit case
> for reads to MSR_IA32_PERF_{STATUS/CTL}.
>
> Restore previous behavior and allow PV guests with the required
> permissions to read the contents of the mentioned MSRs. Non privileged
> guests will get 0 when trying to read those registers, as writes to
> MSR_IA32_PERF_CTL by such guest will already be silently dropped.
>
> Fixes: 322ec7c89f6 ('x86/pv: disallow access to unknown MSRs')
> Fixes: 84e848fd7a1 ('x86/hvm: disallow access to unknown MSRs')
> Signed-off-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
Reviewed-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> ---
> CC: Jan Beulich <JBeulich@xxxxxxxx>
> CC: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> CC: Wei Liu <wl@xxxxxxx>
>
> v2:
> * fix is_cpufreq_controller() to exclude PVH dom0, and collapse to nothing in
> !CONFIG_PV builds
> * Drop the cross-vendor checks. It isn't possible to configure dom0 as
> cross-vendor, and anyone using is_cpufreq_controller() without an exact
> model match has far bigger problems.
I was on the verge of doing this in v1, but wasn't really sure whether
there was any use case to change the vendor for dom0 cpuid.
> * At least Centaur implements these MSRs. Add access.
> ---
> xen/arch/x86/msr.c | 34 ++++++++++++++++++++++++++++++++++
> xen/arch/x86/pv/emul-priv-op.c | 14 --------------
> xen/include/xen/sched.h | 17 +++++++++++++++++
> 3 files changed, 51 insertions(+), 14 deletions(-)
>
> diff --git a/xen/arch/x86/msr.c b/xen/arch/x86/msr.c
> index 9c69ef8792..0a8ae4d22c 100644
> --- a/xen/arch/x86/msr.c
> +++ b/xen/arch/x86/msr.c
> @@ -242,6 +242,25 @@ int guest_rdmsr(struct vcpu *v, uint32_t msr, uint64_t
> *val)
> goto gp_fault;
> break;
>
> + /*
> + * These MSRs are not enumerated in CPUID. They have been around
> + * since the Pentium 4, and implemented by other vendors.
> + *
> + * Some versions of Windows try reading these before setting up a #GP
> + * handler, and Linux has several unguarded reads as well. Provide
> + * RAZ semantics, in general, but permit a cpufreq controller dom0 to
> + * have full access.
> + */
> + case MSR_IA32_PERF_STATUS:
> + case MSR_IA32_PERF_CTL:
> + if ( !(cp->x86_vendor & (X86_VENDOR_INTEL | X86_VENDOR_CENTAUR)) )
> + goto gp_fault;
> +
> + *val = 0;
> + if ( likely(!is_cpufreq_controller(d)) || rdmsr_safe(msr, *val) == 0
> )
> + break;
> + goto gp_fault;
> +
> case MSR_IA32_THERM_STATUS:
> if ( cp->x86_vendor != X86_VENDOR_INTEL )
> goto gp_fault;
> @@ -448,6 +467,21 @@ int guest_wrmsr(struct vcpu *v, uint32_t msr, uint64_t
> val)
> goto gp_fault;
> break;
>
> + /*
> + * This MSR are not enumerated in CPUID. It has been around since
> the
> + * Pentium 4, and implemented by other vendors.
> + *
> + * To match the RAZ semantics, implement as write-discard, except for
> + * a cpufreq controller dom0 which has full access.
> + */
> + case MSR_IA32_PERF_CTL:
> + if ( !(cp->x86_vendor & (X86_VENDOR_INTEL | X86_VENDOR_CENTAUR)) )
> + goto gp_fault;
> +
> + if ( likely(!is_cpufreq_controller(d)) || wrmsr_safe(msr, val) == 0 )
> + break;
> + goto gp_fault;
> +
> case MSR_X2APIC_FIRST ... MSR_X2APIC_LAST:
> if ( !is_hvm_domain(d) || v != curr )
> goto gp_fault;
> diff --git a/xen/arch/x86/pv/emul-priv-op.c b/xen/arch/x86/pv/emul-priv-op.c
> index 7cc16d6eda..dbceed8a05 100644
> --- a/xen/arch/x86/pv/emul-priv-op.c
> +++ b/xen/arch/x86/pv/emul-priv-op.c
> @@ -849,12 +849,6 @@ static inline uint64_t guest_misc_enable(uint64_t val)
> return val;
> }
>
> -static inline bool is_cpufreq_controller(const struct domain *d)
> -{
> - return ((cpufreq_controller == FREQCTL_dom0_kernel) &&
> - is_hardware_domain(d));
> -}
> -
> static uint64_t guest_efer(const struct domain *d)
> {
> uint64_t val;
> @@ -1121,14 +1115,6 @@ static int write_msr(unsigned int reg, uint64_t val,
> return X86EMUL_OKAY;
> break;
>
> - case MSR_IA32_PERF_CTL:
> - if ( boot_cpu_data.x86_vendor != X86_VENDOR_INTEL )
> - break;
> - if ( likely(!is_cpufreq_controller(currd)) ||
> - wrmsr_safe(reg, val) == 0 )
> - return X86EMUL_OKAY;
> - break;
> -
> case MSR_IA32_THERM_CONTROL:
> case MSR_IA32_ENERGY_PERF_BIAS:
> if ( boot_cpu_data.x86_vendor != X86_VENDOR_INTEL )
> diff --git a/xen/include/xen/sched.h b/xen/include/xen/sched.h
> index d8ed83f869..b4d3e53310 100644
> --- a/xen/include/xen/sched.h
> +++ b/xen/include/xen/sched.h
> @@ -1069,6 +1069,23 @@ extern enum cpufreq_controller {
> FREQCTL_none, FREQCTL_dom0_kernel, FREQCTL_xen
> } cpufreq_controller;
>
> +static always_inline bool is_cpufreq_controller(const struct domain *d)
> +{
> + /*
> + * A PV dom0 can be nominated as the cpufreq controller, instead of using
> + * Xen's cpufreq driver, at which point dom0 gets direct access to
> certain
> + * MSRs.
> + *
> + * This interface only works when dom0 is identity pinned and has the
> same
> + * number of vCPUs as pCPUs on the system.
> + *
> + * It would be far better to paravirtualise the interface.
> + */
Would it be useful to add an assert here that the domain cpuid vendor
and the BSP cpuid vendor match?
Is it even possible to fake a different cpuid vendor for dom0?
Thanks, Roger.
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |