On Fri, Oct 18, 2019 at 05:37:23PM +0800, Xiaoyao Li wrote:
> Move the MSR bitmap capability check from vmx_disable_intercept_for_msr()
> and vmx_enable_intercept_for_msr(), so that we can do the check far
> early before we really want to touch the bitmap.
>
> Also, we can move the common MSR not-intercept setup to where msr bitmap
> is actually used.
>
> Signed-off-by: Xiaoyao Li <xiaoyao.li@xxxxxxxxx>
> ---
> Changes in v2:
> - Remove the check of cpu_has_vmx_msr_bitmap() from
> vmx_{disable,enable}_intercept_for_msr (Krish)
> ---
> arch/x86/kvm/vmx/vmx.c | 65 +++++++++++++++++++++---------------------
> 1 file changed, 33 insertions(+), 32 deletions(-)
>
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index b083316a598d..017689d0144e 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -343,8 +343,8 @@ module_param_cb(vmentry_l1d_flush, &vmentry_l1d_flush_ops, NULL, 0644);
>
> static bool guest_state_valid(struct kvm_vcpu *vcpu);
> static u32 vmx_segment_access_rights(struct kvm_segment *var);
> -static __always_inline void vmx_disable_intercept_for_msr(unsigned long *msr_bitmap,
> - u32 msr, int type);
> +static __always_inline void vmx_set_intercept_for_msr(unsigned long *msr_bitmap,
> + u32 msr, int type, bool value);
>
> void vmx_vmexit(void);
>
> @@ -2000,9 +2000,9 @@ static int vmx_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> * in the merging. We update the vmcs01 here for L1 as well
> * since it will end up touching the MSR anyway now.
> */
> - vmx_disable_intercept_for_msr(vmx->vmcs01.msr_bitmap,
> - MSR_IA32_SPEC_CTRL,
> - MSR_TYPE_RW);
> + vmx_set_intercept_for_msr(vmx->vmcs01.msr_bitmap,
> + MSR_IA32_SPEC_CTRL,
> + MSR_TYPE_RW, false);
IMO this is a net negative. The explicit "disable" is significantly more
intuitive than "set" with a %false param, e.g. at a quick glance it would
be easy to think this code is "setting", i.e. "enabling" interception.
> break;
> case MSR_IA32_PRED_CMD:
> if (!msr_info->host_initiated &&
> @@ -2028,8 +2028,9 @@ static int vmx_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> * vmcs02.msr_bitmap here since it gets completely overwritten
> * in the merging.
> */
> - vmx_disable_intercept_for_msr(vmx->vmcs01.msr_bitmap, MSR_IA32_PRED_CMD,
> - MSR_TYPE_W);
> + vmx_set_intercept_for_msr(vmx->vmcs01.msr_bitmap,
> + MSR_IA32_PRED_CMD,
> + MSR_TYPE_W, false);
> break;
> case MSR_IA32_CR_PAT:
> if (!kvm_pat_valid(data))
> @@ -3599,9 +3600,6 @@ static __always_inline void vmx_disable_intercept_for_msr(unsigned long *msr_bit
> {
> int f = sizeof(unsigned long);
>
> - if (!cpu_has_vmx_msr_bitmap())
> - return;
As above, I'd rather keep these here. Functionally it changes nothing on
CPUs with an MSR bitmap. For old CPUs, it saves all of two uops in paths
that aren't performance critical.
> -
> if (static_branch_unlikely(&enable_evmcs))
> evmcs_touch_msr_bitmap();
>
> @@ -3637,9 +3635,6 @@ static __always_inline void vmx_enable_intercept_for_msr(unsigned long *msr_bitm
> {
> int f = sizeof(unsigned long);
>
> - if (!cpu_has_vmx_msr_bitmap())
> - return;
> -
> if (static_branch_unlikely(&enable_evmcs))
> evmcs_touch_msr_bitmap();
>
> @@ -3673,6 +3668,9 @@ static __always_inline void vmx_enable_intercept_for_msr(unsigned long *msr_bitm
> static __always_inline void vmx_set_intercept_for_msr(unsigned long *msr_bitmap,
> u32 msr, int type, bool value)
> {
> + if (!cpu_has_vmx_msr_bitmap())
> + return;
> +
> if (value)
> vmx_enable_intercept_for_msr(msr_bitmap, msr, type);
> else
> @@ -4163,11 +4161,30 @@ static void ept_set_mmio_spte_mask(void)
>
> static void vmx_vmcs_setup(struct vcpu_vmx *vmx)
> {
> + unsigned long *msr_bitmap;
> +
> if (nested)
> nested_vmx_vmcs_setup();
>
> - if (cpu_has_vmx_msr_bitmap())
> + if (cpu_has_vmx_msr_bitmap()) {
> + msr_bitmap = vmx->vmcs01.msr_bitmap;
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_TSC, MSR_TYPE_R);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_FS_BASE, MSR_TYPE_RW);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_GS_BASE, MSR_TYPE_RW);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_KERNEL_GS_BASE, MSR_TYPE_RW);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_SYSENTER_CS, MSR_TYPE_RW);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_SYSENTER_ESP, MSR_TYPE_RW);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_SYSENTER_EIP, MSR_TYPE_RW);
> + if (kvm_cstate_in_guest(vmx->vcpu.kvm)) {
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C1_RES, MSR_TYPE_R);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C3_RESIDENCY, MSR_TYPE_R);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C6_RESIDENCY, MSR_TYPE_R);
> + vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C7_RESIDENCY, MSR_TYPE_R);
> + }
> +
> vmcs_write64(MSR_BITMAP, __pa(vmx->vmcs01.msr_bitmap));
> + }
> + vmx->msr_bitmap_mode = 0;
Zeroing msr_bitmap_mode can be skipped as well.
> vmcs_write64(VMCS_LINK_POINTER, -1ull); /* 22.3.1.5 */
>
> @@ -6074,7 +6091,8 @@ void vmx_set_virtual_apic_mode(struct kvm_vcpu *vcpu)
> }
> secondary_exec_controls_set(vmx, sec_exec_control);
>
> - vmx_update_msr_bitmap(vcpu);
> + if (cpu_has_vmx_msr_bitmap())
> + vmx_update_msr_bitmap(vcpu);
> }
>
> static void vmx_set_apic_access_page_addr(struct kvm_vcpu *vcpu, hpa_t hpa)
> @@ -6688,7 +6706,6 @@ static struct kvm_vcpu *vmx_create_vcpu(struct kvm *kvm, unsigned int id)
> {
> int err;
> struct vcpu_vmx *vmx;
> - unsigned long *msr_bitmap;
> int i, cpu;
>
> BUILD_BUG_ON_MSG(offsetof(struct vcpu_vmx, vcpu) != 0,
> @@ -6745,22 +6762,6 @@ static struct kvm_vcpu *vmx_create_vcpu(struct kvm *kvm, unsigned int id)
> if (err < 0)
> goto free_msrs;
>
> - msr_bitmap = vmx->vmcs01.msr_bitmap;
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_TSC, MSR_TYPE_R);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_FS_BASE, MSR_TYPE_RW);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_GS_BASE, MSR_TYPE_RW);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_KERNEL_GS_BASE, MSR_TYPE_RW);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_SYSENTER_CS, MSR_TYPE_RW);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_SYSENTER_ESP, MSR_TYPE_RW);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_IA32_SYSENTER_EIP, MSR_TYPE_RW);
> - if (kvm_cstate_in_guest(kvm)) {
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C1_RES, MSR_TYPE_R);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C3_RESIDENCY, MSR_TYPE_R);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C6_RESIDENCY, MSR_TYPE_R);
> - vmx_disable_intercept_for_msr(msr_bitmap, MSR_CORE_C7_RESIDENCY, MSR_TYPE_R);
> - }
> - vmx->msr_bitmap_mode = 0;
Keep this code here to be consistent with the previous change that moved
the guest_msrs intialization *out* of the VMCS specific function. Both
are collateral pages that are not directly part of the VMCS.
I'd be tempted to use a goto to skip the code, the line length is bad
enough as it is, e.g.:
if (!cpu_has_vmx_msr_bitmap())
goto skip_msr_bitmap;
vmx->msr_bitmap_mode = 0;
skip_msr_bitmap:
> -
> vmx->loaded_vmcs = &vmx->vmcs01;
> cpu = get_cpu();
> vmx_vcpu_load(&vmx->vcpu, cpu);
> --
> 2.19.1
>
