Re: [PATCH v8 4/4] kvm: vmx: virtualize split lock detection

From: Thomas Gleixner
Date: Wed Apr 15 2020 - 13:43:53 EST


Xiaoyao Li <xiaoyao.li@xxxxxxxxx> writes:
> +/*
> + * Note: for guest, feature split lock detection can only be enumerated through
> + * MSR_IA32_CORE_CAPABILITIES bit. The FMS enumeration is unsupported.

That comment is confusing at best.

> + */
> +static inline bool guest_cpu_has_feature_sld(struct kvm_vcpu *vcpu)
> +{
> + return vcpu->arch.core_capabilities &
> + MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT;
> +}
> +
> +static inline bool guest_cpu_sld_on(struct vcpu_vmx *vmx)
> +{
> + return vmx->msr_test_ctrl & MSR_TEST_CTRL_SPLIT_LOCK_DETECT;
> +}
> +
> +static inline void vmx_update_sld(struct kvm_vcpu *vcpu, bool on)
> +{
> + /*
> + * Toggle SLD if the guest wants it enabled but its been disabled for
> + * the userspace VMM, and vice versa. Note, TIF_SLD is true if SLD has
> + * been turned off. Yes, it's a terrible name.

Instead of writing that useless blurb you could have written a patch
which changes TIF_SLD to TIF_SLD_OFF to make it clear.

> + */
> + if (sld_state == sld_warn && guest_cpu_has_feature_sld(vcpu) &&
> + on == test_thread_flag(TIF_SLD)) {
> + sld_update_msr(on);
> + update_thread_flag(TIF_SLD, !on);

Of course you completely fail to explain why TIF_SLD needs to be fiddled
with.

> @@ -1188,6 +1217,10 @@ void vmx_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
> #endif
>
> vmx_set_host_fs_gs(host_state, fs_sel, gs_sel, fs_base, gs_base);
> +
> + vmx->host_sld_on = !test_thread_flag(TIF_SLD);

This inverted storage is non-intuitive. What's wrong with simply
reflecting the TIF_SLD state?

> + vmx_update_sld(vcpu, guest_cpu_sld_on(vmx));
> +
> vmx->guest_state_loaded = true;
> }
>
> @@ -1226,6 +1259,9 @@ static void vmx_prepare_switch_to_host(struct vcpu_vmx *vmx)
> wrmsrl(MSR_KERNEL_GS_BASE, vmx->msr_host_kernel_gs_base);
> #endif
> load_fixmap_gdt(raw_smp_processor_id());
> +
> + vmx_update_sld(&vmx->vcpu, vmx->host_sld_on);
> +

vmx_prepare_switch_to_guest() is called via:

kvm_arch_vcpu_ioctl_run()
vcpu_run()
vcpu_enter_guest()
preempt_disable();
kvm_x86_ops.prepare_guest_switch(vcpu);

but vmx_prepare_switch_to_host() is invoked at the very end of:

kvm_arch_vcpu_ioctl_run()
.....
vcpu_run()
.....
vcpu_put()
vmx_vcpu_put()
vmx_prepare_switch_to_host();

That asymmetry does not make any sense without an explanation.

What's even worse is that vmx_prepare_switch_to_host() is invoked with
preemption enabled, so MSR state and TIF_SLD state can get out of sync
on preemption/migration.

> @@ -1946,9 +1992,15 @@ static int vmx_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>
> switch (msr_index) {
> case MSR_TEST_CTRL:
> - if (data)
> + if (data & ~vmx_msr_test_ctrl_valid_bits(vcpu))
> return 1;
>
> + vmx->msr_test_ctrl = data;
> +
> + preempt_disable();

This preempt_disable/enable() lacks explanation as well.

> + if (vmx->guest_state_loaded)
> + vmx_update_sld(vcpu, guest_cpu_sld_on(vmx));
> + preempt_enable();

How is updating msr_test_ctrl valid if this is invoked from the IOCTL,
i.e. host_initiated == true?

That said, I also hate the fact that you export both the low level MSR
function _and_ the state variable. Having all these details including the
TIF mangling in the VMX code is just wrong.

Thanks,

tglx