On Mon, Apr 18, 2022, Jon Kohler wrote:
> On vmx_vcpu_load_vmcs and svm_vcpu_load, respect user controlled
> configuration for conditional IBPB and only attempt IBPB MSR when
> switching between different guest vCPUs IFF switch_mm_always_ibpb,
> which fixes a situation where the kernel will issue IBPB
> unconditionally even when conditional IBPB is enabled.
>
> If a user has spectre_v2_user mitigation enabled, in any
> configuration, and the underlying processor supports X86_FEATURE_IBPB,
> X86_FEATURE_USE_IBPB is set and any calls to
> indirect_branch_prediction_barrier() will issue IBPB MSR.
>
> Depending on the spectre_v2_user configuration, either
> switch_mm_always_ibpb key or switch_mm_cond_ibpb key will be set.
>
> Both switch_mm_always_ibpb and switch_mm_cond_ibpb are handled by
> switch_mm() -> cond_mitigation(), which works well in cases where
> switching vCPUs (i.e. switching tasks) also switches mm_struct;
> however, this misses a paranoid case where user space may be running
> multiple guests in a single process (i.e. single mm_struct).
>
> This paranoid case is already covered by vmx_vcpu_load_vmcs and
> svm_vcpu_load; however, this is done by calling
> indirect_branch_prediction_barrier() and thus the kernel
> unconditionally issues IBPB if X86_FEATURE_USE_IBPB is set.
The changelog should call out that switch_mm_cond_ibpb is intentionally "ignored"
for the virt case, and explain why it's nonsensical to emit IBPB in that scenario.
> Fix by using intermediary call to x86_virt_guest_switch_ibpb(), which
> gates IBPB MSR IFF switch_mm_always_ibpb is true. This is useful for
> security paranoid VMMs in either single process or multi-process VMM
> configurations.
Multi-process VMM? KVM doesn't allow "sharing" a VM across processes. Userspace
can share guest memory across processes, but that's not relevant to an IBPB on
guest switch. I suspect you're loosely referring to all of userspace as a single
VMM. That's inaccurate, or at least unnecessarily confusing, from a kernel
perspective. I am not aware of a VMM that runs as a monolithic "daemon" and forks
a new process for every VM. And even in such a case, I would argue that most
people would refer to each process as a separate VMM.
If there's a blurb about the switch_mm_cond_ibpb case being nonsensical, there's
probably a good segue into stating the new behavior.
> switch_mm_always_ibpb key is user controlled via spectre_v2_user and
> will be true for the following configurations:
> spectre_v2_user=on
> spectre_v2_user=prctl,ibpb
> spectre_v2_user=seccomp,ibpb
>
> Signed-off-by: Jon Kohler <jon@xxxxxxxxxxx>
> Cc: Sean Christopherson <seanjc@xxxxxxxxxx>
> Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> Cc: Kees Cook <keescook@xxxxxxxxxxxx>
> Cc: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
> Cc: Waiman Long <longman@xxxxxxxxxx>
> ---
> v1 -> v2:
> - Addressed comments on approach from Sean.
>
> arch/x86/include/asm/spec-ctrl.h | 15 +++++++++++++++
> arch/x86/kernel/cpu/bugs.c | 6 +++++-
> arch/x86/kvm/svm/svm.c | 2 +-
> arch/x86/kvm/vmx/vmx.c | 2 +-
> 4 files changed, 22 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/include/asm/spec-ctrl.h b/arch/x86/include/asm/spec-ctrl.h
> index 5393babc0598..1ad140b17ad7 100644
> --- a/arch/x86/include/asm/spec-ctrl.h
> +++ b/arch/x86/include/asm/spec-ctrl.h
> @@ -85,4 +85,19 @@ static inline void speculative_store_bypass_ht_init(void) { }
> extern void speculation_ctrl_update(unsigned long tif);
> extern void speculation_ctrl_update_current(void);
>
> +/*
> + * Issue IBPB when switching guest vCPUs IFF if switch_mm_always_ibpb.
Extra "if" there.
> + * Primarily useful for security paranoid (or naive) user space VMMs
> + * that may run multiple VMs within a single process.
> + * For multi-process VMMs, switching vCPUs, i.e. switching tasks,
As above, "multi-process VMMs" is very confusing, they're really just separate VMMs.
Something like this?
* For the more common case of running VMs in their own dedicated process,
* switching vCPUs that belong to different VMs, i.e. switching tasks, will also
* ...
> + * will also switch mm_structs and thus do IPBP via cond_mitigation();
> + * however, in the always_ibpb case, take a paranoid approach and issue
> + * IBPB on both switch_mm() and vCPU switch.
> + */
> +static inline void x86_virt_guest_switch_ibpb(void)
> +{
> + if (static_branch_unlikely(&switch_mm_always_ibpb))
> + indirect_branch_prediction_barrier();
> +}
> +
> #endif
> diff --git a/arch/x86/kernel/cpu/bugs.c b/arch/x86/kernel/cpu/bugs.c
> index 6296e1ebed1d..6aafb0279cbc 100644
