On Tue, Apr 11, 2023, Alexey Kardashevskiy wrote:
> SVM/SEV enable debug registers intercepts to skip swapping DRs
> on entering/exiting the guest. When the guest is in control of
> debug registers (vcpu->guest_debug == 0), there is an optimisation to
> reduce the number of context switches: intercepts are cleared and
> the KVM_DEBUGREG_WONT_EXIT flag is set to tell KVM to do swapping
> on guest enter/exit.
>
> The same code also executes for SEV-ES, however it has no effect as
> - it always takes (vcpu->guest_debug == 0) branch;
> - KVM_DEBUGREG_WONT_EXIT is set but DR7 intercept is not cleared;
> - vcpu_enter_guest() writes DRs but VMRUN for SEV-ES swaps them
> with the values from _encrypted_ VMSA.
>
> Be explicit about SEV-ES not supporting debug:
> - return right away from dr_interception() and skip unnecessary processing;
> - clear vcpu->guest_debug at SEV-ES' LAUNCH_UPDATE_VMSA if debugging
> was already enabled; after that point the generic x86's
> KVM_SET_GUEST_DEBUG ioctl disallows enabling debug.
>
> Add WARN_ON to kvm_x86::sync_dirty_debug_regs() (saves guest DRs on
> guest exit) to signify that SEV-ES won't hit that path.
>
> Suggested-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> Signed-off-by: Alexey Kardashevskiy <aik@xxxxxxx>
> Reviewed-by: Tom Lendacky <thomas.lendacky@xxxxxxx>
> ---
> Changes:
> v5:
> * new in the series
> ---
> arch/x86/kvm/svm/sev.c | 6 ++++++
> arch/x86/kvm/svm/svm.c | 10 +++++++++-
> 2 files changed, 15 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
> index 0f4761a57d86..b4365622222b 100644
> --- a/arch/x86/kvm/svm/sev.c
> +++ b/arch/x86/kvm/svm/sev.c
> @@ -639,6 +639,12 @@ static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
> return ret;
>
> vcpu->arch.guest_state_protected = true;
> +
> + if (vcpu->guest_debug)
> + pr_warn_ratelimited("guest_debug (%lx) not supported for SEV-ES",
Note, this needs a newline in the printk, otherwise it'll get buffered until
the next non-cont printk comes along (guess how many times I've been burned by
this).
> + vcpu->guest_debug);
> + vcpu->guest_debug = 0;
Argh, KVM's APIs can be quite frustrating. IIUC, guest_debug can never actually
be consumed because, per Tom[*], "A guest can't run before the LAUNCH_UPDATE process
is complete". But because the fact that the VM is an SEV-ES is communicated to
KVM after KVM_CREATE_VM, userspace can do KVM_SET_GUEST_DEBUG before KVM_SEV_ES_INIT
and before KVM_SEV_LAUNCH_UPDATE_VMSA, and thus get KVM into a state where
guest_debug is non-zero for an SEV-ES guest. Blech.
Instead of a ratelimited warn, can KVM get away with simply rejecting
KVM_SEV_LAUNCH_UPDATE_VMSA if guest_debug is non-zero? That combo can't work,
so it's seems unlikely userspace is relying on being able to do KVM_SET_GUEST_DEBUG.
If we do "have" to keep this approach, I'm generally opposed to any kind of printk
in KVM, but this one does seem to be justified since the most likely scenario is
that there's a human interactively debugging the guest (or at least, trying to
debug the guest). But I would say explicitly call out the ioctl(), "guest_debug"
probably won't mean anything to a random user. And I vote to not print the value,
that implies that the specific value is unsupported, not that debug in general is
disallowed.
Something like this (if we have to)?
pr_warn_ratelimited("Suppressing KVM_SET_GUEST_DEBUG for SEV-ES guest\n"
[*] https://lore.kernel.org/all/7edcf2c3-005f-04bd-7ec6-80baee236f40@xxxxxxx
> +
> return 0;
> }
>
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index dc12de325cca..179952a31d3b 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -1980,7 +1980,7 @@ static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
> {
> struct vcpu_svm *svm = to_svm(vcpu);
>
> - if (vcpu->arch.guest_state_protected)
> + if (WARN_ON_ONCE(sev_es_guest(vcpu->kvm)))
> return;
>
> get_debugreg(vcpu->arch.db[0], 0);
> @@ -2698,6 +2698,14 @@ static int dr_interception(struct kvm_vcpu *vcpu)
> unsigned long val;
> int err = 0;
>
> + /*
> + * SEV-ES intercepts DR7 only to disable guest debugging
> + * and the guest issues a VMGEXIT for DR7 write only. KVM cannot
Wrapping is a bit aggressive (wrap at 80, not earlier).
> + * change DR7 (always swapped as type 'A') so return early.
> + */
> + if (sev_es_guest(vcpu->kvm))
> + return 1;
> +
> if (vcpu->guest_debug == 0) {
> /*
> * No more DR vmexits; force a reload of the debug registers
> --
> 2.39.1
>
