On 6/3/24 20:14, Sean Christopherson wrote:
On Mon, Apr 29, 2024, Alejandro Jimenez wrote:
Even when APICv/AVIC is active, certain guest accesses to its local APIC(s)
cannot be fully accelerated, and cause a #VMEXIT to allow the VMM to
emulate the behavior and side effects. Expose a counter stat for these
specific #VMEXIT types.
Suggested-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
Signed-off-by: Alejandro Jimenez <alejandro.j.jimenez@xxxxxxxxxx>
---
arch/x86/include/asm/kvm_host.h | 1 +
arch/x86/kvm/svm/avic.c | 7 +++++++
arch/x86/kvm/vmx/vmx.c | 2 ++
arch/x86/kvm/x86.c | 1 +
4 files changed, 11 insertions(+)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index e7e3213cefae..388979dfe9f3 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1576,6 +1576,7 @@ struct kvm_vcpu_stat {
u64 guest_mode;
u64 notify_window_exits;
u64 apicv_active;
+ u64 apicv_unaccelerated_inj;
The stat name doesn't match the changelog or the code. The AVIC updates in
avic_incomplete_ipi_interception() are unaccelerated _injection_, they're
unaccelarated _delivery_. And in those cases, the fact that delivery wasn't
accelerated is relatively uninteresting in most cases.
Yeah, this was my flawed attempt to interpret/implement Paolo's comment in the RFC thread:
"... for example I'd add an interrupt_injections stat for unaccelerated injections causing a vmexit or otherwise hitting lapic.c"
so I incorrectly bundled together APIC accesses that result in #VMEXIT and end up requiring additional emulation (while managing to miss the handle_apic_access() case).
And avic_unaccelerated_access_interception() and handle_apic_write() don't
necessarily have anything to do with injection.
apicv_unaccelerated_acccess is perhaps a better name (assuming stat is updated in handle_apic_access() as well)?
On the flip side, the slow paths for {svm,vmx}_deliver_interrupt() are very
explicitly unnaccelerated injection.
Now that you highlight this, I think it might be closer to Paolo's idea. i.e. a stat for the slow path on these can be contrasted/compared with the kvm_apicv_accept_irq tracepoint that is hit on the fast path.
My initial reaction would be to update a stat for the fast path, as a confirmation that apicv is active which is how/why I typically use the kvm_apicv_accept_irq tracepoint, but that becomes redundant by having the apicv_active stat on PATCH 1.
So, if you don't think it is useful to have a general apicv_unaccelerated_acccess counter, I can drop this patch.
Thank you,
Alejandro
It's not entirely clear from the changelog what the end goal of this stat is.
A singular stat for all APICv/AVIC access VM-Exits seems uninteresting, as such
a stat essentially just captures that the guest is active. Maaaybe someone could
glean info from comparing two VMs, but even that is dubious. E.g. if a guest is
doing something function and generating a lot of avic_incomplete_ipi_interception()
exits, those will likely be in the noise due to the total volume of other AVIC
exits.
};
struct x86_instruction_info;
diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
index 4b74ea91f4e6..274041d3cf66 100644
--- a/arch/x86/kvm/svm/avic.c
+++ b/arch/x86/kvm/svm/avic.c
@@ -517,6 +517,8 @@ int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
kvm_apic_write_nodecode(vcpu, APIC_ICR);
else
kvm_apic_send_ipi(apic, icrl, icrh);
+
+ ++vcpu->stat.apicv_unaccelerated_inj;
break;
case AVIC_IPI_FAILURE_TARGET_NOT_RUNNING:
/*
@@ -525,6 +527,8 @@ int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
* vcpus. So, we just need to kick the appropriate vcpu.
*/
avic_kick_target_vcpus(vcpu->kvm, apic, icrl, icrh, index);
+
+ ++vcpu->stat.apicv_unaccelerated_inj;
break;
case AVIC_IPI_FAILURE_INVALID_BACKING_PAGE:
WARN_ONCE(1, "Invalid backing page\n");
@@ -704,6 +708,9 @@ int avic_unaccelerated_access_interception(struct kvm_vcpu *vcpu)
trace_kvm_avic_unaccelerated_access(vcpu->vcpu_id, offset,
trap, write, vector);
+
+ ++vcpu->stat.apicv_unaccelerated_inj;
+
if (trap) {
/* Handling Trap */
WARN_ONCE(!write, "svm: Handling trap read.\n");
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index f10b5f8f364b..a7487f12ded1 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -5657,6 +5657,8 @@ static int handle_apic_write(struct kvm_vcpu *vcpu)
{
unsigned long exit_qualification = vmx_get_exit_qual(vcpu);
+ ++vcpu->stat.apicv_unaccelerated_inj;
+
/*
* APIC-write VM-Exit is trap-like, KVM doesn't need to advance RIP and
* hardware has done any necessary aliasing, offset adjustments, etc...
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 03cb933920cb..c8730b0fac87 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -307,6 +307,7 @@ const struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
STATS_DESC_IBOOLEAN(VCPU, guest_mode),
STATS_DESC_COUNTER(VCPU, notify_window_exits),
STATS_DESC_IBOOLEAN(VCPU, apicv_active),
+ STATS_DESC_COUNTER(VCPU, apicv_unaccelerated_inj),
};
const struct kvm_stats_header kvm_vcpu_stats_header = {
--
2.39.3
