* [PATCH 1/2] KVM: SVM: Use target APIC ID to complete AVIC IRQs when possible
2022-04-20 15:49 [PATCH V2 0/2] KVM: SVM: Optimize AVIC incomplete IPI #vmexit handling Suravee Suthikulpanit
@ 2022-04-20 15:49 ` Suravee Suthikulpanit
2022-04-20 15:49 ` [PATCH 2/2] KVM: SVM: Introduce trace point for the slow-path of avic_kic_target_vcpus Suravee Suthikulpanit
2022-04-20 16:42 ` [PATCH V2 0/2] KVM: SVM: Optimize AVIC incomplete IPI #vmexit handling Paolo Bonzini
2 siblings, 0 replies; 4+ messages in thread
From: Suravee Suthikulpanit @ 2022-04-20 15:49 UTC (permalink / raw)
To: linux-kernel, kvm
Cc: pbonzini, mlevitsk, seanjc, joro, jon.grimm, wei.huang2,
terry.bowman, Suravee Suthikulpanit
Currently, an AVIC-enabled VM suffers from performance bottleneck
when scaling to large number of vCPUs for I/O intensive workloads.
In such case, a vCPU often executes halt instruction to get into idle state
waiting for interrupts, in which KVM would de-schedule the vCPU from
physical CPU.
When AVIC HW tries to deliver interrupt to the halting vCPU, it would
result in AVIC incomplete IPI #vmexit to notify KVM to reschedule
the target vCPU into running state.
Investigation has shown the main hotspot is in the kvm_apic_match_dest()
in the following call stack where it tries to find target vCPUs
corresponding to the information in the ICRH/ICRL registers.
- handle_exit
- svm_invoke_exit_handler
- avic_incomplete_ipi_interception
- kvm_apic_match_dest
However, AVIC provides hints in the #vmexit info, which can be used to
retrieve the destination guest physical APIC ID.
In addition, since QEMU defines guest physical APIC ID to be the same as
vCPU ID, it can be used to quickly identify the target vCPU to deliver IPI,
and avoid the overhead from searching through all vCPUs to match the target
vCPU.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
---
arch/x86/kvm/svm/avic.c | 72 ++++++++++++++++++++++++++++++++++++++---
1 file changed, 68 insertions(+), 4 deletions(-)
diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
index 421619540ff9..c8b8a0cb02b0 100644
--- a/arch/x86/kvm/svm/avic.c
+++ b/arch/x86/kvm/svm/avic.c
@@ -285,11 +285,75 @@ void avic_ring_doorbell(struct kvm_vcpu *vcpu)
put_cpu();
}
-static void avic_kick_target_vcpus(struct kvm *kvm, struct kvm_lapic *source,
- u32 icrl, u32 icrh)
+/*
+ * A fast-path version of avic_kick_target_vcpus(), which attempts to match
+ * destination APIC ID to vCPU without looping through all vCPUs.
+ */
+static int avic_kick_target_vcpus_fast(struct kvm *kvm, struct kvm_lapic *source,
+ u32 icrl, u32 icrh, u32 index)
{
+ u32 dest, apic_id;
struct kvm_vcpu *vcpu;
+ int dest_mode = icrl & APIC_DEST_MASK;
+ int shorthand = icrl & APIC_SHORT_MASK;
+ struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
+ u32 *avic_logical_id_table = page_address(kvm_svm->avic_logical_id_table_page);
+
+ if (shorthand != APIC_DEST_NOSHORT)
+ return -EINVAL;
+
+ /*
+ * The AVIC incomplete IPI #vmexit info provides index into
+ * the physical APIC ID table, which can be used to derive
+ * guest physical APIC ID.
+ */
+ if (dest_mode == APIC_DEST_PHYSICAL) {
+ apic_id = index;
+ } else {
+ if (!apic_x2apic_mode(source)) {
+ /* For xAPIC logical mode, the index is for logical APIC table. */
+ apic_id = avic_logical_id_table[index] & 0x1ff;
+ } else {
+ return -EINVAL;
+ }
+ }
+
+ /*
+ * Assuming vcpu ID is the same as physical apic ID,
+ * and use it to retrieve the target vCPU.
+ */
+ vcpu = kvm_get_vcpu_by_id(kvm, apic_id);
+ if (!vcpu)
+ return -EINVAL;
+
+ if (apic_x2apic_mode(vcpu->arch.apic))
+ dest = icrh;
+ else
+ dest = GET_APIC_DEST_FIELD(icrh);
+
+ /*
+ * Try matching the destination APIC ID with the vCPU.
+ */
+ if (kvm_apic_match_dest(vcpu, source, shorthand, dest, dest_mode)) {
+ vcpu->arch.apic->irr_pending = true;
+ svm_complete_interrupt_delivery(vcpu,
+ icrl & APIC_MODE_MASK,
+ icrl & APIC_INT_LEVELTRIG,
+ icrl & APIC_VECTOR_MASK);
+ return 0;
+ }
+
+ return -EINVAL;
+}
+
+static void avic_kick_target_vcpus(struct kvm *kvm, struct kvm_lapic *source,
+ u32 icrl, u32 icrh, u32 index)
+{
unsigned long i;
+ struct kvm_vcpu *vcpu;
+
+ if (!avic_kick_target_vcpus_fast(kvm, source, icrl, icrh, index))
+ return;
/*
* Wake any target vCPUs that are blocking, i.e. waiting for a wake
@@ -316,7 +380,7 @@ int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
u32 icrh = svm->vmcb->control.exit_info_1 >> 32;
u32 icrl = svm->vmcb->control.exit_info_1;
u32 id = svm->vmcb->control.exit_info_2 >> 32;
- u32 index = svm->vmcb->control.exit_info_2 & 0xFF;
+ u32 index = svm->vmcb->control.exit_info_2 & 0x1FF;
struct kvm_lapic *apic = vcpu->arch.apic;
trace_kvm_avic_incomplete_ipi(vcpu->vcpu_id, icrh, icrl, id, index);
@@ -343,7 +407,7 @@ int avic_incomplete_ipi_interception(struct kvm_vcpu *vcpu)
* set the appropriate IRR bits on the valid target
* vcpus. So, we just need to kick the appropriate vcpu.
*/
- avic_kick_target_vcpus(vcpu->kvm, apic, icrl, icrh);
+ avic_kick_target_vcpus(vcpu->kvm, apic, icrl, icrh, index);
break;
case AVIC_IPI_FAILURE_INVALID_TARGET:
break;
--
2.25.1
^ permalink raw reply [flat|nested] 4+ messages in thread* [PATCH 2/2] KVM: SVM: Introduce trace point for the slow-path of avic_kic_target_vcpus
2022-04-20 15:49 [PATCH V2 0/2] KVM: SVM: Optimize AVIC incomplete IPI #vmexit handling Suravee Suthikulpanit
2022-04-20 15:49 ` [PATCH 1/2] KVM: SVM: Use target APIC ID to complete AVIC IRQs when possible Suravee Suthikulpanit
@ 2022-04-20 15:49 ` Suravee Suthikulpanit
2022-04-20 16:42 ` [PATCH V2 0/2] KVM: SVM: Optimize AVIC incomplete IPI #vmexit handling Paolo Bonzini
2 siblings, 0 replies; 4+ messages in thread
From: Suravee Suthikulpanit @ 2022-04-20 15:49 UTC (permalink / raw)
To: linux-kernel, kvm
Cc: pbonzini, mlevitsk, seanjc, joro, jon.grimm, wei.huang2,
terry.bowman, Suravee Suthikulpanit
This can help identify potential performance issues when handles
AVIC incomplete IPI due vCPU not running.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
---
arch/x86/kvm/svm/avic.c | 2 ++
arch/x86/kvm/trace.h | 20 ++++++++++++++++++++
arch/x86/kvm/x86.c | 1 +
3 files changed, 23 insertions(+)
diff --git a/arch/x86/kvm/svm/avic.c b/arch/x86/kvm/svm/avic.c
index c8b8a0cb02b0..c5b38b160180 100644
--- a/arch/x86/kvm/svm/avic.c
+++ b/arch/x86/kvm/svm/avic.c
@@ -355,6 +355,8 @@ static void avic_kick_target_vcpus(struct kvm *kvm, struct kvm_lapic *source,
if (!avic_kick_target_vcpus_fast(kvm, source, icrl, icrh, index))
return;
+ trace_kvm_avic_kick_vcpu_slowpath(icrh, icrl, index);
+
/*
* Wake any target vCPUs that are blocking, i.e. waiting for a wake
* event. There's no need to signal doorbells, as hardware has handled
diff --git a/arch/x86/kvm/trace.h b/arch/x86/kvm/trace.h
index e3a24b8f04be..de4762517569 100644
--- a/arch/x86/kvm/trace.h
+++ b/arch/x86/kvm/trace.h
@@ -1459,6 +1459,26 @@ TRACE_EVENT(kvm_avic_ga_log,
__entry->vmid, __entry->vcpuid)
);
+TRACE_EVENT(kvm_avic_kick_vcpu_slowpath,
+ TP_PROTO(u32 icrh, u32 icrl, u32 index),
+ TP_ARGS(icrh, icrl, index),
+
+ TP_STRUCT__entry(
+ __field(u32, icrh)
+ __field(u32, icrl)
+ __field(u32, index)
+ ),
+
+ TP_fast_assign(
+ __entry->icrh = icrh;
+ __entry->icrl = icrl;
+ __entry->index = index;
+ ),
+
+ TP_printk("icrh:icrl=%#08x:%08x, index=%u",
+ __entry->icrh, __entry->icrl, __entry->index)
+);
+
TRACE_EVENT(kvm_hv_timer_state,
TP_PROTO(unsigned int vcpu_id, unsigned int hv_timer_in_use),
TP_ARGS(vcpu_id, hv_timer_in_use),
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 547ba00ef64f..d90e4020e9b9 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -12980,6 +12980,7 @@ EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_pi_irte_update);
EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_unaccelerated_access);
EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_incomplete_ipi);
EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_ga_log);
+EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_kick_vcpu_slowpath);
EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_apicv_accept_irq);
EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_vmgexit_enter);
EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_vmgexit_exit);
--
2.25.1
^ permalink raw reply [flat|nested] 4+ messages in thread