Re: [PATCH v2 14/15] kvm: ioapic: Delay update IOAPIC EOI for RTC

[Suthikulpanit, Suravee]

Alex,

On 8/19/19 6:00 AM, Alexander Graf wrote:
> 
> 
> On 15.08.19 18:25, Suthikulpanit, Suravee wrote:
>> In-kernel IOAPIC does not update RTC pending EOI info with AMD SVM /w 
>> AVIC
>> when interrupt is delivered as edge-triggered since AMD processors
>> cannot exit on EOI for these interrupts.
>>
>> Add code to also check LAPIC pending EOI before injecting any new RTC
>> interrupts on AMD SVM when AVIC is activated.
>>
>> Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@xxxxxxx>
>> ---
>> Â arch/x86/kvm/ioapic.c | 36 ++++++++++++++++++++++++++++++++----
>> Â 1 file changed, 32 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/x86/kvm/ioapic.c b/arch/x86/kvm/ioapic.c
>> index 1add1bc..45e7bb0 100644
>> --- a/arch/x86/kvm/ioapic.c
>> +++ b/arch/x86/kvm/ioapic.c
>> @@ -39,6 +39,7 @@
>> Â #include <asm/processor.h>
>> Â #include <asm/page.h>
>> Â #include <asm/current.h>
>> +#include <asm/virtext.h>
>> Â #include <trace/events/kvm.h>
>> Â #include "ioapic.h"
>> @@ -173,6 +174,7 @@ static bool rtc_irq_check_coalesced(struct 
>> kvm_ioapic *ioapic)
>> ÂÂÂÂÂ return false;
>> Â }
>> +#define APIC_DEST_NOSHORTÂÂÂÂÂÂÂ 0x0
>> Â static int ioapic_set_irq(struct kvm_ioapic *ioapic, unsigned int irq,
>> ÂÂÂÂÂÂÂÂÂ int irq_level, bool line_status)
>> Â {
>> @@ -201,10 +203,36 @@ static int ioapic_set_irq(struct kvm_ioapic 
>> *ioapic, unsigned int irq,
>>  * interrupts lead to time drift in Windows guests. So we track
>> ÂÂÂÂÂÂ * EOI manually for the RTC interrupt.
>> ÂÂÂÂÂÂ */
>> -ÂÂÂ if (irq == RTC_GSI && line_status &&
>> -ÂÂÂÂÂÂÂ rtc_irq_check_coalesced(ioapic)) {
>> -ÂÂÂÂÂÂÂ ret = 0;
>> -ÂÂÂÂÂÂÂ goto out;
>> +ÂÂÂ if (irq == RTC_GSI && line_status) {
>> +ÂÂÂÂÂÂÂ struct kvm *kvm = ioapic->kvm;
>> +ÂÂÂÂÂÂÂ union kvm_ioapic_redirect_entry *entry = &ioapic->redirtbl[irq];
>> +
>> +ÂÂÂÂÂÂÂ /*
>> +ÂÂÂÂÂÂÂÂ * Since, AMD SVM AVIC accelerates write access to APIC EOI
>> +ÂÂÂÂÂÂÂÂ * register for edge-trigger interrupts, IOAPIC will not be
>> +ÂÂÂÂÂÂÂÂ * able to receive the EOI. In this case, we do lazy update
>> +ÂÂÂÂÂÂÂÂ * of the pending EOI when trying to set IOAPIC irq for RTC.
>> +ÂÂÂÂÂÂÂÂ */
>> +ÂÂÂÂÂÂÂ if (cpu_has_svm(NULL) &&
>> +ÂÂÂÂÂÂÂÂÂÂÂ (kvm->arch.apicv_state == APICV_ACTIVATED) &&
>> +ÂÂÂÂÂÂÂÂÂÂÂ (entry->fields.trig_mode == IOAPIC_EDGE_TRIG)) {
>> +ÂÂÂÂÂÂÂÂÂÂÂ int i;
>> +ÂÂÂÂÂÂÂÂÂÂÂ struct kvm_vcpu *vcpu;
>> +
>> +ÂÂÂÂÂÂÂÂÂÂÂ kvm_for_each_vcpu(i, vcpu, kvm)
>> +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ if (kvm_apic_match_dest(vcpu, NULL,
>> +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ KVM_APIC_DEST_NOSHORT,
>> +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ entry->fields.dest_id,
>> +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ entry->fields.dest_mode)) {
>> +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ __rtc_irq_eoi_tracking_restore_one(vcpu);
> 
> I don't understand why this works. This code just means we're injecting 
> an EOI on the first CPU that has the vector mapped, right when we're 
> setting it to trigger, no?

Actually, this is similar to the approach in patch 12/15, where we check 
if there is any pending EOI for the RTC_GSI on the destination vcpu. 
Otherwise, the __rtc_irq_eoi_tracking_restore_one() should clear IOAPIC 
pending EOI for RTC.

Suravee