Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
On 14/03/2016 10:41, Suravee Suthikulpanit wrote: >> >> +kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, >> APIC_IRR)); (I think that smp_mb here would make sense, even though we're fine now thanks to re-checking vcpu->mode in kvm_vcpu_kick. >>> >>> Right, though only a smp_mb__after_atomic() is required (which is a >>> compiler barrier). It is similarly required in vmx. >> >> True, kvm_lapic_set_vector uses a lock prefix. >> >> (I thought it behaves like atomic_set, which would require MFENCE for >> correct ordering here ... I don't like smp_mb__after_atomic much >> because of the discrepancy on some atomic operations.) >> > > So, should i just use smb_mb() in this case? You should use smp_mb__after_atomic(). Paolo
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
Hi,
On 03/09/2016 06:10 PM, Paolo Bonzini wrote:
+
>>>+ if (vcpu->mode == IN_GUEST_MODE) {
>>>+ wrmsrl(SVM_AVIC_DOORBELL,
>>>+ __default_cpu_present_to_apicid(vcpu->cpu));
>>>+ } else {
>>>+ kvm_vcpu_kick(vcpu);
>>>+ }
>> >
>And what about
> [...]
> else if (!vcpu->...->is_running)
>kvm_vcpu_kick(vcpu);
>
>?
>The kick isn't needed unless the VCPU is scheduled out.
>Or maybe just
> if (vcpu->...->is_running)
> wrmsrl()
> else
> kvm_vcpu_kick();
>?
>Which doesn't use the information we have on top AVIC, making our logic
>a bit simpler.
Yes, both of this should work. I like the latter.
Ok, I'll modify this to check the is_running bit of the AVIC Physical
APIC ID table entry to determine if we should kick vcpu.
Thanks,
Suravee
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
Hi
On 03/09/2016 11:00 PM, Radim Krčmář wrote:
2016-03-09 12:10+0100, Paolo Bonzini:
On 08/03/2016 22:54, Radim Krčmář wrote:
2016-03-07 16:36+0100, Paolo Bonzini:
On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
+static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
+{
+ struct vcpu_svm *svm = to_svm(vcpu);
+
+ kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
(I think that smp_mb here would make sense, even though we're fine now
thanks to re-checking vcpu->mode in kvm_vcpu_kick.
Right, though only a smp_mb__after_atomic() is required (which is a
compiler barrier). It is similarly required in vmx.
True, kvm_lapic_set_vector uses a lock prefix.
(I thought it behaves like atomic_set, which would require MFENCE for
correct ordering here ... I don't like smp_mb__after_atomic much
because of the discrepancy on some atomic operations.)
So, should i just use smb_mb() in this case?
Suravee.
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
Hi,
Please ignore this. I should have read the whole thread before reply :(
Sorry for spam.
Suravee
On 03/14/2016 12:25 PM, Suravee Suthikulpanit wrote:
On 03/07/2016 10:36 PM, Paolo Bonzini wrote:
On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
+static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
+{
+struct vcpu_svm *svm = to_svm(vcpu);
+
+kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
+
+if (vcpu->mode == IN_GUEST_MODE) {
+wrmsrl(SVM_AVIC_DOORBELL,
+ __default_cpu_present_to_apicid(vcpu->cpu));
+} else {
+kvm_vcpu_kick(vcpu);
+}
You also need to add
kvm_make_request(KVM_REQ_EVENT, vcpu);
before the "if", similar to vmx_deliver_posted_interrupt.
Paolo
Actually, I should only need that just before the kvm_cpu_kick(vcpu)
isn't it. I don't think we need it in the case when sending doorbell.
Thanks,
Suravee
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
On 03/07/2016 10:36 PM, Paolo Bonzini wrote:
On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
+static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
+{
+ struct vcpu_svm *svm = to_svm(vcpu);
+
+ kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
+
+ if (vcpu->mode == IN_GUEST_MODE) {
+ wrmsrl(SVM_AVIC_DOORBELL,
+ __default_cpu_present_to_apicid(vcpu->cpu));
+ } else {
+ kvm_vcpu_kick(vcpu);
+ }
You also need to add
kvm_make_request(KVM_REQ_EVENT, vcpu);
before the "if", similar to vmx_deliver_posted_interrupt.
Paolo
Actually, I should only need that just before the kvm_cpu_kick(vcpu)
isn't it. I don't think we need it in the case when sending doorbell.
Thanks,
Suravee
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
2016-03-09 12:10+0100, Paolo Bonzini:
> On 08/03/2016 22:54, Radim Krčmář wrote:
>> 2016-03-07 16:36+0100, Paolo Bonzini:
>>> On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
+static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
+{
+ struct vcpu_svm *svm = to_svm(vcpu);
+
+ kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
>>
>> (I think that smp_mb here would make sense, even though we're fine now
>> thanks to re-checking vcpu->mode in kvm_vcpu_kick.
>
> Right, though only a smp_mb__after_atomic() is required (which is a
> compiler barrier). It is similarly required in vmx.
True, kvm_lapic_set_vector uses a lock prefix.
(I thought it behaves like atomic_set, which would require MFENCE for
correct ordering here ... I don't like smp_mb__after_atomic much
because of the discrepancy on some atomic operations.)
> Offtopic remark, kvm_make_request() and kvm_check_request() should
> probably have a smp_wmb() and a smp_rmb().
Yeah, noted.
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
On 08/03/2016 22:54, Radim Krčmář wrote:
> 2016-03-07 16:36+0100, Paolo Bonzini:
>> On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
>>> +static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
>>> +{
>>> + struct vcpu_svm *svm = to_svm(vcpu);
>>> +
>>> + kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
>
> (I think that smp_mb here would make sense, even though we're fine now
> thanks to re-checking vcpu->mode in kvm_vcpu_kick.
Right, though only a smp_mb__after_atomic() is required (which is a
compiler barrier). It is similarly required in vmx.
Offtopic remark, kvm_make_request() and kvm_check_request() should
probably have a smp_wmb() and a smp_rmb().
>>> +
>>> + if (vcpu->mode == IN_GUEST_MODE) {
>>> + wrmsrl(SVM_AVIC_DOORBELL,
>>> + __default_cpu_present_to_apicid(vcpu->cpu));
>>> + } else {
>>> + kvm_vcpu_kick(vcpu);
>>> + }
>>
>> You also need to add
>>
>> kvm_make_request(KVM_REQ_EVENT, vcpu);
>>
>> before the "if", similar to vmx_deliver_posted_interrupt.
>
> KVM won't do anything in KVM_REQ_EVENT and I think that the request can
> be avoided because KVM already has to handle IRR writes from AVIC.
Doh, I've made the same remark already and you've given the same answer. :)
> And what about
> [...]
> else if (!vcpu->...->is_running)
> kvm_vcpu_kick(vcpu);
>
> ?
> The kick isn't needed unless the VCPU is scheduled out.
> Or maybe just
> if (vcpu->...->is_running)
> wrmsrl()
> else
> kvm_vcpu_kick();
> ?
> Which doesn't use the information we have on top AVIC, making our logic
> a bit simpler.
Yes, both of this should work. I like the latter.
In any case, the smp_mb__after_atomic() is necessary.
Paolo
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
2016-03-07 16:36+0100, Paolo Bonzini:
> On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
>> +static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
>> +{
>> +struct vcpu_svm *svm = to_svm(vcpu);
>> +
>> +kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
(I think that smp_mb here would make sense, even though we're fine now
thanks to re-checking vcpu->mode in kvm_vcpu_kick.
A comment explaining this optimization would be nice. I'm thinking
about a race where we don't send the doorbell even though the VCPU is
in guest mode, because vcpu->mode was read before writing APIC_IRR.)
>> +
>> +if (vcpu->mode == IN_GUEST_MODE) {
>> +wrmsrl(SVM_AVIC_DOORBELL,
>> + __default_cpu_present_to_apicid(vcpu->cpu));
>> +} else {
>> +kvm_vcpu_kick(vcpu);
>> +}
>
> You also need to add
>
> kvm_make_request(KVM_REQ_EVENT, vcpu);
>
> before the "if", similar to vmx_deliver_posted_interrupt.
KVM won't do anything in KVM_REQ_EVENT and I think that the request can
be avoided because KVM already has to handle IRR writes from AVIC.
And what about
[...]
else if (!vcpu->...->is_running)
kvm_vcpu_kick(vcpu);
?
The kick isn't needed unless the VCPU is scheduled out.
Or maybe just
if (vcpu->...->is_running)
wrmsrl()
else
kvm_vcpu_kick();
?
Which doesn't use the information we have on top AVIC, making our logic
a bit simpler.
Re: [PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
> +static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
> +{
> + struct vcpu_svm *svm = to_svm(vcpu);
> +
> + kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
> +
> + if (vcpu->mode == IN_GUEST_MODE) {
> + wrmsrl(SVM_AVIC_DOORBELL,
> +__default_cpu_present_to_apicid(vcpu->cpu));
> + } else {
> + kvm_vcpu_kick(vcpu);
> + }
You also need to add
kvm_make_request(KVM_REQ_EVENT, vcpu);
before the "if", similar to vmx_deliver_posted_interrupt.
Paolo
[PART1 RFC v2 06/10] svm: Add interrupt injection via AVIC
This patch introduces a new mechanism to inject interrupt using AVIC.
Since VINTR is not supported when enable AVIC, we need to inject
interrupt via APIC backing page instead.
This patch also adds support for AVIC doorbell, which is used by
KVM to signal a running vcpu to check IRR for injected interrupts.
Signed-off-by: Suravee Suthikulpanit
---
arch/x86/kvm/svm.c | 35 ---
1 file changed, 32 insertions(+), 3 deletions(-)
diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
index c703149..8f11200 100644
--- a/arch/x86/kvm/svm.c
+++ b/arch/x86/kvm/svm.c
@@ -71,6 +71,8 @@ MODULE_DEVICE_TABLE(x86cpu, svm_cpu_id);
#define SVM_FEATURE_DECODE_ASSIST (1 << 7)
#define SVM_FEATURE_PAUSE_FILTER (1 << 10)
+#define SVM_AVIC_DOORBELL 0xc001011b
+
#define NESTED_EXIT_HOST 0 /* Exit handled on host level */
#define NESTED_EXIT_DONE 1 /* Exit caused nested vmexit */
#define NESTED_EXIT_CONTINUE 2 /* Further checks needed */
@@ -236,6 +238,8 @@ static bool npt_enabled = true;
static bool npt_enabled;
#endif
+static struct kvm_x86_ops svm_x86_ops;
+
/* allow nested paging (virtualized MMU) for all guests */
static int npt = true;
module_param(npt, int, S_IRUGO);
@@ -978,6 +982,8 @@ static __init int svm_hardware_setup(void)
if (avic) {
printk(KERN_INFO "kvm: AVIC enabled\n");
+ } else {
+ svm_x86_ops.deliver_posted_interrupt = NULL;
}
return 0;
@@ -3071,8 +3077,10 @@ static int clgi_interception(struct vcpu_svm *svm)
disable_gif(svm);
/* After a CLGI no interrupts should come */
- svm_clear_vintr(svm);
- svm->vmcb->control.v_irq = 0;
+ if (!avic) {
+ svm_clear_vintr(svm);
+ svm->vmcb->control.v_irq = 0;
+ }
mark_dirty(svm->vmcb, VMCB_INTR);
@@ -3647,6 +3655,9 @@ static int msr_interception(struct vcpu_svm *svm)
static int interrupt_window_interception(struct vcpu_svm *svm)
{
+ if (avic)
+ BUG();
+
kvm_make_request(KVM_REQ_EVENT, &svm->vcpu);
svm_clear_vintr(svm);
svm->vmcb->control.v_irq = 0;
@@ -3961,7 +3972,7 @@ static inline void svm_inject_irq(struct vcpu_svm *svm,
int irq)
{
struct vmcb_control_area *control;
-
+ /* The following fields are ignored when AVIC is enabled */
control = &svm->vmcb->control;
control->int_vector = irq;
control->v_intr_prio = 0xf;
@@ -4042,6 +4053,20 @@ static void svm_sync_pir_to_irr(struct kvm_vcpu *vcpu)
return;
}
+static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
+{
+ struct vcpu_svm *svm = to_svm(vcpu);
+
+ kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
+
+ if (vcpu->mode == IN_GUEST_MODE) {
+ wrmsrl(SVM_AVIC_DOORBELL,
+ __default_cpu_present_to_apicid(vcpu->cpu));
+ } else {
+ kvm_vcpu_kick(vcpu);
+ }
+}
+
static int svm_nmi_allowed(struct kvm_vcpu *vcpu)
{
struct vcpu_svm *svm = to_svm(vcpu);
@@ -4102,6 +4127,9 @@ static void enable_irq_window(struct kvm_vcpu *vcpu)
* get that intercept, this function will be called again though and
* we'll get the vintr intercept.
*/
+ if (avic)
+ return;
+
if (gif_set(svm) && nested_svm_intr(svm)) {
svm_set_vintr(svm);
svm_inject_irq(svm, 0x0);
@@ -4834,6 +4862,7 @@ static struct kvm_x86_ops svm_x86_ops = {
.sched_in = svm_sched_in,
.pmu_ops = &amd_pmu_ops,
+ .deliver_posted_interrupt = svm_deliver_avic_intr,
};
static int __init svm_init(void)
--
1.9.1
