On Mon, Nov 26, 2018 at 04:47:31PM +0100, Vitaly Kuznetsov wrote:
> Turns out Hyper-V on KVM (as of 2016) will only use synthetic timers
> if direct mode is available. With direct mode we notify the guest by
> asserting APIC irq instead of sending a SynIC message.
>
> The implementation uses existing vec_bitmap for letting lapic code
> know that we're interested in the particular IRQ's EOI request. We assume
> that the same APIC irq won't be used by the guest for both direct mode
> stimer and as sint source (especially with AutoEOI semantics). It is
> unclear how things should be handled if that's not true.
>
> Direct mode is also somewhat less expensive; in my testing
> stimer_send_msg() takes not less than 1500 cpu cycles and
> stimer_notify_direct() can usually be done in 300-400. WS2016 without
> Hyper-V, however, always sticks to non-direct version.
>
> Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
> ---
> - Changes since v1: avoid open-coding stimer_mark_pending() in
> kvm_hv_synic_send_eoi() [Paolo Bonzini]
> ---
> arch/x86/kvm/hyperv.c | 67 +++++++++++++++++++++++++++++++++++-----
> arch/x86/kvm/trace.h | 10 +++---
> arch/x86/kvm/x86.c | 1 +
> include/uapi/linux/kvm.h | 1 +
> 4 files changed, 67 insertions(+), 12 deletions(-)
>
> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
> index eaec15c738df..9533133be566 100644
> --- a/arch/x86/kvm/hyperv.c
> +++ b/arch/x86/kvm/hyperv.c
> @@ -38,6 +38,9 @@
>
> #define KVM_HV_MAX_SPARSE_VCPU_SET_BITS DIV_ROUND_UP(KVM_MAX_VCPUS, 64)
>
> +static void stimer_mark_pending(struct kvm_vcpu_hv_stimer *stimer,
> + bool vcpu_kick);
> +
> static inline u64 synic_read_sint(struct kvm_vcpu_hv_synic *synic, int sint)
> {
> return atomic64_read(&synic->sint[sint]);
> @@ -53,8 +56,21 @@ static inline int synic_get_sint_vector(u64 sint_value)
> static bool synic_has_vector_connected(struct kvm_vcpu_hv_synic *synic,
> int vector)
> {
> + struct kvm_vcpu *vcpu = synic_to_vcpu(synic);
> + struct kvm_vcpu_hv *hv_vcpu = vcpu_to_hv_vcpu(vcpu);
> + struct kvm_vcpu_hv_stimer *stimer;
> int i;
>
> + for (i = 0; i < ARRAY_SIZE(hv_vcpu->stimer); i++) {
> + stimer = &hv_vcpu->stimer[i];
> + if (stimer->config.enable && stimer->config.direct_mode &&
> + stimer->config.apic_vector == vector)
> + return true;
> + }
> +
> + if (vector < HV_SYNIC_FIRST_VALID_VECTOR)
> + return false;
> +
> for (i = 0; i < ARRAY_SIZE(synic->sint); i++) {
> if (synic_get_sint_vector(synic_read_sint(synic, i)) == vector)
> return true;
> @@ -80,14 +96,14 @@ static bool synic_has_vector_auto_eoi(struct kvm_vcpu_hv_synic *synic,
> static void synic_update_vector(struct kvm_vcpu_hv_synic *synic,
> int vector)
> {
> - if (vector < HV_SYNIC_FIRST_VALID_VECTOR)
> - return;
> -
> if (synic_has_vector_connected(synic, vector))
> __set_bit(vector, synic->vec_bitmap);
> else
> __clear_bit(vector, synic->vec_bitmap);
>
> + if (vector < HV_SYNIC_FIRST_VALID_VECTOR)
> + return;
> +
Just noticed that the patch seems to assume that "direct" timers are
allowed to use any vectors including 0-15. I guess this is incorrect,
and instead stimer_set_config should error out on direct mode with a
vector less than HV_SYNIC_FIRST_VALID_VECTOR.
Roman.
