On Tue, May 18, 2021, Wanpeng Li wrote: > From: Wanpeng Li <wanpengli@xxxxxxxxxxx> > > In case of under-committed scenarios, vCPU can get scheduling easily, > kvm_vcpu_yield_to add extra overhead, we can observe a lot of race > between vcpu->ready is true and yield fails due to p->state is > TASK_RUNNING. Let's bail out in such scenarios by checking the length > of current cpu runqueue, it can be treated as a hint of under-committed > instead of guarantee of accuracy. The directed_yield_successful/attempted > ratio can be improved from 50+% to 80+% in the under-committed scenario. The "50+% to 80+%" comment will be a bit confusing for future readers now that the single_task_running() case counts as an attempt. I think the new comment would be something like "30%+ of directed-yield attempts can avoid the expensive lookups in kvm_sched_yield() in an under-committed scenario." That would also provide the real justification, as bumping the success ratio isn't the true goal of this path. > Signed-off-by: Wanpeng Li <wanpengli@xxxxxxxxxxx> > --- > v2 -> v3: > * update patch description > v1 -> v2: > * move the check after attempted counting > * update patch description > > arch/x86/kvm/x86.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > index 9b6bca616929..dfb7c320581f 100644 > --- a/arch/x86/kvm/x86.c > +++ b/arch/x86/kvm/x86.c > @@ -8360,6 +8360,9 @@ static void kvm_sched_yield(struct kvm_vcpu *vcpu, unsigned long dest_id) > > vcpu->stat.directed_yield_attempted++; > > + if (single_task_running()) > + goto no_yield; > + > rcu_read_lock(); > map = rcu_dereference(vcpu->kvm->arch.apic_map); > > -- > 2.25.1 >
