On Sat, Apr 02, 2022, Li,Rongqing wrote:
> > 发件人: Paolo Bonzini <paolo.bonzini@xxxxxxxxx> 代表 Paolo Bonzini
> > On 4/2/22 06:01, Li RongQing wrote:
> > > pi_wakeup_handler is used to wakeup the sleep vCPUs by posted irq
> > > list_for_each_entry is used in it, and whose input is other function
> > > per_cpu(), That cause that per_cpu() be invoked at least twice when
> > > there is one sleep vCPU
> > >
> > > so optimize pi_wakeup_handler it by reading once which is safe in
> > > spinlock protection
There's no need to protect reading the per-cpu variable with the spinlock, only
walking the list needs to be protected. E.g. the code can be compacted to:
int cpu = smp_processor_id();
raw_spinlock_t *spinlock = &per_cpu(wakeup_vcpus_on_cpu_lock, cpu);
struct list_head *wakeup_list = &per_cpu(wakeup_vcpus_on_cpu, cpu);
struct vcpu_vmx *vmx;
raw_spin_lock(spinlock);
list_for_each_entry(vmx, wakeup_list, pi_wakeup_list) {
if (pi_test_on(&vmx->pi_desc))
kvm_vcpu_wake_up(&vmx->vcpu);
}
raw_spin_unlock(spinlock);
> > >
> > > and same to per CPU spinlock
> >
> > What's the difference in the generated code?
> >
>
> This reduces one fifth asm codes
...
> these is a similar patch 031e3bd8986fffe31e1ddbf5264cccfe30c9abd7
Is there a measurable performance improvement though? I don't dislike the patch,
but it probably falls into the "technically an optimization but no one will ever
notice" category.
