On Tue, May 17, 2022, Aaron Lewis wrote:
> When returning from the compare function the u64 is truncated to an
> int. This results in a loss of the high nybble[1] in the event select
> and its sign if that nybble is in use. Switch from using a result that
> can end up being truncated to a result that can only be: 1, 0, -1.
>
> [1] bits 35:32 in the event select register and bits 11:8 in the event
> select.
>
> Fixes: 7ff775aca48ad ("KVM: x86/pmu: Use binary search to check filtered events")
> Signed-off-by: Aaron Lewis <aaronlewis@xxxxxxxxxx>
> ---
> arch/x86/kvm/pmu.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> index eca39f56c231..1666e9d3e545 100644
> --- a/arch/x86/kvm/pmu.c
> +++ b/arch/x86/kvm/pmu.c
> @@ -173,7 +173,7 @@ static bool pmc_resume_counter(struct kvm_pmc *pmc)
>
> static int cmp_u64(const void *a, const void *b)
> {
> - return *(__u64 *)a - *(__u64 *)b;
> + return (*(u64 *)a > *(u64 *)b) - (*(u64 *)a < *(u64 *)b);
On one hand, this is downright evil. On the other, it does generate branch-free
code, whereas gcc does not for explicit returns...
It's a little easier to read if the values are captured in local variables?
u64 l = *(u64 *)a;
u64 r = *(u64 *)b;
return (l > r) - (l < r);
Either way,
Reviewed-by: Sean Christopherson <seanjc@xxxxxxxxxx>
