On Fri, May 8, 2020 at 9:40 AM Alexei Starovoitov <ast@xxxxxx> wrote:
>
> On 5/8/20 12:05 AM, Andrii Nakryiko wrote:
> >
> > base : 9.200 ± 0.319M/s
> > fentry : 8.955 ± 0.241M/s
>
> > +SEC("fentry/__x64_sys_getpgid")
> > +int bench_trigger_fentry(void *ctx)
> > +{
> > + __sync_add_and_fetch(&hits, 1);
> > + return 0;
> > +}
>
> adding 'lock xadd' is not cheap.
> I wonder how much of the delta come from it and from the rest of
> trampoline.
Yeah, could be, though count-global/count-local benchmarks get to
150Mops/s under no-sharing scenario, so effect shouldn't be that high.
Well, in any case, 9Mops/s is more than enough for my cases :)
