Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

On Fri, 25 Sep 2020 00:03:14 +0200 Daniel Borkmann wrote:

static inline u64 gen_cookie_next(struct gen_cookie *gc)
{
         u64 val;

         if (likely(this_cpu_inc_return(*gc->level_nesting) == 1)) {

Is this_cpu_inc() in itself atomic?

Is there a comparison of performance of various atomic ops and locking
somewhere? I wonder how this scheme would compare to a using a cmpxchg.

                 u64 *local_last = this_cpu_ptr(gc->local_last);

                 val = *local_last;
                 if (__is_defined(CONFIG_SMP) &&
                     unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {

Can we reasonably assume we won't have more than 4k CPUs and just
statically divide this space by encoding CPU id in top bits?

                         s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
                                                        &gc->shared_last);
                         val = next - COOKIE_LOCAL_BATCH;
                 }
                 val++;
                 if (unlikely(!val))
                         val++;
                 *local_last = val;
         } else {
                 val = atomic64_add_return(COOKIE_LOCAL_BATCH,
                                           &gc->shared_last);
         }
         this_cpu_dec(*gc->level_nesting);
         return val;
}