Re: [PATCH bpf-next v5 1/6] bpf: Introduce load-acquire and store-release instructions

[Peilin Ye]

On Mon, Mar 03, 2025 at 04:45:45PM -0800, Alexei Starovoitov wrote:
> > > >         switch (insn->imm) {
> > > > @@ -7780,6 +7813,24 @@ static int check_atomic(struct bpf_verifier_env *env, struct bpf_insn *insn)
> > > >         case BPF_XCHG:
> > > >         case BPF_CMPXCHG:
> > > >                 return check_atomic_rmw(env, insn);
> > > > +       case BPF_LOAD_ACQ:
> > > > +#ifndef CONFIG_64BIT
> > > > +               if (BPF_SIZE(insn->code) == BPF_DW) {
> > > > +                       verbose(env,
> > > > +                               "64-bit load-acquires are only supported on 64-bit arches\n");
> > > > +                       return -EOPNOTSUPP;
> > > > +               }
> > > > +#endif
> > >
> > > Your earlier proposal of:
> > > if (BPF_SIZE(insn->code) == BPF_DW && BITS_PER_LONG != 64) {
> > >
> > > was cleaner.
> > > Why did you pick ifndef ?
> >
> > Likely overthinking, but I wanted to avoid this check at all for 64-bit
> > arches, so it's just a little bit faster.  Should I change it back to
> > checking BITS_PER_LONG ?
> 
> In general #ifdef in .c is the last resort.
> We avoid it when possible.

Got it.

> In core.c we probably cannot, but here we can.
> So yes. please respin.

Sure!

> I bet the compiler will produce the exact same code.

Thanks,
Peilin Ye