On Tue, Jul 19, 2022 at 1:21 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Mon, Jul 18, 2022 at 12:07 PM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> >
> > Motivation:
> >
> > Our bpf programs have a bunch of options which are set at the loading
> > time. After loading, they don't change. We currently use array map
> > to store them and bpf program does the following:
> >
> > val = bpf_map_lookup_elem(&config_map, &key);
> > if (likely(val && *val)) {
> > // do some optional feature
> > }
> >
> > Since the configuration is static and we have a lot of those features,
> > I feel like we're wasting precious cycles doing dynamic lookups
> > (and stalling on memory loads).
> >
> > I was assuming that converting those to some fake kconfig options
> > would solve it, but it still seems like kconfig is stored in the
> > global map and kconfig entries are resolved dynamically.
> >
> > Proposal:
> >
> > Resolve kconfig options statically upon loading. Basically rewrite
> > ld+ldx to two nops and 'mov val, x'.
> >
> > I'm also trying to rewrite conditional jump when the condition is
> > !imm. This seems to be catching all the cases in my program, but
> > it's probably too hacky.
> >
> > I've attached very raw RFC patch to demonstrate the idea. Anything
> > I'm missing? Any potential problems with this approach?
>
> Have you considered using global variables for that?
> With skeleton the user space has a natural way to set
> all of these knobs after doing skel_open and before skel_load.
> Then the verifier sees them as readonly vars and
> automatically converts LDX into fixed constants and if the code
> looks like if (my_config_var) then the verifier will remove
> all the dead code too.
Hm, that's a good alternative, let me try it out. Thanks!
