On Tue, Nov 03, 2020 at 12:27:56PM -0800, Yonghong Song wrote:
SNIP
> > > >
> > >
> > > bpf_map iter definition:
> > >
> > > DEFINE_BPF_ITER_FUNC(bpf_map, struct bpf_iter_meta *meta, struct bpf_map *map)
> > >
> > > goes to:
> > >
> > > #define DEFINE_BPF_ITER_FUNC(target, args...) \
> > > extern int bpf_iter_ ## target(args); \
> > > int __init bpf_iter_ ## target(args) { return 0; }
> > >
> > > that creates __init bpf_iter_bpf_map function that will make
> > > it into BTF where it's expected when opening iterator, but the
> > > code will be freed because it's __init function
> >
> > hm... should we just drop __init there?
> >
> > Yonghong, is __init strictly necessary, or was just an optimization to
> > save a tiny bit of space?
>
> It is an optimization to save some space. We only need function
> signature, not function body, for bpf_iter.
>
> The macro definition is in include/linux/bpf.h.
>
> #define DEFINE_BPF_ITER_FUNC(target, args...) \
> extern int bpf_iter_ ## target(args); \
> int __init bpf_iter_ ## target(args) { return 0; }
>
> Maybe you could have a section, e.g., called
> .init.bpf.preserve_type
> which you can scan through to preserve the types.
right, sounds good, will send v3 with that
thanks,
jirka
