On Fri, 2023-12-08 at 09:46 -0800, Alexei Starovoitov wrote: [...] > > I just want to propose to have less work :-) since we are only dealing > > with a few structs in bpf domain. Note that eventually generated > > vmlinux.h will be the same whether we use 'hard code' approach or the > > decl_tag approach. The difference is just how to do it: - dwarf/btf with > > decl tag -> bpftool vmlinux.h gen, or - dwarf/btf without decl tag + > > hardcoded bpf ctx info -> bpftool vmlinux.h gen If we intends to cover > > all uapi data structures (to prevent unnecessary CORE relocation, esp. > > for troublesome bitfield operations), hardcoded approach won't work and > > we may have to go to decl tag approach. > > +1 for simplicity of "hard code" approach. > We've stopped adding new uapi "mirror" structs like __sk_buff long ago. > The number of structs that need ctx rewrite will not increase. Ok, I'll submit V2 with changes in libbpf/bpftool to emit preserve_static_offset for predefined set of types.
