On 08/09/2026 17:01, [email protected] wrote: >> diff --git a/tools/bpf/bpftool/Documentation/bpftool-gen.rst >> b/tools/bpf/bpftool/Documentation/bpftool-gen.rst >> index d0a36f442db7..1cdecf3e4fa5 100644 >> --- a/tools/bpf/bpftool/Documentation/bpftool-gen.rst >> +++ b/tools/bpf/bpftool/Documentation/bpftool-gen.rst > > [ ... ] > >> diff --git a/tools/bpf/bpftool/gen.c b/tools/bpf/bpftool/gen.c >> index a50540ef6521..0fcfe20ff515 100644 >> --- a/tools/bpf/bpftool/gen.c >> +++ b/tools/bpf/bpftool/gen.c >> @@ -583,6 +583,9 @@ static void codegen_attach_detach(struct bpf_object >> *obj, const char *obj_name) >> bpf_object__for_each_program(prog, obj) { >> const char *tp_name; >> >> + if (!bpf_program__autoload(prog)) >> + continue; >> + >> codegen("\ >> \n\ >> \n\ > > [ ... ] > >> @@ -1339,6 +1348,8 @@ static int do_skeleton(int argc, char **argv) >> map_cnt++; >> } >> bpf_object__for_each_program(prog, obj) { >> + if (use_loader && !bpf_program__autoload(prog)) >> + continue; >> prog_cnt++; >> } > > Can the autoload flag change between when this check runs and when > codegen_attach_detach() checks it? > > Looking at the code flow, do_skeleton() samples bpf_program__autoload() here > before calling bpf_object__load(), but codegen_attach_detach() and > codegen_destroy() (which also check autoload at gen.c:586, 635, 654, and 686) > run inside gen_trace() after bpf_object__load() completes. > > In tools/lib/bpf/libbpf.c, bpf_object__load() -> > bpf_object_prepare() modifies prog->autoload for struct_ops programs: > > 1. bpf_object__init_kern_struct_ops_maps() (libbpf.c:1251, 1290) sets > st_ops->progs[i]->autoload = false when the member is missing from kernel > BTF or the slot was replaced/NULLed. > > 2. bpf_object_adjust_struct_ops_autoload() (libbpf.c:1166) sets > prog->autoload = should_load for programs referenced by struct_ops maps, > which can flip SEC("?struct_ops/...") programs from false back to true when > the referencing map has autocreate set. > > For the false->true case (SEC("?struct_ops/foo") with an autocreated map): > do_skeleton() omits both the progs and links struct fields, but > codegen_attach_detach()/codegen_destroy() then emit attach functions > referencing skel->progs.foo.prog_fd and skel->links.foo_fd, which weren't > declared. > > For the true->false case (plain SEC("struct_ops/foo") whose member is missing > from kernel BTF): the fields are emitted but attach/detach/destroy skip the > program, and the loader doesn't reserve a prog_fd slot for it, so subsequent > program FDs get written into preceding fields. > > Would it work to sample autoload once and reuse that decision, or to move the > struct field emission after bpf_object__load()?
It looks like both Sashiko and the bpf-ci bot raised valid concerns. Here, it's probably cleaner to move the struct field emission after bpf_object__load()? Seems a bit more involved though, we'd need to move these struct field emission to gen_trace() for the use_loader case, or move bpf_object__load() out of gen_trace(), not sure which is best. This way we'd have the fields printed after bpf_object__load() has updated (if necessary) the autoload status for "obj", and codegen_attach_detach() would then use the correct fields. Quentin
