On Thu, Feb 3, 2022 at 10:20 PM Rafael David Tinoco
<rafaeldtinoco@xxxxxxxxx> wrote:
>
> >>> As in, do you substitute forward declarations for types that are
> >>> never directly used? If not, that's going to be very suboptimal for
> >>> something like task_struct and any other type that's part of a big
> >>> cluster of types.
>
> >> We decided to include the whole types and all direct and indirect
> >> types referenced from a structure field for type-based relocations.
> >> Our reasoning is that we don't know if the matching algorithm of
> >> libbpf could be changed to require more information in the future and
> >> type-based relocations are few compared to field based relocations.
>
> > It will depend on application and which type is used in relocation.
> > task_struct reaches tons of types and will add a very noticeable size
> > to minimized BTF, for no good reason, IMO. If we discover that we do
> > need those types, we'll update bpftool to generate more.
>
> Just to see if I understood this part correctly. IIRC, we started type
> based relocations support in btfgen because of this particular case:
>
> union kernfs_node_id {
> struct {
> u32 ino;
> u32 generation;
> };
> u64 id;
> };
>
> struct kernfs_node___older_v55 {
> const char *name;
> union kernfs_node_id id;
> };
>
> struct kernfs_node___rh8 {
> const char *name;
> union {
> u64 id;
> struct {
> union kernfs_node_id id;
> } rh_kabi_hidden_172;
> union { };
> };
> };
>
> So we have 3 situations:
>
> (struct kernfs_node *)->id as u64
>
> [29] STRUCT 'kernfs_node' size=128 vlen=1
> 'id' type_id=42 bits_offset=832
> [42] TYPEDEF 'u64' type_id=10
>
> (struct kernfs_node___older_v55 *)->id as u64 (union kernfs_node_id)->id
>
> [79] STRUCT 'kernfs_node' size=128 vlen=1
> 'id' type_id=69 bits_offset=832
> [69] UNION 'kernfs_node_id' size=8 vlen=2
> '(anon)' type_id=132 bits_offset=0
> 'id' type_id=40 bits_offset=0
> [40] TYPEDEF 'u64' type_id=12
>
> (struct kernfs_node___rh8 *)->id = (anon union)->id
>
> [56] STRUCT 'kernfs_node' size=128 vlen=1
> '(anon)' type_id=24 bits_offset=832
> [24] UNION '(anon)' size=8 vlen=1
> 'id' type_id=40 bits_offset=0
> [40] TYPEDEF 'u64' type_id=11
>
> We're finding needed BTF types, that should be added to generated BTF,
> based on fields/members of CORE relo info. How we would know we had to
> add the anon union of the last case if it does not exist in the local
> BTF ? What is your suggestion ?
>
I'd need to see real BPF program code for this situation, but if you
don't have field-based relocation that needs that anonymous union,
then it shouldn't matter if that union is there or not. I suspect you
do have field-based relocations that access fields of struct
kernfs_node___rh8 and kernfs_node___older_v55, so both structs and
necessary fields should be marked as "used" by btfgen algorithm.
> Thanks!
>
> -rafaeldtinoco
