Re: [PATCH] bpftool: output map/prog indices on `gen skeleton`

[sdf@google.com]

On 09/08, Marcelo Juchem wrote:
> The skeleton generated by `bpftool` makes it easy to attach and load bpf
> objects as a whole. Some BPF programs are not directly portable across  
> kernel
> versions, though, and require some cherry-picking on which programs to
> load/attach. The skeleton makes this cherry-picking possible, but not  
> entirely
> friendly in some cases.

> For example, an useful feature is `attach_with_fallback` so that one
> program can be attempted, and fallback programs tried subsequently until
> one works (think `tcp_recvmsg` interface changing on kernel 5.19).

> Being able to represent a set of probes programatically in a way that is  
> both
> descriptive, compile-time validated, runtime efficient and custom library
> friendly is quite desirable for application developers. A very simple way  
> to
> represent a set of probes is with an array of indices.

> This patch creates a couple of enums under the `__cplusplus` section to
> represent the program and map indices inside the skeleton object, that  
> can be
> used to refer to the proper program/map object.

> This is the code generated for the `__cplusplus` section of  
> `profiler.skel.h`:
> ```
>    enum map_idxs: size_t {
>      events = 0,
>      fentry_readings = 1,
>      accum_readings = 2,
>      counts = 3,
>      rodata = 4
>    };
>    enum prog_idxs: size_t {
>      fentry_XXX = 0,
>      fexit_XXX = 1
>    };
>    static inline struct profiler_bpf *open(const struct  
> bpf_object_open_opts *opts = nullptr);
>    static inline struct profiler_bpf *open_and_load();
>    static inline int load(struct profiler_bpf *skel);
>    static inline int attach(struct profiler_bpf *skel);
>    static inline void detach(struct profiler_bpf *skel);
>    static inline void destroy(struct profiler_bpf *skel);
>    static inline const void *elf_bytes(size_t *sz);
> ```
> ---
>   src/gen.c | 32 ++++++++++++++++++++++++++++++++
>   1 file changed, 32 insertions(+)

> diff --git a/src/gen.c b/src/gen.c
> index 7070dcf..7e28dc7 100644
> --- a/src/gen.c
> +++ b/src/gen.c
> @@ -1086,6 +1086,38 @@ static int do_skeleton(int argc, char **argv)
>   		\n\
>   									    \n\
>   		#ifdef __cplusplus					    \n\
> +		"
> +	);
> +

[..]

> +	{
> +		size_t i = 0;
> +		printf("\tenum map_index: size_t {");
> +		bpf_object__for_each_map(map, obj) {
> +			if (!get_map_ident(map, ident, sizeof(ident)))
> +				continue;
> +			if (i) {
> +				printf(",");
> +			}
> +			printf("\n\t\t%s = %lu", ident, i);
> +			++i;
> +		}
> +		printf("\n\t};\n");
> +	}
> +	{
> +		size_t i = 0;
> +		printf("\tenum prog_index: size_t {");
> +		bpf_object__for_each_program(prog, obj) {
> +			if (i) {
> +				printf(",");
> +			}
> +			printf("\n\t\t%s = %lu", bpf_program__name(prog), i);
> +			++i;
> +		}
> +		printf("\n\t};\n");
> +	}

I might be missing something, but what prevents you from calling these
on the skeleton's bpf_object?

   skel = xxx__open();

   bpf_object__for_each_map(map, skel->obj) {
     // do whatever you want here to test whether it's loadable or not
   }

   // same for bpf_object__for_each_program

   xxx__load(skel);

How do these new enums help?