On Fri, 5 Jun 2020 09:58:26 -0700 Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > On Fri, Jun 5, 2020 at 1:23 AM Jesper Dangaard Brouer <brouer@xxxxxxxxxx> wrote: > > > > Great. If we can remove this requirement of -1 init (and let zero mean > > feature isn't used), then I'm all for exposing expose in uapi/bpf.h. > > Not having it in bpf.h doesn't magically make it invisible. > It's uapi because user space C sources rely on its fixed format. > vmlinux.h contains all kernel types. both uapi and kernel internal. > devmap selftest taking uapi 'struct bpf_devmap_val' from vmlinux.h is > an awful hack. > I prefer to keep vmlinux.h usage to bpf programs only. > User space C code should interface with kernel via proper uapi headers. > When vmlinux.h is used by bpf C program it's completely different from > user space C code doing the same, because llvm emits relocations for > bpf prog and libbpf adjusts them. > So doing 'foo->bar' in bpf C is specific to target kernel, whereas > user C code 'foo->bar' is a hard constant which bakes it into uapi > that kernel has to keep backwards compatible. Thank you for taking time to explain this. Much appreciated, I agree with everything above. > If in some distant future we teach both gcc and clang to do bpf-style > relocations for x86 and teach ld.so to adjust them then we can dream > about very differently looking kernel/user interfaces. > Right now any struct used by user C code and passed into kernel is uapi. I like this future vision. I guess this patch is premature, as it operates in the same problem space. It tried to address uapi flexbility, by letting userspace define the uapi layout via BTF at map_create() time, and let kernel-side validate BTF-info and restrict possible struct member names, which are remapped to offsets inside the kernel. I'll instead wait for the future... -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer
