On Fri, Jun 24, 2022 at 2:46 PM Alan Maguire <alan.maguire@xxxxxxxxxx> wrote:
>
> On 23/06/2022 18:56, Andrii Nakryiko wrote:
> > On Thu, Jun 23, 2022 at 4:10 AM Alan Maguire <alan.maguire@xxxxxxxxxx> wrote:
> >>
> >> when CONFIG_NF_CONNTRACK=m, vmlinux BTF does not contain
> >> BPF_F_CURRENT_NETNS or bpf_ct_opts; they are both found in nf_conntrack
> >> BTF; for example:
> >>
> >> bpftool btf dump file /sys/kernel/btf/nf_conntrack|grep ct_opts
> >> [114754] STRUCT 'bpf_ct_opts' size=12 vlen=5
> >>
> >> This causes compilation errors as follows:
> >>
> >> CLNG-BPF [test_maps] xdp_synproxy_kern.o
> >> progs/xdp_synproxy_kern.c:83:14: error: declaration of 'struct bpf_ct_opts' will not be visible outside of this function [-Werror,-Wvisibility]
> >> struct bpf_ct_opts *opts,
> >> ^
> >> progs/xdp_synproxy_kern.c:89:14: error: declaration of 'struct bpf_ct_opts' will not be visible outside of this function [-Werror,-Wvisibility]
> >> struct bpf_ct_opts *opts,
> >> ^
> >> progs/xdp_synproxy_kern.c:397:15: error: use of undeclared identifier 'BPF_F_CURRENT_NETNS'; did you mean 'BPF_F_CURRENT_CPU'?
> >> .netns_id = BPF_F_CURRENT_NETNS,
> >> ^~~~~~~~~~~~~~~~~~~
> >> BPF_F_CURRENT_CPU
> >> tools/testing/selftests/bpf/tools/include/vmlinux.h:43115:2: note: 'BPF_F_CURRENT_CPU' declared here
> >> BPF_F_CURRENT_CPU = 4294967295,
> >>
> >> While tools/testing/selftests/bpf/config does specify
> >> CONFIG_NF_CONNTRACK=y, it would be good to use this case to show
> >> how we can generate a module header file via split BTF.
> >>
> >> In the selftests Makefile, we define NF_CONNTRACK BTF via VMLINUX_BTF
> >> (thus gaining the path determination logic it uses). If the nf_conntrack
> >> BTF file exists (which means it is built as a module), we run
> >> "bpftool btf dump" to generate module BTF, and if not we simply copy
> >> vmlinux.h to nf_conntrack.h; this allows us to avoid having to pass
> >> a #define or deal with CONFIG variables in the program.
> >>
> >> With these changes the test builds and passes:
> >>
> >> Summary: 1/2 PASSED, 0 SKIPPED, 0 FAILED
> >>
> >> Signed-off-by: Alan Maguire <alan.maguire@xxxxxxxxxx>
> >> ---
> >
> > Why not just define expected types locally (doesn't have to be a full
> > definition)? Adding extra rule and generating header for each
> > potential module seems like a huge overkill.
> >
>
> True. I also forgot that if we use vmlinux in the kernel tree as the
> source for vmlinux BTF, this approach won't work since it assumes it
> will find nf_conntrack in the same directory. I'll figure out a simpler
> approach. Thanks for taking a look!
Just define local minimal local definitions of relevant types with
___local suffix, e.g.:
struct nf_conn___local {
long unsigned int status;
/* and whatever else we are using */
};
Keep in mind that with CO-RE triple underscore suffix is ignored,
which can be used to avoid type conflicts
>
> Alan
