On Tue, Apr 5, 2022 at 11:41 PM Nikolay Borisov <nborisov@xxxxxxxx> wrote: > > > > On 6.04.22 г. 2:37 ч., Andrii Nakryiko wrote: > > The problem is that what you've implemented is not a user-space > > equivalent of bpf_core_xxx() macros. CO-RE has extra logic around > > ___<flavor> suffixes, extra type checks, etc, etc. Helper you are > > adding does a very straightforward strings check, which isn't hard to > > implement and it doesn't have to be a set in stone API. So I'm a bit > > hesitant to add this. > > > > But I can share what I did in similar situations where I had to do > > some CO-RE check both on BPF side and know its result in user-space. I > > built a separate very simple BPF skeleton and all it did was perform > > various feature checks (including those that require CO-RE) and then > > returned the result through global variables. You can then trigger > > such BPF feature-checking program either through bpf_prog_test_run or > > through whatever other means (I actually did a simple sys_enter > > program in my case). See [0] for BPF program side and [1] for > > user-space activation/consumption of that. > > > > The benefit of this approach is that there is no way BPF and > > user-space sides can get "out of sync" in terms of their feature > > checking. With skeleton it's also extremely simple to do all this. > > > > [0]https://github.com/anakryiko/retsnoop/blob/master/src/calib_feat.bpf.c > > [1]https://github.com/anakryiko/retsnoop/blob/master/src/mass_attacher.c#L483-L529 > > > > > That's indeed neat, however what is the minimum kernel version required > to have global variables work ? AFAIU one requirement is to use a > recent-enough libbpf which supports the skeleton functionality which is > fine, userspace components can be updated somewhat easily than target > kernels. You need Linux 5.5 for global variables mapping into user-space (added in [0]). But you don't need to use global variables to pass information back to user-space. You can just do a trivial BPF_MAP_TYPE_ARRAY for this. It's a bit more verbose to work with from BPF and user-space side, but effectively is the same for this case. Then you basically have no extra Linux version constraints just because of this feature probing. Only the need to have vmlinux BTF. [0] fc9702273e2e ("bpf: Add mmap() support for BPF_MAP_TYPE_ARRAY")
