On Wed, Aug 7, 2019, at 11:57 AM, Andrii Nakryiko wrote: > On Tue, Aug 6, 2019 at 4:42 PM Daniel Xu <dxu@xxxxxxxxx> wrote: > > > > It is sometimes necessary to perform ioctl's on the underlying perf fd. > > There is not currently a way to extract the fd given a bpf_link, so add a > > helper for it. > > --- > > So I've been going back and forth with this approach and the > alternative one, and I think I'm leaning towards the alternative one > still. > > I think it's better to have a broad "categories" of bpf_links, e.g.: > > - FD-based bpf_link (which is the only one we have right now): > bpf_link_fd. It's not just for perf FD-based ones, raw tracepoint is > not, but it's still FD-based; > - for cgroup-related links (once they are added), it will be > bpf_link_cg (or something along the lines); > - there probably should be separate XDP-related bpf_link with device > ID/name inside; > - etc, whatever we'll need. > > Then we can have a set of casting APIs and getter APIs that extract > useful information from specific type of bpf_link. We can also add > direct bpf_link creation API (e.g., from known FD), for cases where it > makes sense. > > So something like (in libbpf.h): > > struct bpf_link_fd; > struct bpf_link_cg; > > /* casting APIs */ > const struct bpf_link_fd *bpf_link__as_fd(const struct bpf_link *link); > const struct bpf_link_cg *bpf_link__as_cg(const struct bpf_link *link); > > /* getters APIs */ > int bpf_link_fd__fd(const struct bpf_link_fd *link); > int bpf_link_cg__cgroup_fd(const struct bpf_link_cg *link); > > /* link factories (in addition to attach APIs) */ > const struct bpf_link_fd *bpf_link__from_fd(int fd); > const struct bpf_link_cg *bpf_link__from_cg(int cg_fd, /* whatever > else necessary */); > > I think this way it becomes obvious what you can expect to get of each > possible type of bpf_link and you'll have to explicitly cast to the > right type. Yet we still hide implementation details, allow no-brainer > bpf_link__destroy regardless of specific type of link (which probably > will be a common case). > > Thoughts? Makes sense to me. This would probably result in a more predictable API when new types are added. I'll make it this way in V2. > > > tools/lib/bpf/libbpf.c | 13 +++++++++++++ > > tools/lib/bpf/libbpf.h | 1 + > > tools/lib/bpf/libbpf.map | 5 +++++ > > 3 files changed, 19 insertions(+) > > > > diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c > > index ead915aec349..8469d69448ae 100644 > > --- a/tools/lib/bpf/libbpf.c > > +++ b/tools/lib/bpf/libbpf.c > > @@ -4004,6 +4004,19 @@ static int bpf_link__destroy_perf_event(struct bpf_link *link) > > return err; > > } > > > > +int bpf_link__get_perf_fd(struct bpf_link *link) > > this seems like a bit too specific name (and we should avoid "get" > words, as we do in a bunch of other libbpf APIs for getters). Maybe > just `bpf_link__fd`? This especially makes sense with a "file-based > bpf_link" abstraction I proposed above. Ok.
